Backlog

Open defects, design calls and debt for Commander In Chief, as of 2026-08-24 (branch main @ 21290ee).

⚠️ Read the severities sceptically — that is a rule this project paid for. Handed-down reports here have repeatedly named a real smell and got the consequence wrong by a whole class: a "softlock" that was minor legibility, a "sector 4 wall" that was a bot that could not aim, a hitbox complaint that was rotor blades. Before designing a fix, enumerate every call site of the predicate you are about to touch and the actual damage/progress path. See docs/ history and the /triple-a-game skill notes for the worked examples.

⚠️ Line numbers in this file drift; symbols do not. Entries were measured against several different trees across a run, and 3be8149 alone moved sim_world.gd by ~+130 lines below the revive block and main.gd by ~+450 below the HUD block. Spot-checked on 3be8149: _draw_threat_pips is cited at main.gd:13195 and lives at 13647; rifleman_windup is cited at sim_world.gd:3689 and lives at 3819; _detonate_tank is cited at :2900 and lives at 3029. Grep the symbol, do not trust the line. Every citation below was written by the lens that measured it and is kept verbatim; where this merge re-located a symbol on 3be8149 the HEAD line is given alongside.

Provenance. Entries come from /triple-a-game runs — a multimodal reviewer judges screenshots while a second lens reads src/sim/ and drives it headlessly to measure. Measured numbers are copied verbatim; a paraphrased measurement is just an opinion. Where an entry was only reasoned about, the entry says so. This file MERGES: entries that shipped are marked RESOLVED with their commit sha, not deleted — a decision and its outcome belong next to each other. Several lens findings re-flag ground a commit the same run already touched; those carry the sha inline so the next reader starts from the remainder, not the original complaint.

Latest merge (2026-08-24, wf_a7d3399b-122, recovered from a crashed run). The AAA-loop run crashed before its own report stage could write this file, so this merge comes from a reconstructed .aaa/ledger.json rather than a live commit range — treat every "verbatim" claim below as exactly as trustworthy as that reconstruction, no more. Net result: zero commits landed. Three cycles, three different ways to lose the work: cycle 1's screen_fx.gdshader knockdown-smear fix was ACCEPTED by the gate at closeness 96, and then its commit died to an API 529 — the accepted diff is stranded on branch rescue/wf122-smear-mix (be20537) rather than on main; see the RECOVERY entry at the top of §2, which must stay first until that branch is drained. Cycle 2 was blocked outright: its screenshot-capture step also died to a 529 before the dossier could be built. Cycle 3 was REJECTED after 4 attempts (a ground-tint / lane-seal change that measurably cost contrast — see §4/§5), and then its own revert died too, leaving the tree in a mixed state that had to be manually reset rather than recovered by git. Nothing this run touched is on main. What it did bank: 35 new findings and 13 new owner decisions (#40–#52) — the decisions are almost all the design questions the accepted-but-stranded smear fix and the rejected ground-tint attempt raised and did not answer themselves.

Prior merge (2026-08-23, f1601b7..3be8149, 6 commits). A tiny window by commit count and a large one by finding count — the run spent its budget on a behaviour lens driving the sim headlessly with god_mode OFF, which is the first pass to do so since §6 discovered that _god_restore had been hiding the ammo economy from every previous measurement. That single change of instrument produced most of what is new below: the riot shield's impossible flank (re-measured independently, confirming a brief filed a day earlier and still unfixed), the MG nest's non-existent facing, the Triple Shot's inverted surcharge, and the Veteran-Point confiscation on quit. Three new owner decisions (#37–#39), all from 3be8149. Also new: a negative-results record (§9) — four suspicions chased and refuted, banked so nobody re-spends the cycle on them — and the first entry in this file whose counter-evidence is the point (the README assertion badge, §5, which bf27b52 proved is unfalsifiable, not stale).

Prior merge (2026-08-22, 9bb1cdb..f1601b7, 21 commits). A short follow-up window, mostly drain: three of the four owner rulings shipped as code in 0f3f480, eight banked findings closed with shas (§8), one previously "not corroborated" title finally located (the screenshots.gd victory pose — 8005b80 + 042a3c7), and two new owner decisions (#35, #36) plus five new findings banked. Every new finding below was re-verified against f1601b7 by reading the cited line, and says which line.

Counts as of the 2026-08-24 merge: 52 numbered owner decisions (#1–#52, of which #6, #8 and #36 are RESOLVED) and open findings across §2–§7 grown by the run's 35 new findings. None of the 13 new decisions (#40–#52) needed re-numbering against the existing 39 — the reconstructed ledger's 61 raw decision reports consolidate down to those 13 because the other 48 are the run re-confirming (often with fresh, independent measurements) decisions this file already carries as #1–#39; where a re-confirmation added a genuinely new angle it is folded into the existing entry's neighbourhood rather than renumbered. Likewise of the ledger's 115 raw backlog reports, most were the run re-discovering findings already banked below; the 35 additions are new titles only. Zero shas map to this window — see the run-window note above for why.

Counts as of the 2026-08-23 merge: 39 numbered owner decisions (#1–#39, of which #6, #8 and #36 are RESOLVED) and ~100 open findings across §2–§7. This run handed over 80 open findings and 48 owner-decision reports; the 48 consolidate into the 39 numbered decisions because many lenses re-filed decisions #15, #23, #30 and #31 verbatim — those entries keep every lens's measurement rather than picking one. Most of the 80 were already banked and are merged in place, not duplicated; 19 are new (6 sim defects, 1 teaching, 7 visual, 3 tooling, 2 process). Five titled fixes mapped to shas in this window (§8). ⚠️ The ratio moved the wrong way again: ~19 banked against 5 shipped, which is §6's standing complaint and the reason this remains a triage queue, not a work queue.

Counts as of the 2026-08-21 merge: 96 open findings and 37 owner-decision reports handed over from the 2026-08-01 → 2026-08-21 window (390c12d..9bb1cdb, 60 commits). The 37 reports consolidate into 20 numbered decisions (#15–#34) — five lenses filed the HUD-channel call and three filed the deletable-miniboss call, so those entries keep every lens's measurement rather than picking one. Of the 96 findings, 12 were measured against an in-flight diff and are already RESOLVED on 9bb1cdb; each says how it was re-verified and stays in place rather than being deleted. ~120 titled fixes are on the run's shipped list — 29 mapped to shas in this window, and 13 that could NOT be corroborated in the code are listed as such in §8 and remain open in their sections. Older carried entries are marked per-section.

⚠️ Second standing warning, new this snapshot: a finding measured against a work-in-progress diff is not a finding against HEAD. Twelve entries below were filed by review lenses reading a branch mid-flight and were fixed before the commit landed. Every entry that claims a defect on HEAD in this snapshot says whether it was re-verified against 9bb1cdb and by what command. Re-verify before you fix; the predicate may already be gone.

TODO — opening combat and readability follow-up (2026-08-01)

Stagger ordinary riflemen's first-shot phases so a formation cannot resolve as one synchronized bullet wall.
Separate living troops from corpse piles: living threats keep the warm separator rim; bodies settle smaller, flatter, darker and desaturated.
Keep rifleman contact lethal only as a deliberate close-range risk. Riflemen stop at a firing line and never seek body contact; the field manual now states the remaining rule.
Add ARCADE and BOSS RUSH to both normal and large-text versions of the MODES manual.
Reduce opening clutter with a 2-second field-spawn grace, a slower pre-gate cadence, and a delayed first spawn from the landing-zone bunker.
If a suicide bomber is added later, require a unique silhouette, visible fuse tell, dedicated approach behavior and an explosion payoff. Ordinary riflemen are not bombers.
Pin splash skipping with a regression test: the skip press is marked handled and returns before the title menu can activate its default row.
Run the opening with a genuinely new player. Pass target: 8/10 first encounters reach the seawall, identify the rifleman's painted lane, and can explain any death without help. If fewer pass, lengthen the grace by 30 ticks before reducing enemy damage or cadence.

1. Owner decisions — these need a human, not a fix

✅ OWNER RULINGS — 2026-08-21, three of them SHIPPED as code in 0f3f480 on 2026-08-22

Four decisions answered by the owner directly. Do NOT re-open these in a later cycle; a reviewer re-reporting one of them should be dismissed with a pointer here.

Implementation status (0f3f480, "three owner rulings"):

§1.2 hazards — SHIPPED. Held per-PLAYER (a 2P partner who is not shopping still eats it): the endless mast pulse, the foundry vent jet and every in-flight telegraphed strike. Deliberately NOT held, with reasons inline: enemy movement and ordinary bullets, mines/claymores/barrels (those only fire because you WALKED or SHOT — pausing them makes the hold key a minesweeper), tank shells, enemy windups. The exploit was real and is bounded: _update_wheel holds the wheel open with no timeout, so WHEEL_PAUSE_MAX_TICKS = 90 (1.5 s, covers the ~1 s buy) drains while open and refills 1 per 4 ticks only while CLOSED — a full 90 t shield costs 360 t of shopping with the plate down. ⚠️ The sim guard shipped correct, tested and UNREACHABLE: BUY_WHEEL_OPEN existed only in sim_world.gd because the slice that wrote it did not own main.gd, and _update_wheel's held-open path returned a bare 0. One line now returns the sentinel. A guard nothing can reach is not a guard — add that to the read-sceptically list at the top of this file.

§1.9 salvage — SHIPPED, routed through the existing _supply_full predicate rather than a second copy of the rule. The refusal is loud (deny{why:"salvage_full"} → "GRENADES FULL") and returns true so INTERACT is swallowed rather than falling through to arm a claymore underfoot.

Off-frame label tolerance — SHIPPED at 12px, NOT the 24 the ruling suggested, and the number is measured rather than picked: 12.0 is the smallest above-anchor extent of any labelled subject's own art (the downed body's KO ring); 24 would admit labels for subjects with zero pixels on screen. WORLD_LABEL_SUBJECT_FRAME derives from WORLD_LABEL_FRAME so the two cannot drift. The 0% result was re-measured after the change and HOLDS: 0 overlaps across 7,200 frames at 100% AND 200% text scale.

§1.10 airstrike deny — no code; now explicitly DECLINED in this file rather than sitting open.

Goldens did not move, verified rather than hoped: git diff src/sim/ | grep 'rng\.' is empty (no draw added, removed or reordered) and test_determinism.gd is untouched. Suite 1146 → 1154 methods, 37,356 → 37,416 assertions, 0 failures.

# Question Ruling

§1.2 Shop-wheel plate alpha vs. hazards reading through Keep the readable plate; PAUSE the hazards during the buy. The lower alpha was the wrong lever — this is the fix the entry itself named.

§1.9 Hulk salvage at grenade cap Refuse when the strip would deliver 0; allow at >= 1. ⚠️ Accepted consequence: players spawn AT GRENADE_AMMO_MAX, so a fresh player meets a hulk wall they cannot clear until they spend a grenade. The refusal is therefore required to be AUDIBLE (deny{why:"salvage_full"}) so the rule is readable rather than inferred.

AAA-run Off-frame world-label suppression Add a small negative-y tolerance (~Rect2(0,-24,640,384)), so a subject just above the viewport keeps a clamped label. The 0%-overlap result (from 46.32%/47.01% across 14,400 frames) must NOT regress — measure it after.

§1.10 Flat airstrike deny removing the skilled pre-call Keep the deny. The pending_airstrike freeze route is explicitly declined; the trade-off is now permanent by decision, not by omission.

#	Question	Ruling
§1.2	Shop-wheel plate alpha vs. hazards reading through	Keep the readable plate; PAUSE the hazards during the buy. The lower alpha was the wrong lever — this is the fix the entry itself named.
§1.9	Hulk salvage at grenade cap	Refuse when the strip would deliver 0; allow at >= 1. ⚠️ Accepted consequence: players spawn AT `GRENADE_AMMO_MAX`, so a fresh player meets a hulk wall they cannot clear until they spend a grenade. The refusal is therefore required to be AUDIBLE (`deny{why:"salvage_full"}`) so the rule is readable rather than inferred.
AAA-run	Off-frame world-label suppression	Add a small negative-y tolerance (~`Rect2(0,-24,640,384)`), so a subject just above the viewport keeps a clamped label. The 0%-overlap result (from 46.32%/47.01% across 14,400 frames) must NOT regress — measure it after.
§1.10	Flat airstrike deny removing the skilled pre-call	Keep the deny. The `pending_airstrike` freeze route is explicitly declined; the trade-off is now permanent by decision, not by omission.

tools/probe_concussion_hud.gd runs in NO automated gate. The plan said to wire it into "the CI step that already runs gl_compatibility pixel smoke", but no such CI step exists (verified .github/workflows/ci.yml — every job is --headless). The CI-gated coverage is smoke.gd's structural layer pin (mutation-proven); the pixel probe is local-only. Adding a GL-capable CI job (macOS runner, Xvfb+gl_compatibility on Linux) is a cost/flakiness tradeoff only the owner can call. (Re-flagged 2026-07-31, unchanged.)
Shop-wheel plate alpha overturns a deliberate design. The c2 author cut the plate to 0.55 on purpose so mast, scars, drops and hazards read through during a buy; a later cycle raised it. Measured cost: a ~100px donut of ground around a stationary soldier is masked for the ~1s buy, and terrain contrast through the disc drops from ±24.3 to ±3.8 luma. If hazard-reading-during-buy is a real requirement, the honest answer is pausing hazards or offsetting the wheel — not a lower alpha. (2026-07-29: ee25263 shipped a shop-overlay scrim fix, and the lens still flags the unmasked radial menu — see §4. The alpha-design question above is still unanswered.)
The cue row is flush-cut at the arena edge. It is ~167px wide including its plate, but _draw_wheel clamps c.x to [78, 562] on a 640px viewport, so at the left edge the plate starts at about x = -8. Reads as deliberate framing or as a bug. If it should never touch the edge, raise the clamp floor to ~87 (and drop the ceiling symmetrically).
test_no_hostile_stalls_an_endless_wave samples in 800-tick windows because it force-clears the field by fiat. It therefore pins "no wedge within any 800-tick window", not "no wedge, ever" — and the mutation run logged a 764-tick streak, comfortably inside one window. (Note: the stated reason — "there is no combat-capable headless bot" — is no longer true; demo_input gained endless targeting on 2026-07-26. This test can now be strengthened.)
The mg_nest teaches a second lane grammar. Sniper/elite/ghillie/technical paint "this line is committed, sidestep it"; the nest paints "this line follows you". Locking the sim to match the others would delete the tracking rake the sim comments call deliberate, and move test_determinism GOLDEN. Keep two grammars, or lose the rake. (2026-07-29: 37e8287 shipped "MG nest telegraph never fires down the drawn lane" — the honesty half. The two-grammars design call itself is still open.)
~~Should the nest lead its target?~~ RESOLVED 2026-07-26 — ca3ed3f added lead (SimWorld.mg_nest_led_aim), no new sim state, view follows via main.telegraph_dir(), goldens re-recorded (only the trailing sample moved).
HOW TO PLAY ignores the 200% TEXT SIZE accessibility setting. Art.text_scale routes only through Art.fs(), which no HOWTO call site uses. Honouring it would overflow every band the FRAME_INNER_* constants just tightened — a real scope decision, not an oversight.
~~duplicate of #6~~ RESOLVED.
NEW 2026-07-31 — Hulk salvage at grenade cap is a cover-destroying no-op. The untracked tools/probe_salvage.gd in this tree measures that a player at GRENADE_AMMO_MAX salvaging a smoldering hulk gets _try_salvage_hulk returning true, event n=0 (+0 grenades), and the hulk's burn_ticks stripped 100 → 0 — the cover is destroyed for zero gain. The sim docstring (sim_world.gd:2392-2395) frames strip-vs-keep as a deliberate trade ("keep the wall or take the ammo"), and the shop's own _supply_full rule refuses no-op purchases everywhere else ("none of them can bill for a no-op", tests/test_shop.gd:249) — so either salvage should refuse at cap (keep the wall) or the docstring's trade framing should be narrowed. A +0 "trade" is not a decision the player can read, but reversing the documented choice is the owner's call. The defect entry lives in §2. (probe_colossus.gd, the other orphan probe, runs clean: seed 0xC0FFEE engage_t=288 fight_ticks=683 downs=3 victory=true.)
NEW 2026-07-31 — The flat airstrike deny removes the skilled "pre-call". Buying in the last ~44 intermission ticks so the strike lands mid-trickle clips ~4 enemies; 8425bd7's deny kills that line. The plan weighed freezing pending_airstrike instead and rejected it (view countdown + lead-prediction + checksum ordering). If the owner considers the pre-call a feature worth keeping, the freeze alternative is the only honest route back — the deny makes that trade-off permanent.
NEW 2026-07-31 — ghillie and sapper still have no working separator rim. Their _LIGHT_RIM entries remain dead behind the OUTLINE gate, same root cause as the m_soldier2 defect this cycle fixed. The plan deliberately deferred them: ghillie's cloak subtlety is the mechanic and sapper is a non-shooter with a satchel pip. If a future pass rims them, the same three-registry pattern (OUTLINE + _UNIT_RIM + _LIGHT_RIM) and the same source-derived test shape apply.
NEW 2026-07-31 — The warm-light rim visibly enlarges/brightens the infantry silhouette slightly (measured ~120 changed px per ~18px sprite). This is the a1-02 separator class working as designed, but it is an aesthetic call: the rim reads as a soft glow outline rather than a hard keyline. If the owner prefers subtler separation, _LIGHT_RIM's 2.2px width is the single knob.
NEW 2026-07-31 — An abandoned daily now leaves NO Hall/leaderboard trace of the attempt. d45132a deliberately chose arm+demote over the finding's alternative of banking the quit with its standing score — the daily board only ever shows the one completed run. If the owner wants "attempted, abandoned" visible on the daily leaderboard, that is a product call, not a defect.
NEW 2026-07-31 — A spent-daily re-deal keeps the SAME board as unranked practice rather than re-seeding a fresh layout — d45132a's chosen behavior; a fresh-seed practice would play differently and is equally defensible.

New 2026-08-21 — from the `390c12d..9bb1cdb` window

Several lenses reported the same decision independently. Where that happened the entry keeps every lens's measurement verbatim rather than picking one, because independent measurements of the same number are the only evidence that the number is real.

NEW 2026-08-21 — The HUD's stacked-chrome channel is now governed by one hand-picked number, and paying for it moved two shipped screens. Shipped in 66d146a. Five lenses reported this; all five measurements are kept.
- BOSS_BAR_TOP moved 64.0 → 77.0, pushing the whole top-center boss/mini HP-bar stack 13px DOWN into the play field (slot-2 bar bottom moves 121 → 134). That is a visible layout shift in every shipped screenshot and marketing frame, and it trades 13px of world visibility at the top of the screen for the 6px channel under the corner panel. The alternative — leaving the bar line at 64 and shrinking the corner panel or the label plate instead — was not costed. Measured: panel_bottom() maxes at 60 in 1P-endless/2P-campaign/2P-endless, so the minimum BOSS_BAR_TOP that clears it by MIN_HUD_CHANNEL=6 is 77; anything less needs the panel to give ground instead.
- MIN_HUD_CHANNEL was set to 6.0 and BOTTOM_RESERVE_GAP to MIN_HUD_CHANNEL+2 = 8.0, overturning a documented, deliberate BOTTOM_RESERVE_GAP := 3.0 and its band-contract docblock. 6 and 8 are asserted as hard floors across ~12,700 pair checks, so every future piece of HUD chrome is now bound by them. Measured basis: the true channel on HEAD was 1px between two dark plates overlapping 225px horizontally. Whether 6/8 is the right house minimum at 640x360 (vs 4, which would leave BOSS_BAR_TOP at 75) is a taste call the reviewer cannot make.
- Second lens on the same number: MIN_HUD_CHANNEL at 6.0 px is what moved the verb-chip/colossus channel from a measured 1 px to a measured 8 px. Nothing in the repo says what the intended channel is — 6 was chosen by the implementer. At 640x360 upscaled 2x-3x this reads clean in my capture of 05-foundry-colossus-last-stand, but it also pushed the whole bottom overlay cluster up 7 px (lift 48 → 55) and the boss bars down 13 px, a visible layout change to two shipped screens. If the owner has a house number for HUD breathing room, set it once here and let everything derive.
- Third lens: the measurement justifying BOTTOM_RESERVE_GAP 3.0 → 8.0 is sound (the true channel was 1px, not 3, because COLOSSUS_BLOCK_TOP was a 2px-optimistic ascent mirror), and 8px is screenshot-confirmed as readable separation. But 6.0 was chosen by eye, not against a legibility criterion. Owner call: is 6px the house minimum, or should it be derived from the font's line box?
- Fourth lens, on the boss bars specifically: BOSS_BAR_TOP 64.0 → 77.0 pushes the bars and their phase labels 13px down on every boss encounter and in Boss Rush. Re-captured 04-bridge-gunship.png and confirmed the result is legible and clear of the corner panel (on HEAD the panel buried the top of the label) — but this is a shipped-layout change made to fix a defect the reviewer never reported. The alternative fix — shrinking the corner panel so panel_bottom() clears the existing block top — was not attempted. Is 13px of play field the right currency to pay, or should the panel shrink instead?
NEW 2026-08-21 — LAST_STAND_Y moved 350 → 353, not the 352 the plan specified, putting the banner plate at 341..358 with 2px to the 360 viewport floor. The extra pixel buys MIN_HUD_CHANNEL exactly against COLOSSUS_BLOCK_BOTTOM=335 rather than 5px. Deliberate and documented in the constant's comment, but it is a 1px deviation from the approved plan on the element closest to the screen edge — worth a nod before it becomes the reference value. Second lens, same element: the LAST STAND banner plate now occupies y 341..358 of a 360-px-tall viewport — 2 px from the floor. That is deliberate (the comment says "still on the viewport floor") and no test complains, but it is the tightest margin on screen and there is no constant expressing "how close to the floor is acceptable". Worth an explicit floor-clearance constant if the owner cares.
NEW 2026-08-21 — Neither boss phase label is localized, and neither scales with TEXT SIZE. Two separate exemptions on the two highest-stakes strings on screen, both banked deliberately by the implementer and both re-verified open on 9bb1cdb (grep -n 'GUNSHIP_PHASE_NAMES\|COLOSSUS_PHASE' src/main.gd → main.gd:11026 and :11193 interpolate the bare consts; no TranslationServer.translate(), no Art.fs()):
- Localization: main.gd draws the English literal at the gunship and colossus draw sites, so the localization suite cannot see them, and a longer translation would widen the plate past everything the new 12,700-pair ratchet measures. The new test pins the related Art.fs() gap explicitly but deliberately does not pin this one. Whether shipped locales are allowed to leave boss names in English is a product call.
- Accessibility: neither label calls Art.fs(), so the TEXT SIZE setting does not scale them — the new test asserts this as a known gap (label width byte-identical at text_scale 1.0 and 2.0 while Art.fs(10) correctly returns 20). Putting them on Art.fs() would widen the plates and require the ratchet to grow a text-scale arm. Owner decides whether the two most important strings on screen are exempt from the accessibility setting.
NEW 2026-08-21 — The shutdown-leak gate is set at ZERO with no floor and no exemption, and every measurement behind it was taken on macOS/arm64 under the headless dummy renderer. CI runs a 3-OS matrix (Linux · macOS-arm64 · Windows) and those two other runners are unmeasured. If Godot's Linux or Windows headless build reports even one environment-specific engine leak at exit, the gate goes red on a clean commit — and CLAUDE.md itself warns that a randomly-red gate stops being read (that is exactly why test_perf.gd was moved to OPT_IN_SUITES). Owner call: push and find out, or soften the CI copy of the gate to a warning for one cycle while the other two OSes are characterised.
NEW 2026-08-21 — The plan named TWO probes for promotion into tools/; only one shipped. probe_rooted_cover.gd (the census) shipped; probe_rooted_trap.gd (the offset x-position trap sweep) was folded into ratchet B (test_rooted_unit_born_on_cover_is_killable_from_the_south) and discarded. Defensible (the ratchet has no sampling window and the probe would) — but the repo now has no standalone measuring tape for "is THIS unit at THIS position shootable", which is the question the next cover/hitbox change will want to ask. Owner call whether to ship it.
NEW 2026-08-21 — _spawn_clear_x duplicates the endless spawn-x domain as a magic number. It bounds its search to a literal [24, 616] * F_ONE (sim_world.gd:4325, :4328), mirroring the endless wave spawner's raw rng.range_i(24, 616) roll. There is no existing named constant for that domain — ARENA_MARGIN is 96 * F_ONE and means something else (colossus/hazard corridor), and SCREEN_W_FP is the full 640. So the pair is a duplicated magic number that will silently desync if the spawn-x roll is ever retuned. Whether to introduce a shared SPAWN_X_MIN/MAX (and re-point the roll at it) is a design call the implementer did not make alone.
NEW 2026-08-21 — The cover seam displaces MOVERS as well as rooted units, and that is what cost a golden re-record. _spawn_enemy (rushers/elites) and every archetype _spawn_special birth now get shoved out of all four armor families, not just rooted ones. MEASURED: reverting ONLY _spawn_enemy reproduces all six retired goldens byte-for-byte, so the rooted calls are golden-inert and the walker seam owns 100% of the churn. A walker born in cover self-corrects in a few ticks (it moves), so the walker half is prophylactic rather than a defect fix, and it cost a golden re-record. If the house rule is "behavior changes only at phase boundaries", shipping a golden re-record for a prophylactic change is the owner's call.
NEW 2026-08-21 — The wave-5 supply pod's rim rocks land 4px from the rooted spawn row and nothing in the tree pins that relationship. Rim rocks at y = -312 (ARENA_L_SLOTS[5] = [320,-312]) while _rooted_spawn_y() is pinned at camera_top + 52 = -308 in endless. Four pixels is the whole margin between "legitimate cover" and "trap" for a 10px-reach ghillie. Neither constant references the other and no test asserts the gap. Whether to make the rooted row derive from (or assert against) the arena slot geometry is a product/architecture call.
NEW 2026-08-21 — The wave-5 miniboss can be destroyed before it ever lands, and nothing pins whether that is intended. Shipped in 016b4cd (the fly-in became damageable). Three lenses measured this independently; all three sets of numbers are kept, because they disagree in rate and agree in kind.
- Lens A, shipped bot over 5 endless seeds: killed mid-approach on seeds 11 (phase_t=-95) and 99 (phase_t=-16); reduced to hp 8/40 by arrival on seed 7; 31/40 on seed 42. The plan asked for a ratchet pinning "cannot be deleted before it lands"; the implementer correctly measured that claim as false and substituted a tracking-reward pin instead, which is the right call. Boss HP at wave 5 is 40 (BOSS_HP, sim_world.gd:708) against a 420-tick window in which a perfectly-tracking 1P shooter lands ~53 rounds.
- Lens B, demo_input bot, god_mode, 7 seeds, SimWorld.new(seed, 1, 'endless') driven to the miniboss: 3 of 7 seeds destroy it mid-approach (seed 11 at phase_t -170, seed 3 at -227, seed 23 at -33); the 4 that land arrive at 8, 4, 32, 22 of 40 HP — median 15/40. The coder's own comment claims 2 of 6 and median 23/40 without god_mode; directionally the same. So roughly 40% of the time the "GUNSHIP INBOUND" banner, alarm and _trauma 0.4 announce a boss that dies before it is even clearly visible through the HUD panel.
- Lens C, shipped demo_input bot, 7 seeds, 6 field a miniboss: 2 of those 6 destroy it mid-approach (seed 11 at phase_t -95, seed 99 at phase_t -16) — the arrival beat, the strafing run and the whole engaged cycle never happen. Of the 4 that land, hp@arrival is 8 / 15 / 31 / 38 (median 23 of 40).
- Instrument limit, stated by two of the three lenses: these are the shipped BOT's numbers with open-loop aim (sector_probe.gd:12-20), not a human's. A human tracking the hull is strictly faster, so the real skip rate is higher than 2-in-6.
- DECIDE: is a deletable approach the intended skill payoff — in which case pin a FLOOR (e.g. the bot must not average better than N HP of damage during the window, so a later radius/HP edit cannot quietly make it free) — or must the gunship always land, in which case it needs damage resistance or a shorter/faster ramp during phase_t < 0? Right now it is pinned in NEITHER direction.
NEW 2026-08-21 — The fly-in ground shadow is pinned to the LANDING PAD, not to the hull, and it is now the only thing left at a point the helicopter is not at. main.gd:10892 (re-verified on 9bb1cdb): _ground_shadow(ground + Vector2(0, 30), 8.0 + eta_f * 18.0, 0.12 + eta_f * 0.30) where ground is the arrival point — so at phase_t -420 the shadow sits at screen (320, 80) while the hull is at (470, -5), the same 159.8 px the hitbox fix just closed. It grows 8→26px / 0.12→0.42 alpha as the hull descends, which reads as a deliberate landing-zone telegraph. But it is now the only piece of the gunship still anchored to a point the helicopter is not at, up to 159.8 px from the caster. Pinned telegraph or tracking caster-shadow — it cannot honestly be both now that the hitbox moved to the hull. If it stays pinned it probably wants to stop looking like a shadow (a reticle/landing-ring reads as a telegraph; a soft ellipse reads as a shadow and lies). Not a defect either way; it needs intent.
NEW 2026-08-21 — The fly-in is no longer a cinematic, and that overturns a stated intent on purpose. sim_world.gd:6303-6306's stated intent — "phase_t starts NEGATIVE: a 7s fly-in (8v ramp smoothing — the gunship no longer lands on top of the wave_start card)" — was overturned by 016b4cd, and the diff does say so in its replacement comments. Confirm you want that: the smoothing beat survives (ramp, haze, scale and rotor spin are untouched), but the gunship is now a live target for the entire 7 s rather than a cinematic. If the original card-collision problem is still a concern, the fix for it is now the spawn-tick banner ordering, not invulnerability.
NEW 2026-08-21 — Field Manual page-count vs page-density. Entry-granular pagination adds a leaf at 125% WAR CHEST (2 → 3), 200% WAR CHEST (5 → 6) and 200% SPECIALS (4 → 5), and leaves pages as sparse as 14% and 34% fill, in exchange for never breaking a sentence. The alternative — allow a mid-entry cut when the break lands on a comma or clause boundary — would keep pages full at the cost of the reviewer's original complaint recurring in a milder form. No test currently pins either side, so whichever way this goes it should get a ratchet.
NEW 2026-08-21 — Should the broadcast mast get a visual arrival at all? Its broadcast_pulse fires on its first stepped tick (measured campaign n=20, min=max=1 tick) and main.gd:3180-3186 already draws that as an 8→140 px shockwave, so the plan called it covered and deliberately excluded the mast from the opacity ramp. But the mast is a 43px radio_tower plus two 48px base arcs and it is now the only rooted kind that appears at 100% opacity with no ramp — a player watching one wave will see the mast pop while the nest beside it fades. The cycle goal's wording ("every turret, mast and hidden sniper") would read as covering the mast too. If the owner wants it to fade, its draw branch needs to read a ramp key that _consume_events currently and intentionally does not register. Only the owner can say whether the ring is enough of an arrival for that silhouette.
NEW 2026-08-21 — Rooted arrival audio borrows the STRUCTURAL-breach channel. The arrival cue is alarm_low at -15 dB / 0.55 pitch, sharing a sample with flank_warn (-11 dB), flank_breach (-6 dB) and rear_warn (-12 dB) — all of which the file's own a1-13 taxonomy comment reserves for STRUCTURAL breaches. Up to three rooted units land per endless wave, so this adds a quiet structural rumble ~3x/wave to a channel the taxonomy says means "the walls answer". Whether a grunt-tier emplacement belongs in that vocabulary at all, versus a dedicated dirt/thud sample, is a sound-design call.
NEW 2026-08-21 — Rooted units can now be born as far south as screen row 287, i.e. BEHIND the player. Measured on the god-mode bot in endless: 146 of 168 births (86.9%) land south of the player, vs 93 of 115 (80.9%) at HEAD — so it was already the norm, but the southern tail is new. In campaign the camera scrolls past them and _step_enemies culls at camera_top + 420 (sim_world.gd:3542), so the off-screen-but-still-firing window is 60 px of advance either way, identical to HEAD — checked; it is not a regression. The product call is whether a rooted MG that materialises behind the player is the fight you want; a y > player_y exclusion (or a southern cap around row ~240) is the lever if not.
NEW 2026-08-21 — How far is the ground repeat worth chasing? Two lenses, same conclusion, different numbers; both kept.
- Lens A: the residual 64px repeat measures 0.62–0.70 lag-64 autocorrelation on real frames but only ~4.4% luminance amplitude (3.29 std on a 74.02 mean at the sector-1 stop), and at 10x contrast boost I could see no lattice in the frame either before or after this diff. Killing it properly means touching the BASE layer (a second seamless sand strip at 0.75 scale / 96px pitch at partial alpha, or a hashed per-row x phase on the existing tiled strip), which is a bigger visual risk than the tell it removes. Ship the −5% and close the finding, or spend a cycle on the base layer?
- Lens B: the repeating signal is ~3.3 luminance units on a 74 mean (4.4%), and at 6x high-pass contrast boost I could not see a grid in EITHER the before or the after frame. Killing the residual 96px repeat costs a second source tile (art + ASSETS.md provenance) or a shader on the base strip. If the owner's read is that the original reviewer over-claimed this tell — which the verification's own numbers support — then the shipped diff is already the right stopping point and this should be closed as won't-fix rather than worked.
NEW 2026-08-21 — The off-frame world-label gate deliberately overturns the documented keep-place rule: a label whose subject the player cannot see is now SUPPRESSED, not relocated. Shipped in 9bb1cdb. Four lenses reported the same trade; all four framings are kept because they photograph different casualties.
- The gate can now suppress a label whose SUBJECT is on-screen. A ford/objective sign anchored ~20px above its subject, with that subject at screen y < 20, produces a want-rect entirely above y=0 and is dropped; previously it clamped to y=0 and printed (under the HUD panel). Net effect at the very top edge is a brief label flicker as the world scrolls in, traded against the 46%-of-frames overprint the gate removes. Confirm the flicker is the trade you want, or add a small negative-y tolerance (e.g. intersect against Rect2(0,-24,640,384)) for anchored objective signage only.
- Confirm the trade is what you want for ALL 15 producers (LOW FUEL, ESCAPING!, RESCUE +N, SILENCE THE SPOTTER, rally countdown, GET UP, REVIVE…), not just the crate labels the reviewer photographed — an off-screen ESCAPING! prisoner now has only the chevron cue.
- It is right for a crate 40px above the frame, but a subject 1px off the top edge loses its label entirely with only the threat chevron as a cue. Is suppression correct at the boundary, or should a subject within ~8px of the frame edge keep a clamped label?
- Measured payoff, so the trade is priceable: 0 label-on-label or label-on-player overlaps remain across 14,400 sampled bot frames, versus 46.32% / 47.01% at HEAD. The casualty is that e.g. a downed partner just off the top edge no longer shows REVIVE <cost> at all, where HEAD clamped it to y=0. Routing those through the existing off-screen chevron / CALLOUT_OVERFLOW tail with an explicit indicator is the alternative.
NEW 2026-08-21 — The repo now has TWO GL-only rendered-pixel probes with zero callers and no CI job that can run either. grep -rn ground_lag . --exclude-dir=.git finds only the file itself and two test comments; tools/ground_profile.py has zero references anywhere (both re-verified on 9bb1cdb). .github/workflows/ci.yml runs only --import, lint_sim, lint_assets, i18n_check, smoke, run_tests and the export/soak jobs — no gl_compatibility job, no xvfb (grep -n 'gl_compatibility\|xvfb' .github/workflows/ci.yml → no matches). GL does work on the dev box (14 signature shots in under 3 minutes), so the gap is specifically the Linux runner needing xvfb. Owner call: wire a --rendering-method gl_compatibility + xvfb job that runs both at a fixed threshold, or accept that rendered-composite autocorrelation is permanently a manual developer-machine check on a 4.4%-amplitude signal.
NEW 2026-08-21 — Ground dressing now costs +25 draw calls/frame (+9.5%) as a standing price, and 8 base strips cost 4.00 MB of raw pixel data. Two knobs, one budget question.
- Draw budget: tools/perf_probe.gd on real GL, 121 frames, driven by demo_input: draw_calls_avg 259.760 → 284.512, draw_calls_max 272 → 309, cpu_ms_avg 0.801 → 0.883, cpu_ms_max 2.557 → 2.681. Fine against a 16.7 ms budget on the machine measured (0.88 ms of frame), but the export targets include Windows/Linux boxes nobody here has profiled, and test_perf.gd is sim-tick only and cannot see this at all.
- Memory: GROUND_BASE_VARIANTS = 8 costs 4.00 MB of raw pixel data (measured; ~5.6 MB VRAM per the diff's profiler note; perf_probe confirms 62.57 MB → 68.15 MB) to attenuate a residual the diff itself measures at +0.041 → +0.007 going from N=4 to N=8. A real but small win for 2 MB. Nothing in the repo states a VRAM budget, so whether that trade is right for the lowest-spec target is the owner's.
NEW 2026-08-21 — The base sand pitch is now 96px while the dressing cell stays 64px, so a y-axis residual lives permanently at their lcm of 192. Measured +0.058 / +0.055. Driving it out means picking a second non-commensurate dressing pitch or accepting it. The current choice trades a 4x reduction for a surviving 192px period; whether that is the right stopping point is the owner's.

New 2026-08-22 — from the `9bb1cdb..f1601b7` window

NEW 2026-08-22 — The OPTS "DEFAULTS RESTORED" banner was demoted from font 9 to font 8 to buy the focused-row halo its 2px channel. src/view/menu.gd, _draw_opts_header — the draw is at menu.gd:5945 on f1601b7 (_center_text("DEFAULTS RESTORED", OPTS_SUBLINE_Y, 8, ...)). Measured: at 9 its ink reached y96 against a ring top of 97 — a 1px overrun; at 8 it clears. The in-code note explains why the alternative was rejected: raising HEADER_CLEAR_COMPACT (menu.gd:324, currently 8.0) to 9 costs the DIRTY 11-row OPTS column a plate row (bh 18→17) and halves its icons (16→8). Trading a confirmation banner's emphasis for a layout row is a taste call and the owner may prefer the other side of it — the green is arguably carrying the emphasis either way. (Reasoned + measured by the implementing lens; the font-8 call site is verified on f1601b7, the y96/y97 overrun is the lens's measurement.)
~~NEW 2026-08-22 — tools/versions.lock pins CI to Godot 4.7.1-stable while local dev now runs 4.7.2, and the lock's own justification is therefore false.~~ RESOLVED 2026-08-23 — cebc3b7 bumped the pin to 4.7.2-stable (owner call: "bump godo"). Both goldens reproduce UNCHANGED on 4.7.2, so no re-record; the new sha256 was authenticated against the release's published SHA512-SUMS.txt before being hashed, and the docs/badges follow the lock rather than being "corrected" independently. The all-platforms half of the lock's own upgrade rule is CI's 3-OS matrix, and the commit message commits to reverting rather than re-recording if any platform disagrees. Original entry follows. Surfaced by a39c1f3 and deliberately NOT executed there, because it is policy rather than chore. Verified on f1601b7: tools/versions.lock still reads godot_version=4.7.1-stable, and its comment still claims "4.7.1 is what local dev already runs, so this aligns CI with the toolchain the suite is actually verified on" — which is now the opposite of true. The suite, goldens and both lints pass on 4.7.2 locally. But the lock states upgrades happen "only at phase boundaries after the replay regression suite passes on all platforms", and a bump changes the engine the determinism goldens are verified against. Owner call: bump the lock (and re-verify goldens on all three CI OSes), or correct the comment so it stops asserting something false.

New 2026-08-23 — from the `f1601b7..3be8149` window

All three come out of 3be8149 (the compounding broke rally). The fix is landed and green; these are the three balance questions it deliberately did not answer, and each one is a knob the implementer set by judgement rather than by measurement.

NEW 2026-08-23 — BROKE_WAIT_MAX_MULT = 4 freezes the brake at exactly the depth where the price accelerates hardest. sim_world.gd:257 on 3be8149 (const BROKE_WAIT_MAX_MULT := 4), consumed by broke_wait_ticks() at :2117 (return BROKE_RESPAWN_TICKS * clampi(p["deaths"], 1, BROKE_WAIT_MAX_MULT)). The free-rally wait tops out at 1200t (20.0 s) from the FOURTH death on, while revive_cost keeps compounding uncapped (wave 60 / deaths 12 = 7800 coin). MEASURED post-fix, deaths 1..6 → 300 / 600 / 900 / 1200 / 1200 / 1200 t. Assertion (a) in the new ratchet only requires monotone >=, so the flat ceiling is deliberately permitted by the test. Whether the cap should be higher, or removed so the wait tracks the price all the way up, is a balance call.
NEW 2026-08-23 — The 2P paid self-revive still buys ONLY time; the plan declined to change that, so the reviewer's complaint is answered on the time axis only. The coin revive lands at _checkpoint_y() with the same stripped body as the free rally — MEASURED dx 0 px, dy 0 px, vest/triple/claymores all stripped on both paths. The fix widened what the coin buys from 299 saved ticks to 1199 saved ticks at deaths 6, and nothing else. The plan deliberately declined to move it to target["y"] the way the SOLO self-revive already does (sim_world.gd:2174-2179 on 3be8149, whose own comment records the change), on the grounds that the checkpoint penalty is what makes waiting for the partner rescue — which lands you at the partner's x AND y, a jump of up to the full 608 px arena — the better play. That is an intent call, not a defect. But the original finding's second sentence ("the paid revive buys only 5 seconds, nothing ELSE") remains true about everything except the clock.
NEW 2026-08-23 — Early endless waves are still a free rally by default, and the fix does not touch them. MEASURED, zero-spend run, per-wave income for waves 1-6: 55 / 130 / 260 / 385 / 380 / 525 against a revive_cost of 50-100 — so the chest naturally sits at or under the price there, rally_is_free() holds, and a first death costs the unchanged 5.0 s. Wave duration under the demo bot in that same range measures ~470t (7.8 s) median, so a first death in waves 1-3 is ~64% of a wave and the compounding never starts. Whether the early game should carry any coin brake at all — or whether the loadout strip, the burned Commendation and (from wave 2 on only) the 40-coin Clean Wave forfeit are the intended entire bill — is unaddressed by this fix and is a difficulty call.

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json` after the run crashed)

All thirteen come out of the same run's two cycles that touched code: cycle 1's screen_fx.gdshader knockdown-smear fix (ACCEPTED at gate closeness 96, then stranded on rescue/wf122-smear-mix — see the recovery entry at the top of §2) and cycle 3's ground-tint / lane-seal attempt (REJECTED, reverted, tree manually reset). Numbers 40–49 are the smear-fix design questions the accepted diff deliberately did not answer; 50–52 are the ground-tint questions the rejected attempt raised before it was thrown away.

NEW 2026-08-24 — Concussion warp duty cycle — how often should a full-screen post-process fire at all? Concussion warp duty cycle: my bot-driven census (SimWorld + main.gd::demo_input, 1P, 4 seeds x 5400 ticks per mode, replaying the real arm/decay envelope) puts the shader rect visible for 17.9% of play in campaign AND arcade (71 knockdowns, 3856/21600 ticks) and 6.1% in endless (24 knockdowns, 1320 ticks). Shaping the falloff fixes the centre-readability defect but does not change how OFTEN the effect fires. Whether a fifth of frames carrying any full-screen post-process is the intended feel — or whether the 0.92 s envelope should also be shortened once the centre is clear — is a taste call, not a defect. Caveat that decides how much weight to give it: demo_input is open-loop, so 71 downs measures the bot's death rate, not a player's.
NEW 2026-08-24 — The vest-break warp is armed at a fixed 0.7 with no down_self_scale pass, unlike the knockdown warp. The vest-break warp (src/main.gd:4372) is armed at a fixed 0.7 with no down_self_scale pass, while the knockdown warp at :3155 IS scaled down to 0.35 when a squadmate is still standing. Intent question: should a co-op partner's vest break warp YOUR screen at full strength when their outright knockdown deliberately does not? The de-escalation comment at :3143-3145 argues the partner's knockdown is 'genuinely everyone's' feel — that reasoning points the opposite way for the smaller vest event, which suggests the omission is an oversight rather than a decision, but only the owner knows which envelope they meant.
NEW 2026-08-24 — CENTRE_KEEP_RATIO's target is undefined: "the centre is untouched" and "the centre is merely legible" are different floors. CENTRE_KEEP_RATIO floor value: the shader's master blend mix(clean, col, amt) (screen_fx.gdshader:67) still folds 70% of the processed image in at the probe's concussion 0.7, and the vignette multiplies the centre by ~~0.945 (vig=1, pulse~~0.9-1.0, 0.55 mix) independently of the three terms this fix ramps — so even a perfect distance ramp may not reach the proposed 0.85 centre-edge retention. Whether the target is 'the centre is untouched' (which would require ramping the master blend or the vignette too, changing the effect's character) or 'the centre is merely legible' (a floor set empirically from the post-fix measurement, likely well under 0.85) is a design call, not a defect. Pre-fix baseline for whoever decides: 408->68 strong edges, keep 0.167, in the 200-440 x 240-330 logical box that contains the soldier.
NEW 2026-08-24 — The probe's amplitude sample (0.7) is below the shipped solo-knockdown peak (1.0), and raising it fails the floor. The knockdown effect's intensity at concussion 0.7 is what the probe pins, but _concussion is set to down_scale which is 1.0 solo (src/main.gd:3155 via down_self_scale at :3728-3735). If the intent is that a solo knockdown is MORE readable than the probe's 0.7 sample, the probe's sample point is below the worst case actually shipped and should be raised to 1.0 — but raising it also makes any centre-retention floor harder to hit. Which of the two the ratchet should pin is an owner call.
NEW 2026-08-24 — Should the concussion warp follow the PLAYER or the SCREEN? This game has no horizontal camera, so they are not the same thing. Should the concussion warp follow the PLAYER or the SCREEN? This game has no horizontal camera (src/main.gd:6899 _to_screen = fx * PX), so the two are only the same when the soldier happens to stand mid-field. "Peripheral vision tunnels toward the centre of the SCREEN" and "the player stays legible while downed" are different effects and the diff conflates them. Measured cost of the current screen-centred choice: a downed soldier at x=60 keeps 0.04 of his strong edges at the shipped solo peak, i.e. no better than before the fix.
NEW 2026-08-24 — If the warp follows the player, should the VIGNETTE follow too? If the warp follows the player, does the VIGNETTE follow too? The diff deliberately couples them (float vig = 1.0 - periph;, sold as "one falloff, used twice"). A vignette whose dark ring slides across the frame as the player strafes is a different look from a fixed one; only the owner can say which is intended.
NEW 2026-08-24 — _blast_warp shares the concussion shader via maxf — one falloff currently serves two different intents. _blast_warp (src/main.gd:4252, 0.30 on marquee detonations) shares this shader with _concussion via maxf at src/main.gd:1468. For a knockdown you want the warp AWAY from the player; for a kill heat-shock you arguably want it centred ON the blast. One shader currently serves both with one falloff — split it, or accept that marquee kills lost their centre punch.
NEW 2026-08-24 — The concussion warp is now materially WEAKER overall, not merely redistributed. The concussion warp is now materially WEAKER overall, not merely redistributed: whole-world gradient retention goes 0.35 (pre-fix) -> 0.50 at an edge pose and -> 0.60 with the soldier at screen centre, because the spared ellipse is smoothstep(0.3, 0.75) in UV = 384x216 logical px fully unwarped out of 640x360. The far edge and the top strip are genuinely untouched (0.28 -> 0.28, 0.32 -> 0.32), so the redistribution claim is true at the margins, but a player standing mid-screen now gets a visibly gentler knockdown than the one that shipped. If the intended feel is 'ears ringing, world hostile', the owner may want the ramp's inner radius pulled in (0.3 -> ~0.2) to trade legibility back for punch — that is a taste call, not a defect, and the probe's 0.85 floor would still hold at 0.2.
NEW 2026-08-24 — The warp's apparent intensity now depends on WHERE the player is standing, because the spared ellipse moves with him. The warp's apparent intensity now depends on WHERE the player is standing, because the spared ellipse moves with him and the screen is only 640 wide. Measured: whole-world retention 0.50 at x=40 vs 0.60 at x=320 — a knockdown at the screen edge reads noticeably harsher than one dead centre, since more of the frame is far from the focus. Consistent within itself and arguably correct (you see your own corner clearly either way), but it means the same event has two different looks. Owner call whether that positional variance is wanted or whether the ramp should be normalised so total warp energy is pose-independent.
NEW 2026-08-24 — At edge poses the screen-centred vignette dims the soldier even though the smear now spares him. At edge poses the soldier sits inside the screen-centred vignette's dark ring and is dimmed to ~66-73% of clean luma (probe's vignette gain 1.37 at x=40 and x=600 vs 1.12 at x=320). He is legible — I looked, the crops are unambiguous — but 'spared from the smear' and 'spared from the darkening' are two different promises and only the first was made. If the intent is 'the downed soldier stays readable', the owner may want a small focus-anchored brightness lift to sit alongside the screen-anchored vignette; the diff deliberately did not do this, and its stated reason (a vignette chasing the player reads as a wandering blob) is sound.
NEW 2026-08-24 — World-ground tints (the lane-seal / trench-tint work) were self-scoped beyond the cycle's stated goal. Tell [2] (world-ground tints) was self-scoped by the plan and is not in the cycle's goal, which named only the label dodge and the deck gunner. It is the half that carries the regression. Whether it ships at all in this commit, or gets split out and redone with a zone-appropriate card, is a scoping call.
NEW 2026-08-24 — The lane seal's 0.92 fill and hard edge were an INTENTIONAL, commented choice — overturning it costs the three-state read. The lane seal's 0.92 fill and hard edge were an INTENTIONAL, commented choice ('so the cycle is learnable', 'art == collision'). Overturning the opacity costs the three-state read: measured interior contrast vs adjacent open ground went SEALED -50 -> -6, WARN +30 -> +12, SCAR -3 -> +2, so SEALED and SCAR are now 8 R-points apart. Whether 'no debug quad' outranks 'the blocked lane is obvious at a glance' is a design call only the owner can make.
NEW 2026-08-24 — With the dark base gone, the choke wall now reads as a boulder field only. With the dark base gone, the choke wall now reads as a boulder field only (measured slab-region luminance 56.6 -> 74.1 vs ~100 outside, and most of the remaining darkening is the rocks themselves). rock1/rock2 are also used as ordinary scatter litter elsewhere in the frame at smaller scale. Whether a boulder field alone communicates 'this flank is impassable' - a hard _choke_bounds collision - or needs a distinct wall material is an art-direction call.

2. Sim / gameplay defects — player-facing

RECOVERY (2026-08-24) — accepted work stranded off HEAD, not yet landed

Knockdown smear fix was ACCEPTED (gate perfect, closeness 96) but its commit died on API 529 — the blessed diff is stranded on branch rescue/wf122-smear-mix. Run wf_a7d3399b-122 cycle 1 (2026-08-24): gate verdict 'perfect: true, closeness: 96' on the peripheral-ramp screen_fx.gdshader fix — 'The knockdown smear is gone from the thing the player is looking at' — 5 tracked files, +444/-32 vs 21290ee. commit:1 and ledger:1 then died on API 529, cycle 3's revert also died, and the tree ended as a MIXTURE of this accepted work and cycle 3's REJECTED attempt-1 (lane-seal SEALED contrast -88%, trench tint -61% — do not cherry-pick blindly). Whole mixture preserved on branch rescue/wf122-smear-mix (be20537); the 4 scratch probes (probe_two/probe_ammo_econ/probe_bash_rate/probe_offscreen_fire) ride along there. RECOVERY: extract the smear fix from that branch (screen_fx.gdshader + its main.gd application + probe_concussion_hud.gd centre-readability ratchet + tests), re-gate it, and re-check the gate's four banked residuals before calling it done. Do NOT re-derive from scratch — the accepted diff exists.

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

Run wf_a7d3399b-122, three cycles, zero commits landed (API 529 killed cycle 1's commit and cycle 3's revert both; see the run-window note above §1). Items 80–98 below are cycle 1's behaviour-lens sweep and the concussion-warp measurements that came out of reviewing the stranded smear fix; 105–108 are cycle 3's behaviour-lens sweep.

TRIPLE SHOT / TRENCH GUN halve your gun at every range the game fights at — and you cannot take them off. WHERE: src/sim/sim_world.gd:1512-1543 (the fan spawn + its ammo surcharge at :1528-1529), constants SPREAD_COS/SPREAD_SIN at :38-41, SHOP_TRIPLE_COST at :450, CRATE_POOL at :460, the strip-on-death at :2239; src/main.gd:2604-2605 (the hint copy), :7058 (the celebratory callout), :6701 (always-fire). WHAT HAPPENS: the fan is a fixed +-12deg (and +-24deg for the 5-shot) spray. BULLET_HIT_RADIUS is 10px, so the off-axis pellets stop being able to reach a body dead ahead at 10/tan(12deg) = 47px. MEASURED (tools/probe_two.gd, one pull into a 60hp dummy, flight time excluded): at 16-47px triple lands 3 pellets for 2 rounds billed (1.5 hp/round, a real 50% gain); at 56px and beyond it lands exactly 1 pellet for 2 rounds, and triple+spread lands exactly 1 for 3 — measured identically at 56, 70, 100, 150 and 240px. That is 50% and 33% of the plain gun's ammo efficiency. Now put that against the roster's own standoff constants: RIFLEMAN_STANDOFF 100, ELITE_STANDOFF 120, DRONE_STANDOFF 130, GRENADIER_STANDOFF 150, SNIPER_STANDOFF 240. EVERY ranged enemy in the game stands beyond the 47px cliff — the nearest of them at 2.1x it. The only things that come inside 47px are a rusher reaching contact-kill range and a technical ramming you, i.e. the moments you are already dying. WHY IT OUTS THE GAME: this is a reward that punishes, and the game dresses it as the opposite. The Trench Gun is a free elite drop that pops a golden capsule and shouts 'SPREAD SHOT!' (main.gd:7058) — and for the next 480 ticks your DPS-per-round is halved at every range you are actually shooting from. Triple Shot is worse: it is the shop's most expensive item (SHOP_TRIPLE_COST 120, commented 'crate-only big ticket... above the airstrike'), it is PERMANENT — sim_world.gd:2239 is the only line that removes it, and it fires on death — and endless crates are bought on PROXIMITY with no confirm (_collect_pickups, :1958-2003), so walking within PICKUP_RADIUS (12px) of it debits _econ_scale(120) — 420 coins by wave 30 — from the shared chest that also funds revives, and permanently downgrades your gun. The only way to undo a purchase you did not want is to die. Compounding it: always-fire (main.gd:6701 sets p1.fire = true unconditionally, there is no trigger) means you cannot holster the surcharge — MEASURED (tools/probe_ammo_econ.gd) a full 99-round mag lasts 13.2s of wall clock at base, 6.6s on triple, 4.4s on triple+spread; and _respawn hands you MG_AMMO_MAX/2 = 49 rounds, which on triple is 24 pulls, 3.3 seconds. The hint copy ('TRIPLE SHOT — 3-ROUND FAN, 2 AMMO A PULL') discloses the PRICE and conceals the only thing that matters, which is that two of those three rounds cannot reach anything past 47px. WHAT A REAL STUDIO DOES: makes the fan's geometry answer the roster it ships. Either converge the pellets (fixed lateral offset in px rather than a fixed angle, so the fan is a real shotgun inside its useful band and collapses to the aimed line beyond it), or widen the pellet hit radius for fan pellets only, or — cheapest and most honest — bill the fan at 1 round and let it be a pure crowd-clear whose value falls off naturally, since the ammo surcharge is currently buying the player nothing at 9 out of every 10 shots they take. And a permanent, un-refusable, auto-purchased mod needs either a decline input on the crate or a way to unequip; 'die to remove your upgrade' is not a mechanic.
The two commonest shooters in the game fire aimed rounds from off-screen, and the anti-cheap-death pip system explicitly excludes them. Lens A (behaviour, cycle 1): WHERE: src/main.gd:13646-13661 (_draw_threat_pips, kind filter at :13660); src/sim/sim_world.gd:4420 and :4407 (field spawn row at camera_top - 24), :1951 (_clamp_actor pins the player at camera_top + 16), :153 (RIFLEMAN_STANDOFF 100), :145 (ELITE_STANDOFF 120), :3818-3819 and :3862 (the windup commit). WHAT HAPPENS: _draw_threat_pips exists, in its own docstring's words, so that 'a lethal shot from beyond the 640x360 viewport reads as a threat, not a cheap death'. Its kind filter is {sniper, grenadier, ghillie, drone, technical, mg_nest}. It does not include kind 'rusher' (the RIFLEMAN — the base infantryman, the single most numerous unit in the game) and it does not include the ELITE. Both of those lock an aim vector, telegraph it, and fire a lethal round. Meanwhile the field spawner drops every walker at camera_top - 24, i.e. 24px ABOVE the drawn viewport (the game renders exactly 640x360, project.godot:27-28), while _clamp_actor lets the player stand as high as camera_top + 16. Whenever the ratchet camera lags an advancing player — and it is speed-limited to MAX_CAM_STEP so it routinely does — that separation is as little as 40px, comfortably inside both the rifleman's 100px and the elite's 120px firing standoff. So the enemy commits and fires from a body the view never draws. MEASURED (tools/probe_offscreen_fire.gd, 4 seeds x 9000 campaign ticks, god_mode armed so the run cannot end): 131 of 1039 rifleman/elite windups — 12.6% — began at a position outside the drawn 640x360 frame. Direction breakdown: 131 of 131 above the top edge, worst case 24px above, 0 below, 0 off the sides. Per-seed off-screen windups were 39/26/38/28 against 233/226/232/217 on-screen, so this is not one seed's quirk. WHY IT OUTS THE GAME: this is difficulty from concealment, which the brief names directly — a round crosses the top edge with no source and no telegraph attached to anything the player can see, and one-hit-kill design means it is a death with no visible cause roughly one windup in eight. The game clearly KNOWS this is unacceptable: it built a whole clamped-edge-arrow subsystem for exactly this failure, gave it a colour-blind-safe hollow-vs-filled shape split for area-vs-aimed, and routed six archetypes through it. It then left out the two that produce the overwhelming majority of aimed fire. A player never reads that as a missing filter entry; they read it as the game cheating. WHAT A REAL STUDIO DOES: two independent fixes, either sufficient. View-only, zero golden churn, one line: add 'rusher' and 'elite' to the filter at main.gd:13660 so the same red filled arrow already drawn for a sniper is drawn for them. Sim-side and stronger: gate the fire-commit on visibility — refuse to start a windup while e["y"] < camera_top, the way _step_bunkers already gates spawning on the live band — so nothing the view does not draw is ever allowed to commit an aimed shot. The second changes checksums and needs a golden re-record; the first does not, and should ship regardless.

Lens B (behaviour, cycle 3, independently re-measured): WHERE: src/sim/sim_world.gd:4407 and :4420 (campaign spawner) and :6180-6220 (endless waves) all spawn walkers at camera_top - 24 * F_ONE; src/sim/sim_world.gd:3671 keeps every enemy alive and stepping out to camera_top + 420; the viewport is 640x360 (project.godot:27) and src/main.gd:6951 maps world->screen as (fx, fy - camera_top), so the only visible band is screen-y 0..360 — which is the game's OWN definition, used verbatim at src/main.gd:10042 to gate the first-sighting teaching card. Every ranged archetype then ROOTS itself for the whole windup: sim_world.gd:3872 return # rooted while winding up (elite), :3811 (rifleman), :3936 (sniper), :3962 return # holds position while painting (drone), :4772 (mg_nest). WHAT I SEE: measured over 4 seeds x 12000 ticks per mode, driving the sim directly. ENDLESS: 574 firing windups, 94 of them (16.4%) had the shooter outside screen-y [0,360] for EVERY SINGLE TICK of its own telegraph — elite 21/49 (42.9%), drone 9/34 (26.5%), sniper 7/36 (19.4%), rusher 49/254 (19.3%), grenadier 7/52 (13.5%). CAMPAIGN: 185/2083 (8.9%) fully blind — mg_nest 95/339 (28.0%), elite 31/317 (9.8%), rusher 49/899 (5.5%). Because the shooter is rooted, 'partially visible' measures at 0.0% in Endless and <=3.5% in Campaign: it is all-or-nothing. Tracking every enemy bullet from its muzzle to the tick it connects with a player, 19 of 119 campaign connects (16.0%) came from a muzzle that was never drawn (per-seed 17/63, 1/29, 1/17, 0/10 — skewed, but the mechanism is uniform). The game has a system built precisely for this: src/main.gd:13698 _draw_threat_pips — 'so a lethal shot from beyond the 640x360 viewport reads as a threat, not a cheap death'. Its kind filter at main.gd:13711-13713 is {sniper, grenadier, ghillie, drone, technical, mg_nest}. _spawn_enemy (sim_world.gd:4486) tags units "elite" and "rusher" — neither is in that list. So 70 of Endless's 94 blind telegraphs (74.5%) and 80 of Campaign's 185 (43.2%) get no pip at all. The fallback, _draw_threat_edges (main.gd:12837), does draw them, but puts elite and rusher in the non-danger tier (orange, 4px spread vs 6px red), caps the row at 6 threats, carries no aim lane, and parks the chevron at screen-y 28 for a body that is at screen-y -24. WHY IT OUTS IT: this is the exact death a player cannot learn from. You are hit by a bullet whose muzzle flash you never saw, down a lane that was drawn for 24 ticks on a rectangle of screen that does not exist. The game's whole design thesis is that a drawn line is a promise — it has a test suite for that — and then it draws the promise where you cannot read it. Note the codebase ALREADY diagnosed this for one class of unit: sim_world.gd:4405 says 'Rooted kinds never move, so the walk-in-from-the-top y is where they STAY while the camera is held — a lethal gun above the reachable field', and _rooted_spawn_y (:4655) was written to fix it. Nobody generalised the insight to the mobile shooters, which are equally rooted for the 20-55 ticks that matter. AAA VERSION: a shooter may not OPEN a windup unless its own position is inside the drawn band. That is a one-line guard sitting right next to the not _concealed(target) gate each archetype already has — the same shape as the concealment rule the game already enforces. Add the spawn row's own fix as a backstop: bring walkers in at a y inside the frame and fade them in with rooted_arrival_alpha, which main.gd:9930 already implements for exactly this 'READS as arriving rather than as having always been there' reason. As a floor, add "rusher" and "elite" to _draw_threat_pips' filter so the system that exists to prevent cheap off-screen deaths actually covers the units causing three quarters of them.
Deep Endless: the free empty-clip bash out-kills the paid machine gun from wave 31, so the shop's cheapest staple becomes a downgrade you pay for. WHERE: src/sim/sim_world.gd:1487-1506 (the bash), specifically the armour exemption at :1492-1493; :6044-6053 (_wave_armor); :26 (FIRE_COOLDOWN_TICKS 8) and :31 (BASH_COOLDOWN_TICKS 40); :2524-2543 (_supply_cost) and :2465-2489 (_econ_scale); the view's scolding cue at src/main.gd:2825-2835 and src/view/hud.gd:2773-2780. WHAT HAPPENS: from wave 13, _wave_armor gives ordinary infantry hp = 1 + (wave-13)/6 extra, growing without bound. The MG pays FIRE_COOLDOWN_TICKS per hp. The bash does not: its armour check at :1493 exempts every 'veteran bulk' kind whose hp comes only from wave armor, so one swing kills regardless. MEASURED (tools/probe_bash_rate.gd, wave machinery parked so nothing else spawns, fresh armoured body re-planted on each kill, 1200-tick window each): the bash returns a flat 30 kills per 1200 ticks at EVERY depth tested — wave 1, 13, 19, 25, 31, 37, 43, 50, 75, 100 — while the gun falls 149 -> 74 -> 49 -> 37 -> 29 -> 24 -> 21 -> 18 -> 12 -> 9. The crossover is wave 31 (hp 5, gun 29 vs bash 30) and the gap reaches 3.3x by wave 100. Direct armour-bypass check in the same probe: a wave-100 rusher at hp 16 dies to ONE bash tick with its hp field still reading 16 — untouched, not decremented. The economy inverts first and separately: 30 rounds of shop ammo costs _econ_scale(30), which is 90c at wave 25 and buys 30/4 = 7.5 bodies paying COIN_RUSHER 10 each = 75c back. Ammo goes coin-NEGATIVE at wave 25, six waves before it goes throughput-negative. By wave 100 it is 277c for 1.9 bodies and 19c back. WHY IT OUTS THE GAME: the bash requires mg_ammo == 0. So past wave 31 the optimal play is to be permanently broke of ammo and never buy it — and the game screams at you the whole time you are doing the right thing. Running dry fires a dry_fire event that plays a click_dry sound, a 'vo_clip_dry' voice line, a grey muzzle puff and a red CLICK (main.gd:2825-2835); the HUD flips the ammo stat to a warn colour and draws a draining cooldown ring on a DRY icon (hud.gd:2773-2780); the shop lists AMMO as its first and cheapest item. Every surface says empty = failure while the sim says empty = your best weapon. That is a rule contradicting what the game tells the player, and it is compounded by finding #1: Triple Shot doubles the burn rate, so the 120-coin premium upgrade's real function at depth is to get you into bash range faster. HONEST CAVEATS, stated because they bound the finding: the bash kills only ONE body per swing (it breaks), so this is a survival tool, not a clear — at wave 31 the endless spawn interval has floored at 8 ticks and a wave carries 4 + 2*wave = 66 bodies, so both weapons lose the race, the gun just loses it harder. The bash also pays no coin and no score (deliberately, _kill_enemy(e, true, true)), so at depth the player is forced to choose between killing things and scoring. And wave 31 is deep — the pacing data supplied with this task only reached wave 20 and the driver capped there, so I cannot show a player reaching the crossover, only that the crossover exists and that the coin inversion at wave 25 sits well inside plausible depth. WHAT A REAL STUDIO DOES: make the last-resort weapon respect the armour system it exists inside — let the bash strip hp like a bullet (it would still be free and still a real panic button, just no longer a strictly better gun), or scale BASH_COOLDOWN_TICKS with _wave_armor so its throughput tracks the curve everything else is on. And if being dry is going to be viable, stop having four separate surfaces punish the player for it.
The base infantryman's telegraph is 20 ticks — below the 24-tick reaction floor this project pins in five constants and asserts in three tests. WHERE: src/sim/sim_world.gd:155 (RIFLEMAN_WINDUP_TICKS := 20). The floor it violates: :83 TECHNICAL_REV_TICKS := 24 with the comment 'the 24t reaction floor (REAR_WARN/VENT_WARN precedent)'; :335 REAR_WARN_TICKS := 90 '(>= the 24t reaction floor)'; :566 VENT_WARN_TICKS := 30 '>= the 24t reaction floor'; :622 FLANK_WARN_TICKS := 45 '(> the 24t reaction floor)'; plus FROGMAN_SURFACE_TICKS. It is asserted as a rule in tests/test_mechanics.gd:581 ('the rev clears the 24-tick reaction floor the rest of the game pins'), tests/test_mechanics.gd:1896, and tests/test_biomes.gd:109. WHAT HAPPENS: MEASURED (tools/probe_two.gd) every archetype's tell against that floor: rifleman 20t, elite 24t, drone 24t, technical rev 24t, mg-nest aim 30t, frogman surface 30t, vent warn 30t, grenadier 40t, flank warn 45t, strike ring 45t, sniper/ghillie 55t, rear warn 90t. Exactly one archetype is under the line, and it is the one you meet most — kind 'rusher' is the game's default body, spawned by both the campaign field spawner and every endless wave. grep -rn RIFLEMAN_WINDUP tests/ returns nothing: no test pins it, which is why the floor the project enforces everywhere else never caught it here. WHY IT OUTS THE GAME: a project that writes its reaction floor into five constants and three test assertions has decided that below 24 ticks a telegraph stops being a promise. Shipping the most common enemy in the game 4 ticks under that line means the rule is aspirational rather than real, and it is the kind of inconsistency a player feels as 'sometimes I just get shot' without ever being able to name it. WHAT A REAL STUDIO DOES: raise RIFLEMAN_WINDUP_TICKS to 24 for parity with the elite (goldens move, so re-record per the documented procedure), and add the same floor assertion the technical and the frogman already carry so the next archetype cannot slip under it either. HONEST CAVEAT, and the reason this is ranked last rather than higher: the tell is not the player's whole budget. ENEMY_BULLET_SPEED is 3px/tick, so flight time is additive. I measured warn-event to hit on a stationary player at five ranges: 25 ticks at 20px, 31 at 40px, 38 at 60px, 45 at 80px, 51 at 100px (the standoff). Every total clears 24. And because the sim commits aim_lx/aim_ly at windup START and fires that stale vector, ENEMY_BULLET_HIT_RADIUS is 7px and PLAYER_SPEED is 2.4px/tick, so three ticks of any lateral movement clears the lane. Practically this shot is very dodgeable — the defect is that the project's own stated invariant is false, not that the rifleman is unfair. I looked for the severe version of this and it is not there; report it as a consistency and test-coverage gap, not as a difficulty problem.
Concussion warp is applied at full strength at screen centre (the residual the brief named, still unfixed). src/view/screen_fx.gdshader. Three of the four warp channels have no distance falloff, so they hit the exact screen centre — where the player soldier sits — as hard as the corners. Measured analytically on a 640x360 logical frame at amt=1.0: chroma split (:47) is normalize(to_center) * 0.0045 * amt, and normalize() discards dist, giving a CONSTANT +/-2.88 px x / +/-1.62 px y R/B separation at every pixel including dist=0; the wobble (:36, vec2(wob) * 0.0022 * amt, |wob| <= 2) is a constant +/-2.82 px x / +/-1.58 px y whole-image displacement; the blur fold (:53, mix(col, blur, 0.5)) applies a flat 50% weight everywhere, which is what carries the centre chroma smear into the output. Only the radial blur taps (:41) correctly fall to 0 px at centre. PREDATES THIS DIFF — the diff is empty and HEAD is the baseline 21290ee, so this is verbatim HEAD state, verified by reading the file in the target worktree. Fix is ~3 lines: a smoothstep(0.12, 0.5, dist) ramp multiplied into :36, :47 and the :53 fold weight.
_ev_vest_break arms the same 0.7 warp for 0.73 s and is exempt from the co-op de-escalation the knockdown path gets. src/main.gd:4372 _concussion = maxf(_concussion, 0.7) inside _ev_vest_break, paired with only _hitstop_frames = maxi(_hitstop_frames, 2) at :4368. Replaying main.gd's own envelope (hold while _hitstop_frames > 0 per :5989, then x0.9/tick to the 0.01 floor per :5990) gives 2 held + 41 multiplies + 1 zeroing tick = 44 ticks = 0.73 s at 70% intensity, on every vest break. Unlike the player_down site at :3155 this value is NOT passed through down_self_scale (src/main.gd:3728-3735), so a co-op partner's vest break warps your screen at full 0.7 while their knockdown would have been scaled to 0.35. PREDATES THIS DIFF (empty diff, HEAD == baseline). Not triggered by my census instrument — 0 vest_breaks across 12 bot runs — so its real-play frequency is unmeasured, but the code path is unconditional on the event.
Knockdown smear is unchanged for a player anywhere but screen centre. The peripheral ramp is keyed on distance from SCREEN centre, and src/main.gd:6897-6900 _to_screen() applies no horizontal camera — the player's world x is his screen x across the full 16..624px field (WORLD_RIGHT := 624 * F_ONE, src/sim/sim_world.gd:19). Measured with the shipped probe instrument, fix in place, soldier's sim x overridden immediately before capture and the box re-derived from his real screen position — player-box strong-edge retention at concussion 0.7 / 1.0: x=280 0.86/0.72, x=140 0.55/0.33, x=560 0.35/0.10, x=60 0.15/0.04. Pre-fix baseline at x=280 was 0.07/0.08, so at x=60 the fix buys nothing. Captured and viewed the x=60 / amt=1.0 frame: soldier, facing arrow, aim reticle and revive ring are a ghosted chroma-fringed mush. Reproduced 3/3 runs at x=60. NEW with this diff in the sense that the diff claims to have fixed it; the underlying smear predates the diff.
"The periphery is untouched (~0.98 at the corners)" overstates by a lot. Claimed in the src/view/screen_fx.gdshader comment block and echoed in the probe's const block. True only at the literal UV corner (dist 0.707); smoothstep(0.3, 0.75, dist) is 0.417 at a mid-edge (dist 0.5) and ~0.38 at the visible corner regions of the world band. Measured gradient-sum retention (concussion vs clean frame, |dI/dx| summed over the region) pre-fix -> post-fix: left edge (x 0-80, y 60-330) 0.407 -> 0.583; right edge (x 560-639) 0.403 -> 0.567; top strip (y 60-110) 0.432 -> 0.755; bottom strip (y 280-330) 0.449 -> 0.731; TL corner region 0.383 -> 0.557; BR corner region 0.405 -> 0.470. The blur is measurably weaker across the whole mid-field. The GRADE is genuinely untouched (BR corner mean luma 52.09 pre-fix vs 52.07 post-fix), so the claim is right about darkening and wrong about blur.
_blast_warp silently rebalanced by this diff, undocumented. src/main.gd:1468 is var amt := maxf(_concussion, _blast_warp) * maxf(_motion, 0.25) — the same shader serves the marquee-kill heat-shock (_blast_warp = maxf(_blast_warp, 0.30), src/main.gd:4252, decaying x0.86/tick at :5991). The peripheral ramp removes its centre punch too. No comment in the shader, no assertion in test_concussion_warp_spares_the_screen_centre, and no coverage in the probe mentions this second consumer.
Triple Shot — the most expensive thing you can buy — halves your magazine on a gun with no trigger, and the only way to remove it is to die. WHERE: src/sim/sim_world.gd:1526-1529 charges fan_cost (+1 ammo for a 3-fan, +2 for a 5-fan) on top of the base round, at the same 8-tick cadence; SHOP_TRIPLE_COST is 120 (sim_world.gd:450), the dearest item in CRATE_POOL_BASE and above the 100-coin airstrike. src/main.gd:4734 states it outright — 'ALWAYS-FIRE: there is NO "fire" binding any more' — and _gather_inputs sets p1.fire = true unconditionally at main.gd:6752. sim_world.gd:2230 strips triple only on respawn. WHAT I SEE: isolated burn test, invulnerable player, always-fire, no resupply, from a full 99: base 785 ticks (13.1 s), Triple 393 ticks (6.5 s), Trench/spread 393 ticks (6.5 s), Triple+Spread 257 ticks (4.3 s). Buying the 120-coin upgrade exactly halves the magazine; stacking the Trench Gun capsule on it cuts it to a third. In a live A/B — same seed, same input stream, mod granted at tick 600, 9000 ticks, 4 seeds — the BASE loadout spent 0.0% of the run dry on all four seeds with zero empty clicks. With Triple, seed 1 went to 8.0% dry including one CONTINUOUS 720-tick stretch (12.0 seconds) of holding a gun that clicks, 554 dry-fire events. With Triple+Spread all four seeds went dry: 0.9% / 1.2% / 3.2% / 3.8%, longest continuous 287 ticks (4.8 s). WHY IT OUTS IT: three things compound into a reward that punishes. (a) There is no trigger, so a player who notices they are burning out cannot ration — the drain is a metronome they have no input on. (b) triple is held until death, so the only mechanism in the game for removing an upgrade you can no longer feed is to get killed. (c) The base gun never ran dry at all, so ammo only becomes a resource AFTER you spend 120 coins; the 30-coin AMMO slot on the five-slot supply wheel is solving a problem that the default loadout does not have. The player's read is 'I bought the big one and now my gun stops working', which is the classic shape of a trap purchase — and the code comment at sim_world.gd:1521 anticipates it ('If Triple then feels punitive rather than powerful, drop the 5-fan to +1') without ever measuring it. Caveat I owe you: the driver kills 268-415 enemies per 9000 ticks and elite drops are its resupply, so a weaker player gets FEWER pickups — the dry share moves against the player, not toward them. AAA VERSION: either give the fan its pellets for one round of ammo and pay for the power somewhere the player can steer (a longer cadence, a tighter cone, a duration), or make the mod droppable — a hold-to-unequip, or an auto-revert below some ammo floor. A permanent upgrade whose only removal path is death is not an upgrade, it's a status effect. If the +1/+2 charge stays, the ammo pickup must scale with it: +30 rounds means 15 shots with Triple and 10 with Triple+Spread, so the resupply grammar the game teaches you in sector 1 quietly stops meaning what it meant.
The most-fired telegraph in the game is the one constant that sits below the project's own stated reaction floor. WHERE: src/sim/sim_world.gd:155 const RIFLEMAN_WINDUP_TICKS := 20. The project asserts a 24-tick reaction floor four separate times as a design invariant: :83 'TECHNICAL_REV_TICKS := 24 # the 24t reaction floor (REAR_WARN/VENT_WARN precedent)', :335 'REAR_WARN_TICKS := 90 ... (>= the 24t reaction floor)', :566 'VENT_WARN_TICKS := 30 # >= the 24t reaction floor (KIMK r4 precedent)', :622 'FLANK_WARN_TICKS := 45 ... (> the 24t reaction floor)'. It is also re-asserted as a fairness rule in _apply_supply's flashbang re-arm at :2377, which deliberately restores each frozen windup to 'ITS OWN archetype's full tell' rather than a flat 24. WHAT I SEE: every other archetype clears the floor — elite 24, drone 24, technical 24, mg_nest aim 30, grenadier 40, sniper/ghillie 55. The rifleman is the only one under it, at 20 ticks (333 ms), and it is by a wide margin the most common tell in the game: 899 of 2083 campaign windups (43%) and 254 of 574 endless windups (44%). WHY IT OUTS IT: this is the tell a player sees more than any other, so it is the one that teaches them what a telegraph is worth, and it is the only one the project's own rule says is too short. HONEST CAVEAT — I checked whether this is actually unfair and it mostly is not: the rifleman locks its vector at windup START (:3818-3819) and stands off at <=100px firing a 3px/tick round (:785), so the full budget from tell to impact is ~53 ticks, and the player moves 2.4px/tick. You can clear the lane. So I am filing this as a consistency defect, not a fairness one — but a 20 in a file that says 24 four times is exactly the kind of drift that a later pass reads as intentional and builds on. AAA VERSION: either raise it to 24 and re-record the goldens (the change is a 4-tick nudge to the opening field's cadence, and RIFLEMAN_FIRE_CD_TICKS 132 absorbs it), or write the exemption down at the constant — 'below the 24t floor on purpose: the locked lane plus 33t of bullet flight gives a 53t total budget' — so the floor stops reading as a rule with one silent violation. A number that contradicts its own documented invariant with no note is how a codebase starts lying to the next person who edits it.
Things I went hunting for that did NOT hold — reported so they are not re-investigated. WHERE / WHAT I SEE, five suspicions I tested and had to drop: (1) UNCONSUMED EVENTS. I enumerated all 100 {"t": "..."} event kinds emitted by src/sim/sim_world.gd and grepped every one against src/main.gd and src/view/*.gd. Zero unhandled. There is no beat the sim raises that the view silently eats. (2) ROLL INVULNERABILITY. The how-to (menu.gd:5244) promises 'ROLL to dodge — you can't be hit mid-roll'. I traced every call site of _hurt_player/_kill_player (sim_world.gd:1646, 2999, 3007, 3429, 4294, 4323, 4978, 6787, 7109, 7322) — every one that a rolling player can reach is guarded on p["roll_iframe"], including contact death, explosions, mines, vents, the mast, colossus, enemy bullets and mortar strikes. The two unguarded sites (2999, 3007) are the tank crew ring, which you cannot roll inside. The promise is kept, and sim_world.gd:3421 documents the one historical hole (a check keyed on roll_ticks == 0 instead of roll_iframe, off by exactly one tick) as already closed. (3) THE TECHNICAL'S REV LINE. _step_technical (sim_world.gd:4009-4014) writes aim_lx/aim_ly only when the windup REACHES zero, not at its start — so the stored vector is stale or absent for the whole 24-tick rev, and the docstring at main.gd:9918 is wrong to name the technical among shooters that 'commit aim_lx/aim_ly once'. But the VIEW never uses it: main.gd:10305-10312 draws the rev as a DASHED line off live facing with the comment 'dashed = still aiming, the solid charge corridor = committed', and switches to telegraph_dir only once lunge_ticks > 0. The player-facing telegraph is honest; only the comment is wrong. (4) DRY-FIRE AUDIO SPAM. The empty-mag branch (sim_world.gd:1450) has no cooldown, so it emits dry_fire EVERY tick — I measured 554 events in a 720-tick dry stretch. I expected a 60Hz click loop; main.gd:2841 throttles it to one per 14 frames. Held. (5) PILOT / TEACHING COVERAGE. The downed pilot pays 0 coin, 0 score and no streak when shot (sim_world.gd:3443-3450), does not block wave clear (:6316), and does not tick kills — matching the how-to's 'Not a HOSTILE, so the tally skips him'. All eleven specialist kinds have a _KIND_TEACH card (main.gd:388-401). WHY IT MATTERS: this codebase is unusually good at closing these exact holes and writing down why, which is what makes rank 1 stand out — it is the same class of bug (a lethal unit acting from where the player cannot see), already diagnosed and fixed at sim_world.gd:4405 for rooted units, and simply never generalised to the mobile shooters that are equally rooted for the 20-55 ticks their windup lasts. AAA VERSION: when a fix like _rooted_spawn_y lands, ask what OTHER state makes a unit effectively rooted — and put the invariant in the test suite (test_view_honesty.gd is already the right home) rather than in a comment on one constant.

New 2026-08-23 — from the `f1601b7..3be8149` window

The behaviour lens drove the sim headlessly with god_mode OFF for all of these — the first pass to do so since §6 established that _god_restore refills the magazine every 60 ticks against an 8-tick fire loop. Cited lines are the lens's; where the symbol moved on 3be8149 the HEAD line is given too.

The Riot Shield's advertised flank is kinematically impossible at every range the game fights at — and a green "safe" arc is drawn on his back promising otherwise. ⚠️ This is the second independent filing. A brief (docs/riot-shield-*, filed a day earlier) had already measured it; the behaviour lens re-measured from scratch rather than trusting the doc and confirmed it, and it is still unfixed on bf27b52/3be8149. Three live promises, all present on HEAD: the first-contact banner "shield": "RIOT SHIELD — FLANK OR BLOW IT OPEN" (src/main.gd:391, verified); the Field Manual line SHIELD — front eats bullets. Flank it, blast it, or use Rend. (src/view/menu.gd:5636, verified — cited as :5630); and a PERSISTENT green rear SAFE-arc drawn on his back every frame at src/main.gd:10331-10335 (verified — the comment there reads "Rear SAFE-arc: the shield only eats the front cone, so its back is the…", i.e. the art is explicitly teaching "get around him"). The behaviour is _turn_shield_toward (sim_world.gd:4596 on HEAD, cited as :4556) with SHIELD_TURN_STEP = F_ONE / 32 (:180, verified), whose own comment at :178-179 claims "a player circling at safe standoff can get outside the 120-degree block cone". _shield_blocks is sim_world.gd:4572 on HEAD (cited :4532-4553), a front 120° cone at dot < -0.5; the turn is ≈1.8°/tick (the sim says so itself at :3158). The flank never opens at ELITE_STANDOFF or GRENADIER_STANDOFF. The sim's own comment at :3117-3121 asserting that "a committed close-range lateral…" gets around him is part of the same false claim. (Fix scope note before anyone reaches for it: the honest levers are lowering SHIELD_TURN_STEP, adding a turn deadzone, or deleting all three promises — but the green arc is the loudest of them and the cheapest to make honest.)
Quitting an Endless run silently confiscates every Veteran Point it earned — while the same code path banks the wave count the VP is computed from. src/main.gd:5424 on HEAD (cited :5418) — vet_points += sim.wave is the ONLY line in the codebase that ever increases the currency (verified: grep -n 'vet_points +=' src/main.gd returns exactly one hit). It lives inside _record_run() (main.gd:5371 on HEAD, cited :5365), called from exactly one place, _apply_score_verdict() (main.gd:5353 on HEAD, cited :5347/:5357), itself called from exactly one place: the debrief transition, guarded by if sim.victory or sim.wiped or (_down_fr… (report text truncated at source). Quit-to-title never reaches it. So a player who quits a deep endless run keeps the leaderboard row and loses the meta-progression currency.
The airborne Recon Drone is shot down by knee-high rocks — the view sells altitude the sim does not have. _step_bullets (sim_world.gd:3043 on HEAD, cited :3003-3145) resolves in the order bunkers → sandbags → tank hulks → ROCKS → enemies, and a bullet that dies on cover never reaches the enemy scan. _step_drone (sim_world.gd:3948 on HEAD, cited :3908-3930) moves by writing e["x"] += … directly and never calls _advance_toward, so unlike every ground mover it is not stopped by cover — it flies over rocks it can then be blocked behind. The sim is inconsistent with itself in both directions at once: the drone ignores cover for movement and respects it for damage.
A maxed Veteran-Perk save posts to the shared Endless leaderboard with no marker, in a game that already tags ASSIST and NG-HARD. src/main.gd:1737-1744 applies owned perk tiers post-construction in _reset(): perk_level(PERK_VEST) > 0 forces pl["vest"] = true, sim.war_chest += PERK_CHEST_BONUS * perk_level(PERK_CHEST) (60 per tier, 3 tiers), and sim.tokens += perk_level(PERK_TOKEN) (1 per tier, 2 tiers). src/main.gd:5375-5382 builds the Hall-of-Fame / Steam entry and records "daily", "assist", "hard", "gr…" (report text truncated at source) — but nothing for perks. The board therefore flags the two toggles that lower or raise difficulty by choice and stays silent on the one that hands a run a free vest, up to 180 coin and 2 tokens at tick zero.
A quarter of the MG nest's rounds are fired from outside the drawn viewport, telegraph included. The sweep is sim_world.gd:3671 on HEAD (cited :3631) — if not e["alive"] or e["y"] > camera_top + 420 * F_ONE — while the drawn viewport ends at camera_top + 360 (project.godot:27-28, viewport 640x360; _to_screen maps screen y = world y − camera_top). That leaves a 60 px band below the bottom edge where rooted shooters stay alive, keep stepping, and keep firing. The nest's own tell, mg_nest_aim (cited :4750), is emitted at the nest's position, so it plays off-screen too, and its 30-tick MG_NEST_AIM_TICKS telegraph is drawn where nobody can see it. MEASURED by attributing every enemy_shot / sniper_fire event to the enemy standing at that exact position, 4 campaign seeds (report text truncated at source — the headline "a quarter" is the lens's own summary of that attribution). Note this is the same 60 px band §1 #29 measured as "identical to HEAD" for rooted spawn position; this entry is about firing, which nothing bounds.
The 120-coin Triple Shot triples your no-weapon time, and the surcharge that pays for it assumes a trigger the game deleted. The fan ammo surcharge is sim_world.gd:1528 on HEAD (cited :1527, verified byte-for-byte: var fan_cost := 2 if (p["spread_ticks"] > 0 and p["triple"]) else 1), justified in the comment at :1519-1525. The missing trigger is src/main.gd:6701 p1.fire = true (verified; :6765 for P2), whose own comment at main.gd:6685 reads "ALWAYS FIRE. There is no fire key, no fire trigger and no fire pad button" (verified verbatim). SHOP_TRIPLE_COST is 120 (sim_world.gd:450, verified) — above the 100-coin airstrike, the most expensive thing in the economy. MEASURED in isolation, one player, always-fire, no pickups, from a full 99: base 785 ticks (13.08 s, 7.57 rounds/s); Trench Gun alone OR Triple alone 393 ticks (6.55 s); both stacked 257 ticks (4.28 s). The surcharge is priced as if a player could choose not to spend the fan — and there is no such choice. (Cross-ref: 4d493b7 earlier stopped TRIPLE charging 2x ammo for zero extra damage; this is the remaining half — the charge is real now, and so is the fact that you cannot decline it.)

New this run (2026-08-21), measured unless noted

These come from the 390c12d..9bb1cdb window. The behaviour lens drove the sim headlessly for all of them; where an entry was reasoned from code instead, it says so.

The campaign has no fail state until its last 3%, then ends in under 5 seconds. WHERE: src/sim/sim_world.gd:1936 (_step_dead_player: if last_stand: return — no timer, no coin reader), :1955-1961 (the broke fallback: campaign always calls _respawn(p, _checkpoint_y()), only endless latches the wipe), :6433 (last_stand = true on colossus engage), :2030 (_try_revive returns immediately under last_stand). MEASURED, god_mode OFF, 4 seeds, demo_input driver: seed 0xC0FFEE Last Stand began at tick 6468 of a 6510-tick run (99%), 19 knockdowns before it, run over 42 ticks later. Seed 1: t=5957 of 6127 (97%), 20 before, dead in 170 ticks. Seed 2: t=6704 of 7016 (96%), 21 before, dead in 312 ticks.
Dying is the cheapest resupply in the game, and the gap TRIPLES as you go deeper. WHERE: src/sim/sim_world.gd:2117-2118 (_respawn hands back MG_AMMO_MAX / 2 = 49 rounds and 4 grenades), :997 (campaign revive_cost = REVIVE_BASE_COST * mini(p["deaths"], 3), halved solo — hard-capped, never scales with depth), :2369 (_econ_scale: every SHOP price is base + base * _econ_depth() / 4, +25% per gate opened, explicitly uncapped), :2074-2078 (solo self-revive stands you up where you FELL, not at the checkpoint), :2129 (plus VEST_IFRAME_TICKS = 90 ticks of mercy). MEASURED by stepping a real campaign and printing both curves at every depth change: depth 0 ammo 30 grenade 30 | … (report text truncated at source — the surviving figures are the constants and the depth-0 row.)
Endless shoots you from off the top of the screen — 30% of ranged-enemy time is spent outside the drawn viewport. WHERE: src/sim/sim_world.gd:5929-5972 (every endless spawn lands at camera_top - 24 * F_ONE), :1877-1885 (_clamp_actor pins the player's ceiling at camera_top + 16), :5900-5906 comment ("endless never runs _step_camera, so camera_top is pinned at -VIEW_H forever"), and the standoff branches that then never fire: :3665 (if dlen > RIFLEMAN_STANDOFF, 100px), :3727 (ELITE_STANDOFF 120px), :3758 (GRENADIER_STANDOFF 150px), :3785 (SNIPER_STANDOFF 240px), :3815 (DRONE_STANDOFF 130px). WHAT HAPPENS: in Endless the camera is frozen and the spawn line sits 24px ABOVE the viewport, while the player's own reachable ceiling is 16px BELOW it. (Report text truncated at source.)
The Claymore is a strictly dominated 50-coin crate sitting next to a 30-coin crate that does the same thing four times. WHERE: src/sim/sim_world.gd:459-461 (CRATE_POOL = [0,1,2,6,8], CRATE_POOL_BASE = [30, 30, 60, 120, 50] — the endless shop draws 3 of these 5 every intermission), :450 (SHOP_CLAYMORE_COST := 50), :446 (SHOP_GRENADE_COST := 30), :2260 (p["claymores"] = mini(CLAYMORE_CAP, p["claymores"] + 1) — ONE charge), :2247 (p["grenade_ammo"] + 4 — FOUR grenades), :4143-4147 (a tripped claymore calls _explode(m["x"], m["y"])), :3157-3160 (a landed grenade calls the SAME _explode). Both items detonate through the identical _explode with the identical BLAST_KILL_RADIUS of 30px (sim_world.gd:134). (Report text truncated at source at the "what differs" clause.)
The kill streak stops paying at 20 and the counter keeps climbing to 121, so the chain has no tension to hold. WHERE: src/sim/sim_world.gd:3337-3360 (_kill_enemy's streak block: tiers at 5 / 10 / 20 for 25% / 50% / 100%, and if kill_streak == 20 mints the Commendation and fires the surge), :264 (KILL_STREAK_WINDOW_TICKS = 90 — 1.5 seconds), :244 (SPAWN_INTERVAL_TICKS = 45). MEASURED, 4 seeds, god_mode on: best campaign kill streak = 121; best endless streak = 24. The last tier is at 20. So for 101 consecutive kills the HUD counted up while nothing about the reward changed. The window is the reason: 90 ticks against a field that spawns every 45 means the chain effectively cannot lapse during normal advance.
The BAIL OUT countdown is a race with no finish line — a tank cooking off on top of you does exactly nothing. WHERE: src/sim/sim_world.gd:2900-2909 (_detonate_tank → _explode(tank["x"], tank["y"])). _explode itself (:3188) scans enemies, sandbags, rocks, barrels, tanks, bunkers, the observer and every boss — and has no player scan at all. The crew ring at :2863-2872 only hits players whose in_tank == ti. Compare _detonate_barrel (:3296-3301) and the claymore/mine trigger (:4161-4166), which both explicitly hurt every exposed player inside GRENADE_RADIUS before calling the same _explode. The prompt the player is reading while this resolves is src/view/hud.gd:847, "BAIL OUT! %ds", over the tank_ignite alarm (src/main.gd:517). MEASURED directly (.aaa/c4/blast.gd) — a player standing on a cooking-off tank takes nothing.
The COURIER — the game's advertised 4x-bounty chase — runs off the BOTTOM of the arena in 1.1 seconds and is deleted with no feedback at all. Two lenses, kept together because the second measured it over more seeds. WHERE: src/sim/sim_world.gd:4432-4438 (_spawn_courier), :3556-3568 / :3574-3586 (the flee step and the courier_escape event), :3524-3526 / :3542 (the silent sweep at y > camera_top + 420). Player-facing copy: src/main.gd:390 "COURIER — 4x BOUNTY, GUN IT DOWN". The lost-it sting is wired at src/main.gd:575 ("courier_escape": ["deny", -5.0, 0.7]) and handled at :3286. WHAT HAPPENS: the flee vector is fx = -dx; fy = -dy - 40 * F_ONE — "run directly away from the nearest player, with a fixed 40px northward bias". The escape line is NORTH (y < camera_top - 30). But the courier spawns at camera_top + 300, and couriers only exist in ENDLESS (measured: 0 spawns across 4 campaign seeds x 20,000 ticks). MEASURED over 6 seeds (.aaa/c4/courier.gd): n=11 couriers, all in Endless, all leave by the bottom sweep, which emits nothing.
The gold BOUNTY +N¢ fountain fires with a real number on three kill paths that bank exactly zero coins. WHERE: src/sim/sim_world.gd:3312 emits {"t": "bounty_kill", ..., "coin": coin} unconditionally; the actual payment is five lines later at :3317, gated behind if not no_coin: war_chest += coin. The view is src/main.gd:2767-2770 — a five-coin gold fountain plus a distinct milestone fanfare, printing "BOUNTY +%d¢" straight off the event. Three kill paths pass no_coin = true: the airstrike screen-wipe (_fire_mission, :2233), the empty-magazine bash (:1434), and barrel/blast kills (:3189). All three still run the marked-target branch and still emit the toast with the full tripled figure. MEASURED on a marked elite killed three ways: the wipe printed "BOUNTY +75¢" for a +0 chest.
The 100-coin airstrike also silently confiscates every bounty it kills — its true cost is more than double the sticker price, and nothing on screen says so. WHERE: src/sim/sim_world.gd:2231-2233 (_fire_mission calls _kill_enemy with no_coin = true for every enemy on screen). Player-facing surfaces: src/main.gd:467 (wheel label "AIRSTRIKE", cost SHOP_AIRSTRIKE_COST), :471 ("AIRSTRIKE INBOUND"), :3337 (banner "AIRSTRIKE INBOUND — KEEP FIRING"), and the hint at :5396. None of them mentions coin. MEASURED against a 12-enemy screen — endless wave 10, strike price after _econ_scale = 150¢, chest delta after the wipe = +0, score delta = +579. Killing the same twelve by hand pays roughly 195¢ into the chest and ~1,950 score. So the real price of the button is 150 spent plus ~195 forfeited = ~345¢, and it hands back about 30% of the score.
The MG has no trigger, so the ammo counter is a 13-second stopwatch — and it runs out worst exactly where the game is hardest. WHERE: src/main.gd:6480 (p1.fire = true, unconditional) · src/main.gd:6513 (the ONE refusal: an open supply wheel, which also hijacks your aim vector) · src/sim/sim_world.gd:26 (FIRE_COOLDOWN_TICKS := 8) · :242 (MG_AMMO_MAX := 99) · :1445-1447 (the shot debits a round) · :1383-1440 (the empty-clip bash) · :30 (BASH_RADIUS := 16) · :31 (BASH_COOLDOWN_TICKS := 40) · src/view/menu.gd:5358-5360 ("The weapon has no trigger — it fires on its own"). There is no fire key, no pad button, no trigger: _gather_inputs writes p1.fire = true every tick a run is live, so with ammo in the gun the sim runs a permanent 8-tick fire loop. 99 rounds ÷ 8 ticks = 13.2 s. The only way to stop firing is to hold the supply wheel.
Endless's "threat-free" shop breather quietly spends half your magazine, and the shop sells it back to you. WHERE: src/sim/sim_world.gd:4769-4775 and :6982 (the sim's own comments: "The intermission shop is sold as threat-free (the breather grade, the wheel scrim, the airstrike refusal all say so)" — and it sleeps the mast hazard and the Spotter to keep that promise) · :5868-5872 (_intermission_len: 300 ticks at wave 1 down to a 120-tick floor; the comment even says "a flat 5s of standing in an empty arena is just noise") · :445 (SHOP_AMMO_COST := 30) · :444 (READY_HOLD_TICKS := 20) · src/main.gd:6480 (nothing sleeps the player's trigger) · src/main.gd:8668 (view: "the shop is sold threat-free, so a warn ring drawn during the buy would lie"). The game goes to real trouble to make the window safe and then lets the permanent fire loop drain the magazine through it.
ARCADE's deepest chapter is a 4.8-second walk into a no-revive Last Stand with an empty War Chest — the mode's own menu copy promises a run. WHERE: src/view/menu.gd:5134 (["mi_play", "ARCADE", "Choose a starting zone; play to the finale."]) · src/sim/sim_world.gd:953-973 (jump_to_chapter — it moves the camera, the player and every streaming cursor, and touches NOTHING the player carries) · :6495 (last_stand = true the instant the final gate scrolls into view) · :2032-2033 (_try_revive: if last_stand: return — the coin reader is dead, silently) · :1059 (all players down under last_stand latches wiped) · :2418-2437 (_supply_cost: prices creep with _econ_depth, which a chapter jump correctly primes). CHAPTER SELECT relocates you to the mouth of the chosen zone with the tick-0 loadout: 0¢ War Chest, 99 … (report text truncated at source.)
An Endless wave can be held open forever by a ghillie that the game's own anti-stall has frozen into a target dummy. WHERE: src/sim/sim_world.gd:6140-6163 (the all_cloaked force-reveal in _step_waves), :4028-4029 (_step_ghillie's range re-cloak), :6172 (_wave_hostiles_cleared), step order at :1092-1094 (_step_enemies runs BEFORE _step_waves). A ghillie whose nearest player is beyond GHILLIE_NOTICE_RADIUS (210px) hits line 4029 and sets submerged = true. Later in the SAME tick, _step_waves sees that every remaining hostile is a cloaked ghillie and force-reveals it: submerged = false, surface_ticks = GHILLIE_REVEAL_TICKS (26). For the next 26 ticks _step_ghillie returns at its surface_ticks > 0 early-out. (Report text truncated at source.)
Riflemen and elites shoot you from above the top edge with no committed-shot warning — the six other aimed archetypes all get one. WHERE: src/main.gd:13195 (_draw_threat_pips()); the kind filter at :13208-13210 admits only sniper, grenadier, ghillie, drone, technical, mg_nest. rusher and elite — the two most numerous shooters in the game — are excluded. Their windups are emitted at src/sim/sim_world.gd:3689 (rifleman_windup) and :3751 (elite_windup); nothing clamps an enemy's y back into the visible band, and enemies spawn at camera_top - 24 (:4284) while the player is clamped to camera_top + 16 (:1896). A rifleman spawning 24px above the top edge, with the player hugging the band ceiling, is already inside RIFLEMAN_STANDOFF (100px).
The trench is tall grass without the anti-camp counter that was added because tall grass was too strong. WHERE: src/sim/sim_world.gd:1647-1655 _concealed() returns true for smoke OR _in_grass() OR _in_trench() — one predicate, three sources, hard-gating every aimed shooter in the game (call sites at :3685 rifleman, :3747 elite, :3811 sniper, :3906 technical charge, :4031 ghillie paint, :6589 colossus spray, :6789 gunship). :4776-4802 _step_grass_flush() is the counter — a telegraphed flush grenade every FLUSH_CD_TICKS (600) while an enemy is within FLUSH_RADIUS (100px) — and its guard at :4785 is not _in_grass(p). The trench never enters it. Trench geometry: :4854-4869, one hash-placed 120x48px ditch per band from COVER_VARIETY_SEG (2) on.
Priced ground crates spend the shared War Chest on proximity — no button, no confirm, no way to decline. WHERE: src/sim/sim_world.gd:1903-1948 _collect_pickups(), called unconditionally from _step_players every tick. Line 1912 skips a crate only when the chest cannot afford it; line 1929 war_chest -= cost fires on distance alone (PICKUP_RADIUS, or TANK_CRUSH_RADIUS while riding). There is no interact gate anywhere on the path. The endless shop plants three of these at fixed x 190/350/510 every intermission (:6123-6127). In 2P the chest is shared, so one player crossing a 150-coin vest crate spends the coin the other player was saving.
The endless revive price compounds on lifetime deaths while every other price creeps on depth — past a point, coin can only buy gear, and the advertised choice is gone. WHERE: src/sim/sim_world.gd:1000-1017 revive_cost() — endless is `REVIVE_BASE_COST * maxi(deaths,
1. - (1 + wave / 5), **uncapped in BOTH factors**, halved solo. Supply prices scale on depth only, via _econ_scale (:2371) and _supply_cost (:2430), and the vest is explicitly capped at **120 base** (:2447). MEASURED (.aaa/probe_economy.gd`, endless, solo, god mode so the run survives): seed 0xC0FFEE — wave 12 chest 4,105 / revive 900; wave 15 chest 2,115 / revive 2,200 (unaffordable); wave 16 chest 3,165 / revive 4,400. Seed 2 — wave 16 chest 2,385 / revive 4,700; wave 20 chest 7,520 / revive 7,250, against an ammo price of 75 and a vest price of 150.
_cover_blocked never checks enemies — arena drops can plant geometry on a LIVE rooted unit (pre-existing, plan-banked, latent not live). src/sim/sim_world.gd:6137-6162 _cover_blocked(bx, by, recycle) — the 20px dedupe shared by the every-3rd-wave L-drop and the wave-5 supply pod — scans sandbags and rocks only. It never looks at enemies. The spawn seam 4d704cc adds guards BIRTH; it cannot guard geometry arriving later. Verified by reading the function (identical on HEAD — the diff does not touch it). The plan measured the only ARENA_L_SLOTS cell that reaches the rooted row, [320,-312], and a kind-0 rock centred at y=-312 with a unit at y=-308 is south-killable at tick 33 — containment WITHOUT a trap. It becomes live the moment a slot, a rock kind, or the rooted row moves.
Player-built sandbag nests have no enemy check — a player can re-create the in-cover trap by hand. src/sim/sim_world.gd:2283-2309, the supply-wheel nest plant: sbx/sby come straight off p["x"/"y"] + CLAYMORE_PLANT_OFFSET * aim, and the four sandbags.append(...) calls run with no test against enemies. A sandbag is SANDBAG_HALF_W = 18px half-width — WIDER than a kind-0 rock's 16 — against a ghillie's 10px BULLET_HIT_RADIUS (:224, no KIND_HIT_RADIUS entry), which is exactly the arithmetic that makes the rock case a 13px trap window. Verified by reading the plant path (identical on HEAD, untouched by 4d704cc). NOT measured for real-run reachability — it needs the player to plant on a surfaced ghillie.
In ENDLESS _step_camera is never called, so camera_top is pinned for the whole run and NOTHING is ever culled by the camera_top + 420 sweep. Rooted units born at row 287 therefore persist at row 287 for the entire run. Fine today (287 < the 360 px drawn band), but it means ROOTED_SPAWN_Y_SPAN is load-bearing in a way the campaign path is not: raise it past ~344 and endless gains permanent un-killable-because-off-screen turrets with no cull to save it. Worth a one-line note or a static assert if the span is ever tuned. (Reasoned from the step order plus the measured span; the ~344 threshold is arithmetic, not driven.)

Carried from the 2026-07-31 snapshot — not re-verified this run, measured unless noted

2P ready-up is a unanimous party vote with no tally — holding E alone does nothing, forever, with zero on-screen explanation. Where: rule at src/sim/sim_world.gd:5405-5418 (_ready_up — "True while EVERY living player holds REVIVE and nobody is down"), counter at :5439-5445 (ready_hold resets to 0 any tick the unanimity breaks; 20 ticks to deploy, READY_HOLD_TICKS at :402), hint at src/main.gd:4913-4915 (HOLD [%s] TO DEPLOY EARLY). MEASURED: headless 2P endless probe — reached a real intermission (300t window), had P1 hold revive alone for 180 ticks: intermission ran 300→120, ready_hold never left 0 (P1's entire 3-second hold was discarded every tick). Both players then holding: deployed in exactly 20 ticks. And ready_hold has ZERO view reads (grep over src/main.gd + src/view/hud.gd — none). (Report text truncated at source.)
Daily Run teaches "one attempt" — but the lock only arms at the debrief, so QUIT TO TITLE / RESTART makes today's seed infinitely scoutable. Where: taught rule at src/view/menu.gd:5136 ("One shared seed a day, one attempt."); the lock's only write at src/main.gd:4753-4758 (_daily_done_seed = _current_seed inside _record_run); the only caller chain is _apply_score_verdict (src/main.gd:4709) ← the debrief trigger at src/main.gd:4928-4943, which fires on sim.victory or sim.wiped or last-stand-all-down. The exit paths — pause menu "QUIT TO TITLE" (src/view/menu.gd:3185-3188) and "RESTART" (:3183-3184) — call main._reset() directly, which runs _flush_bests() (src/main.gd:1855) that touches only best/hint sections, never the daily lock. (d45132a shipped the arm+demote behavior this run — owner decisions #13/#14. The lens still flags the scouting remainder; the open item is what survived d45132a.)
DAILY RUN's "one attempt" is only enforced at the debrief — R/RESTART before the wipe retries the same seed, unlocked, forever. Where: copy at src/view/menu.gd:5136; the lock write at src/main.gd:4753-4759 (_daily_done_seed written ONLY in _record_run); the restart path at src/main.gd:1434-1452 (_reset() → _flush_bests() persists only best/seen — the daily seed is untouched) and :1770-1776 (R key → _reset()). MEASURED live against the real main scene headlessly: start_daily() → seed 1017050458, daily_done()==false; play 240 real ticks; call _reset() (the R-key path) → same seed 1017050458, daily_done()==false, _daily still true — unlimited retries. (Near-duplicate of the entry above from a second lens — kept because its measurement is independent. d45132a shipped; this is the flagged remainder.)
**NG+ HARD scores ~1.58x a normal campaign and lands on the same Hall of Fame and Steam leaderboard with no marker — the board flags ASSIST and DAILY but not the toggle that inflates score. Where: src/main.gd:1459 (sim.hard = _hard and not _endless and ... — campaign only, and _hard is a free RUN SETUP toggle at src/view/menu.gd:3090-3092 with no unlock gate); the board entry at src/main.gd:4726-4733 (_record_run tags assist and daily — no hard key); the Steam upload at :4787+ (_steam.upload_score(sim.mode, score) — 'campaign' for both rulesets). MEASURED: same scripted bot, same 4 seeds, god mode, 12,000 ticks each: normal campaign scores [131775, 104669, 142093, 151981] (mean ~132.6k, 1136 kills) vs hard [265621, 172361, 206247, 194341] (mean ~209.6k, 1934 kills) — +58% score, +70% kills, because harder spawning is more income (elites pay 25c vs 10c). (Report text truncated at source.)
Grenade-family hits on bosses emit no boss_hit feedback event (bullets do). _explode's campaign-gate branch (sim_world.gd:3002-3005) and endless branch (:3012-3016) call _damage_boss without emitting boss_hit, while _bullet_hits_boss (~:6300) emits it — so grenade/barrel/mine hits on a boss produce none of the hit-flash/hit-stop feedback bullets get. Confirmed pre-existing by reading HEAD: the fly-in fix only added the phase_t conjunct and left the event asymmetry untouched (the plan's brief explicitly named it out of scope). Fix shape: emit the same boss_hit event from _damage_boss's non-lethal arm, or from both _explode boss branches, and extend the fly-in test's post-arrival arm to assert the event. (Reasoned from code, not driven.)
A salvaged or expired tank hulk keeps burning and drawing as cover while bullets fly straight through it. Where: sim cover predicate src/sim/sim_world.gd:2750-2756 (player bullets) and :6443-6449 (enemy bullets) — both not hk["alive"] and hk["burn_ticks"] > 0, lifetime HULK_TICKS=1050 (17.5s) at :445, instant-strip at :2405 (salvage sets burn_ticks=0). View: _hulks pool src/main.gd:5409-5416 (appended on death, persists to an 8-cap eviction at :5435), wreck sprite + scorch drawn unconditionally at :10702-10715, and the big flame envelope at :10531-10539 keyed on view-time h["t"] (~8s fade), which NOTHING in the sim touches. What happens: the game teaches "dead tanks are cover while they smolder" — but the visible fire runs on a view-side 8s clock, the sim's cover runs on a 17.5s sim clock, and … (report text truncated at source)
Salvaging a hulk at full grenades strips your cover for +0 ammo — the one supply path that still bills for a no-op. Where: src/sim/sim_world.gd:2397-2408 — _try_salvage_hulk clamps p["grenade_ammo"] to GRENADE_AMMO_MAX and unconditionally sets tank["burn_ticks"] = 0 (ends the cover), even when the clamp grants +0; the view receipt at src/main.gd:2793-2795 ("FULL UP — COVER STRIPPED"). The rule it breaks is written in the sim's own _supply_full docstring at :2239-2241: "shared by every path that hands out a supply so none of them can bill for a no-op". MEASURED: headless probe (tools/probe_salvage.gd) — player at 12/12 grenades interacts with a smoldering hulk: salvage returns true, grenades 12 → 12 (event grants n=0), hulk burn_ticks 100 → 0. The cover cost is real: dead hulks block bullets both ways. (Fix direction is owner decision #9.)
The Colossus siege restocks grenades in total silence — the one fight balanced on grenade income never announces the drop. Where: src/sim/sim_world.gd:6131-6136 — the siege drop block in _step_colossus does a bare pickups.append (kind 1, no event, no drop field); compare the endless wave drop at :5857-5861 which fires supply_drop → cargo chime (src/main.gd:462), "SUPPLY DROP — HOLD IT" toast + ground light (:2845-2849) and the parachute/TTL drawing (:8315+, keyed on the drop field the siege pack lacks). MEASURED headlessly (tools/probe_colossus.gd, 4 seeds, god_mode, arcade jump to FINAL_GATE_INDEX): the finale runs 663-899 ticks with 3-5 downs and 2 core windows (healthy pressure — no tuning claim), and spawns 3-4 grenade packs per fight with ZERO announce; the sim comment at :6131 says these drops "keep the g…" (report text truncated at source)
The shop bills full price for air it partially delivers: a "+30 AMMO" buy at 90/99 grants 9 rounds for the full 30 coins, "+4 GRENADES" at 10/12 grants 2 — and the label, the receipt floattext, and the token-drop callout all still print the full amount. Where: src/sim/sim_world.gd:2093-2095 (_apply_supply caps with mini(MG_AMMO_MAX, +30) / mini(GRENADE_AMMO_MAX, +4)), :2239-2247 (_supply_full — the no-op guard shipped in e57191b — only returns true AT the cap, never near it), :2300-2342 (_try_buy debits the full _supply_cost regardless of headroom), :2262-2298 (_try_token_drop filters on the same hard-cap-only test); view side: static labels "AMMO +30" / "GRENADES +4" at src/main.gd:419-421, full-amount receipts BUY_FLOAT "+30 AMMO" / "+4 GRENADES" at src/main.gd:426 printed on every buy (:2384) and token drop (:2792). MEASURED by stepping the sim headlessly: ammo buy at 90/99 → now 99 (delivered 9), charged 30; grenade buy at 10/12 → del… (report text truncated at source)
The revive's whole go-to-the-body grammar — body beacon, dashed tether, off-screen chevron "so the revive has a spatial target" — signposts a dangerous rescue run the sim does not require: E revives from ANY distance and teleports the partner to your row. Where: src/sim/sim_world.gd:1928-1962 (_try_revive — loops all dead players, checks only war_chest >= cost; no distance test anywhere; the revived player lands at reviver["y"]) vs the view's spatial rescue kit at src/main.gd:10139-10148 (rising beacon "pulls their eye to the body"), :10899-10915 (off-screen chevron whose comment says the cue exists "so the revive has a spatial target"), :9815-9818 (off-screen partner chevron). MEASURED: 2P sim, killed P2, walked P1 300px north, called _try_revive(0, p1) — P2 alive:true, teleported ~284px to P1's row, chest debited 50. The game invests three separate rendering systems in telling the player WHERE the body is and building the fi… (report text truncated at source)
A tank parked directly on a free supply crate collects nothing and says nothing — the biggest object on the field rolls over a glowing pickup with zero effect or feedback. Where: src/sim/sim_world.gd:1107-1109 (_step_players: the in_tank branch calls _drive_tank and continues before _collect_pickups at :1454-1456, so riders and gunners can never collect), vs the gate-cache crate the sim plants dead-center in the only path north at :4246-4248 (x = SCREEN_CX). MEASURED: boarded a tank directly onto a free grenade crate at 0 grenades, stepped 10 ticks — grenades still 0, crate still on the ground, no deny event, no cue. The tank's tread grammar already touches everything else it rolls over — it crushes infantry, flattens sandbags, detonates barrels, and even RESCUES the pilot on contact — so "drove over the crate, nothing happened" reads as a bu… (report text truncated at source)
Closed gates pin the camera — and the spawner keeps planting rooted MG nests 24px above the viewport, so 4-6 invisible, unflankable turrets fire lead-computed bursts into the game's longest fights (including the no-revive finale). Where: src/sim/sim_world.gd:3969-3971 — _step_spawner plants rooted units at camera_top - 24 * F_ONE (_spawn_mg_nest for SECTOR_SPECIALS sectors 3 and 6, _spawn_broadcast for sector 5). The player's northern clamp is camera_top + 16 (_clamp_actor, :1784) and a closed gate pins the camera at g["y"] - GATE_CAMERA_PAD (:4624), so while any gate fight lasts, a rooted spawn sits 24px above the top edge and 40px above the furthest point the player can ever stand — for the whole fight. MEASURED (6 campaign seeds via .aaa/probe_offscreen_nests.gd + .aaa/probe_offscreen_attr.gd): peaks of 4-6 rooted nests/broadcasts piled up above the pinned edge during stage-3 (gunship) gate … (report text truncated at source)
Endless 2P: a broke death respawns FREE after 5s whenever a partner is up — the mode's only death brake (the compounding revive price) is waived exactly when you can't pay, so spend-to-zero is the dominant strategy. Where: src/sim/sim_world.gd:1863-1870 — _step_dead_player's broke-timer expiry calls _respawn(p, _checkpoint_y()) (free) whenever rally_is_free() (:1913-1921), which is true in endless the moment any partner is alive; only a full-party down latches the wipe. MEASURED (.aaa/probe_broke_respawn.gd): endless 2P, wave 10, revive_cost 900 vs chest 10 — the downed player respawns at exactly tick 300 with cost=0, a fresh 49-round clip and 4 grenades. Dying solvent costs the full compounding price (50 × deaths × (1+wave/5), uncapped — 4600 by wave 16 in the economy probe); dying broke costs 5 seconds and pays a partial restock worth ~45+ coins at wave-10 shop prices. (Report text truncated at source.)
"WAVE CLEARED — SHOP OPEN" fires while a live Spotter is still shelling the shop window. Where: src/sim/sim_world.gd — _wave_hostiles_cleared (:5670) iterates only enemies; the Mortar Observer is a separate top-level dict spawned on SPOTTER waves (:5836) and exempted from every despawn in endless ("the observer living until it is shot is the documented, intended pressure", :6562-6566); its strike loop (:6575-6585) runs every tick regardless of intermission; the wave-clear + Clean Wave bonus evaluate at :5583-5600; deaths_this_wave resets at _start_wave (:5712). View: "WAVE CLEARED — SHOP OPEN" banner (src/main.gd:2946). MEASURED (.aaa/probe_c6_spotter.gd, seed 7): wave 3 rolls SPOTTER, observer up; wave-clear fired with observer_alive=true; 276 shop ticks followed … (report text truncated at source. 390c12d shipped "the endless milestone shop window is a kill zone the game itself calls threat-free" this run — the lens still flags the Spotter remainder; the open item is what survived 390c12d.)
Miniboss fly-in airstrike window remains a partial whiff. The endless intermission is now gated because enemies is provably empty for the whole 45t telegraph (_step_waves early-returns; strike resolves before _start_wave) — that half shipped in 8425bd7. The wave-5+ miniboss fly-in is a second window where _fire_mission (:2071) spares bosses, but the trickle keeps spawning so it is NOT provably empty — the plan explicitly banked it. A strike bought during the fly-in still kills only trickle infantry. (Reasoned, not driven — banked by the plan.)
A tank driver's (or gunner's) revive key is a fully swallowed input the HUD is actively prompting — the arbitration mutes the cannon for a rescue the sim never performs. Where: src/sim/sim_world.gd:1114-1117 (if p["in_tank"] >= 0: _drive_tank(...); continue) — the continue skips the only alive-player revive read at :1402 (if inp.revive: _try_revive(i, p)), and neither _drive_tank (:2446) nor _ride_as_gunner (:2549) ever reads inp.revive. The arbitration that sends the key there is main.gd:5897-5909 (revive_context rule 3: "YOU ARE UP AND A PARTNER IS DOWN -> revive") and main.gd:5983-5993 (shared_e mutes the grenade route when the context is live). MEASURED (tools/probe_tank_revive.gd, headless, steps the sim directly): 2P campaign, P1 boarded in a live tank, P2 downed, chest 500 vs cost 50 — revive_… (report text truncated at source)
Grenades and every blast kill the cloaked, bullet-immune ghillie — the sim's own "only the reveal window can kill it" rule, its HOWTO card, and the airstrike's submerged exemption all say otherwise. Where: src/sim/sim_world.gd:2961-2974 (_explode's enemy scan: if e["alive"] and e["kind"] != "pilot" and _dist_lte(..., BLAST_KILL_RADIUS) — no submerged exemption), vs the same file's _fire_mission at :2071-2087 which deliberately spares e.get("submerged", false) ("Spares the submerged (1986 rule)"), vs the archetype's own contract at :3717-3722 ("Killing it during the reveal/paint window defuses the shot — and that window is the only time it can be killed at all"), vs the HOWTO card src/view/menu.gd:5262 ("GHILLIE — hidden sniper; only its laser gives it away. Close in."). MEASURED (tools/probe_ghillie_blast.gd, headless): a cloaked (submerged=true) ghillie 20px from the … (report text truncated at source)
The supply receipt lies on partial stocks: wheel buy and SUPPLY CALL print "+30 AMMO" / "+4 GRENADES" while the sim clamps the grant to as little as +1 — at full price. Where: src/sim/sim_world.gd:2098-2101 (_apply_supply clamps: mg_ammo = mini(99, +30), grenade_ammo = mini(12, +4)), :2310-2360 (_try_buy charges the FULL _supply_cost and emits {"t":"buy","kind":kind} with no granted amount), :2283-2309 (_try_token_drop spends the Commendation and emits the same amount-less token_drop event), src/main.gd:429 (BUY_FLOAT = ["+30 AMMO", "+4 GRENADES", ...]), src/main.gd:2419-2421 (the buy floattext prints BUY_FLOAT[kind] verbatim), src/main.gd:2828-2830 (SUPPLY CALL prints "SUPPLY CALL — " + BUY_FLOAT[kind]). The _supply_full guard only denies at the hard cap (99 / 12), so the whole partial window is live: ammo 70-98, gre… (report text truncated at source. Near-duplicate of the full-price-partial-delivery entry above from a second lens — kept because it pins the receipt-lie half, the other pins the charge half.)
Deep-endless veteran armor silently kills the empty-clip bash — the game's only free defense — exactly in the depth band that needs it. Where: src/sim/sim_world.gd:1335-1339 (the bash routes through e["hp"] exactly like a bullet: hp>1 bodies eat the swing as an armor_block chip) × :5427-5436 (_wave_armor gives EVERY rusher/elite/special +1 hp from wave 13, +1 more every 6 waves) × :1298-1299 (the bash's own contract: "running dry is a beat of danger, not pure helplessness"). MEASURED (controlled ring, tools probe stepping the sim: 12 identical rushers closing on a dry player holding fire, god mode, only hp varied): hp=1 → 8 kills / 8 swings / 3 downs in 2400t; hp=2 → 8 kills / 16 swings / 5 downs; hp=3 → 8 kills / 24 swings / 8 downs. Swings-per-kill scales 1:1 with hp — the melee's lethality is 1/hp, i.e. a bullet's. (f6666b8 shipped the invisibility half this run — banner/chip for the armor term. The bash-consequence above is the open remainder.)
Colossus closed core: MG tracers phase through the final boss with zero contact feedback for ~73% of the fight. Where: src/sim/sim_world.gd:2862-2867 — a player bullet only dies on the colossus while core_open > 0; the conjunct failing leaves dead=false, so the round flies on THROUGH the boss body and dies at range. Core timing (:639-640): 240t closed / 90t open per 330t cycle = closed ~73% of the finale. MEASURED (colossus stood up directly, player parked south firing north, god mode): closed core — 72 rounds fired, 0 boss_hit, 0 armor_block, 59 rounds observed PAST the boss's position, hp unmoved; open core — 27 fired, 18 boss_hit, hp 60 → 6 (the chip path itself works). The run-long grammar for armor is a VISIBLE ricochet — bunkers ping armor_block, shields ping it, wave-13 veterans ping it … (report text truncated at source)
Endless miniboss is drawn and shootable for a 7s fly-in while bullets pass through it unanswered. Where: sim guard at src/sim/sim_world.gd:6446 (endless_boss["phase_t"] >= 0 required for a bullet to register) + the visible presence kit at src/main.gd:9176-9196 (the helicopter drawn on-screen, alpha-ramping 0.35→1.0 across the 420-tick approach). Measured: for 7 full seconds the miniboss is on-screen, growing, with a real shadow — and player tracers pass clean through it with no spark, no ping, no armor_block, nothing (the bullet simply isn't dead). The sim comment says "unhittable and silent until arrival"; the view comment says the haze is meant to read "high-alt / not here yet". A visible target that eats a mag with zero feedback is the classic "my hits aren't regis…" (report text truncated at source. a512bee shipped the grenade half this run — "fly-in is bullet-proof but grenade-soft". The bullet pass-through above is the open remainder.)

Carried from the 2026-07-29 snapshot — re-flagged 2026-07-31, still open

Five teaching hints stamp hardcoded key letters (E/F/Q) into verbs the rebind system can move — including the one that fires while you're bleeding out. Where: src/main.gd:2625 (FEED THE WAR CHEST TO REVIVE — [E]), :2206 (claymore plant, [F]), :4904 (HOLD [Q] FOR THE SUPPLY WHEEL), :4909 (HOLD [E] TO DEPLOY EARLY), :4914 (airstrike wheel, [Q]). All five read Art.pad_label(...) if Art.use_pad else "<hardcoded letter>". The project ships a full keyboard-rebind system (BIND_DEFAULTS at main.gd:4020-4040, c1-18), and two sibling hints in the SAME file already read the live bind for exactly this reason — :2303 uses GameMenu.key_label(bind("grenade")) and :2888 uses OS.get_keycode_string(bind("revive")), with the comment "a stamped key would have lied about it (and about every rebind) exactly like the old one did." (Reasoned from code, not measured — the rebind path was read, not driven. Teaching entry; filed here in 2026-07-29, belongs in §3 — moved there this pass.) → see §3.
~~Downed-pilot ransom is geometry-locked: he walks AWAY from every reachable position, and the game's own camera anchor puts him out of reach.~~ RESOLVED 2026-07-31 — 66e57e5. (Was: eject floored at camera_top+120, walks north at PILOT_SPEED 1.4px/t, captured at camera_top-30; rescue = touch within 14px after the 36t grace; 120 strategy/offset/seed combinations measured; capture fuse after grace ~107 ticks.)
~~The grenade landing telegraph lies on held throws: marker + kill circle draw the 99px full lob while a >=0.27s press airbursts at 51px — a 48px lie, bigger than the blast ring, taught nowhere.~~ RESOLVED 2026-07-31 — 1ea8d25 ("telegraph now tracks the real release trajectory"). (Was: tapped lob 99px every time; held pops airburst at 51px and throws exactly once in 400 ticks while 6 taps landed; apex trips at ~16 ticks (0.27s).)

Carried from the 2026-07-26 snapshot — not re-verified this run

The tank is strictly dominant and it eats the game. Measured 3.6× safer AND 1.44× deadlier per tick, occupying 38.7% of a campaign. It makes the on-foot game — the game this actually is — the worse way to play. (Cycle 2's first fix attempt made the metric move the WRONG way and was rejected at closeness 56; re-measure before trusting any fix here.)
Deep Endless plays a water-splash sound up to 15×/second, forever, on dry land, in a mode with no water. Audible within ten seconds of deep Endless.
A frogman that beaches on dry land can never re-submerge, and is stuck in that state.
~~Unspent War Chest scores zero in Endless.~~ RESOLVED — d6c2aa9 ("Unspent War Chest worth zero score in Endless mode"). (The 2026-07-29 lens still flags a related teaching gap — the HOWTO hides the 10x victory conversion; that half shipped 2026-07-31 as 7e3d174.)
The Riot Shield's advertised counterplay is impossible. It says FLANK OR GRENADE; flanking is geometrically impossible solo because the shield re-aims at you on every hit test. 5,040 trials, 774 rounds reached the body, 774 blocked, zero kills. 2P or the REND buff both kill it in 7–9 ticks, so the block is implemented correctly — the advice is false.
The anti-camp mortar punishes you for doing the objective. A closed gate hard-pins camera_top, and the stall detector's only reset is the camera moving north — so every tick inside a gate arena scores as stalling while the HUD orders you to push north, which the sim's own clamp forbids. 46–58% of every campaign measured as camera-pinned.
The victory card's letter grade is a constant. RANK S locks in about 12% into every campaign (measured 9.3% / 13.2% / 14.5% across three seeds) and ~75% of the real score range sits above the top grade. (Note: 254b2a9 made the graded kill streak the simulated one — a different defect. The constant-grade curve above was not touched by it.)
Death silently deletes your Commendation token and Claymores with no event or cue RESOLVED 2026-07-29 — the loss beat now covers them. (sha still not located — find it and inline it.)
Crate-chasing rusher can move TWICE in one tick — introduced by the wedge slide fix.
Two movement rules revert your step with no persistent visual at all; the one-way ledge has literally zero lines of view code. (Suspicion only partly held — frequency of player contact was not established. Measure before acting.)
The airstrike is the one supply the shared no-op guard forgot. Buy it twice inside its own telegraph and you pay 200 coins for one strike — and the second purchase pushes the first later. (Note: 254b2a9 measured the airstrike's net score by screen density — break-even ~10-11 bodies — and left the double-buy untouched. 8425bd7 (2026-07-31) gated the intermission whiff window; the double-buy-during-telegraph above is a different window — re-verify against HEAD.)
CHAPTER SELECT / ARCADE plays a de-fanged campaign: three movement hazards, the whole rear-pressure system, and every price increase are silently switched off.
A quarter of all enemy gunfire in the campaign is fired by a shooter drawn off the top of the screen, and one telegraph in six is drawn where the player cannot see it. (Related: the 2026-07-31 camera-pin nests finding above adds measured peaks of 4-6 rooted turrets piled above the pinned edge during gate fights.)
The Recon Drone is the only lethal archetype with no telegraph.
The Colossus finale is the one stage no pacing table has ever covered. (2026-07-31: probe_colossus.gd now runs clean — seed 0xC0FFEE engage_t=288 fight_ticks=683 downs=3 victory=true — so the finale is at least observable headlessly now.)

3. Teaching / content honesty — the game says one thing, the sim does another

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

The War Chest tutorial tells you to spend, and the numbers pay better for hoarding — the file's own two comments assert opposite things about it. WHERE: src/view/menu.gd:5253 is the how-to paragraph the player reads: 'Spend it — 6x score. What's left when you fall salvages at only 3x. WIN, and what's left banks at 10x — plus a 5,000 bonus. Nothing pays like the chest you carry home.' The constants: SPEND_SCORE_MULT 6 (sim_world.gd:468), WIPE_SCORE_MULT 3 (:473), VICTORY_SCORE_MULT 10 + VICTORY_SCORE_BONUS 5000 (:476-477), applied at :1996 and :2654 (spend), :1079 (wipe salvage), :6825 (victory: score += war_chest * VICTORY_SCORE_MULT + VICTORY_SCORE_BONUS). WHAT I SEE: on any run that finishes, an unspent coin is worth 10 and a spent coin is worth 6 — hoarding pays 67% more per coin. The comment at sim_world.gd:473 states the design intent as 'Half of spend keeps spending strictly dominant — hoarding can never out-earn the shop', which is only true on a LOSING run (3 < 6) and is false on the winning run the whole campaign is aimed at. Thirty lines away, the comment at sim_world.gd:2646 states the opposite as the target: 'test: a hoard run should out-score an all-buy run on the same seed by 10-25%. If the gap exceeds 40% (nobody ever buys), raise to 8.' Both comments cannot be right, and the player-facing paragraph splits the difference by giving an imperative ('Spend it') and then closing four sentences later with a line that points the other way ('Nothing pays like the chest you carry home' — which is the TRUE statement, 10x being the best rate in the economy). WHY IT OUTS IT: a shop is only a decision if the player can reason about the trade. Here the tutorial's opening word and its closing sentence recommend opposite behaviours, and neither the tutorial nor the HUD ever says the thing that would settle it — that the 10x only fires if you actually reach the finale, which is what makes spending correct at low win rates and wrong at high ones. A player who reads carefully ends up distrusting the copy; a player who doesn't ends up hoarding, because the last sentence is the one that sticks. Scale check so I'm not overselling it: on a 400-coin chest the gap is 1,600 points against measured run totals of 108k-167k, so this is roughly a 1% scoring effect, not a run-defining one — the defect is the contradiction, not the magnitude. AAA VERSION: pick one and make the copy state the actual decision. Either drop VICTORY_SCORE_MULT to at or below SPEND_SCORE_MULT so 'spend it' is unambiguously true, or keep 10x and rewrite the paragraph as the gamble it is: 'Coin you spend banks at 6x, guaranteed. Coin you carry banks at 10x — if you make it. Fall, and it's worth 3x.' Then reconcile the two code comments, because right now the file documents two incompatible economies and a future tuning pass will read whichever one it happens to open.

New 2026-08-23 — from the `f1601b7..3be8149` window

"MG NEST — BREAK ITS LINE OR FLANK": the nest has no facing, no arc, no LOS gate and no maximum range — bearing changes literally nothing. The promise is a full-screen banner, _KIND_TEACH["mg_nest"] at src/main.gd:394 (verified verbatim on 3be8149: "mg_nest": "MG NEST — BREAK ITS LINE OR FLANK"), fired once per enemy kind by show_banner(_KIND_TEACH[ekind], …) (cited main.gd:9994). The behaviour is sim_world.gd:4785 on HEAD (cited :4745, verified verbatim): the entire engagement gate is if e["fire_cd"] == 0 and dlen > F_ONE and target["alive"] — i.e. the only range condition is "more than one pixel away". _spawn_mg_nest (sim_world.gd:4731 on HEAD, cited :4691) stores no face_x/face_y at all, and mg_nest_led_aim (sim_world.gd:7155 on HEAD, cited :7115) re-acquires the NEAREST player with velocity lead on every round of the burst (cited :4720-4726). MEASURED: a player stood at 12 bearings, 30° apart, at 140 px from a nest on a scrubbed field (no rocks, sandbags, bunkers or other cover) (report text truncated at source). "Flank" is not a partial truth here — there is no quantity in the code that bearing could change. ⚠️ Before fixing: §1 #5 is a standing owner decision that the nest deliberately teaches a second lane grammar ("this line follows you") and that locking it to match the other archetypes deletes the tracking rake and moves test_determinism GOLDEN. The cheap honest fix is the banner text, not the sim. This entry is filed as a copy defect for that reason.

New this run (2026-08-21)

The one card that teaches MG-nest counterplay promises a flank the nest has no facing to be flanked from. WHERE: src/main.gd:386 ("mg_nest": "MG NEST — BREAK ITS LINE OR FLANK", in _KIND_TEACH, which src/main.gd:9630-9636 calls out as "the ONLY place this game states counterplay"), against src/sim/sim_world.gd:4510-4549 (_step_mg_nest). _step_mg_nest has no facing field, no arc test and no range test. Its arm condition is literally if e["fire_cd"] == 0 and dlen > F_ONE and target["alive"] (:4543) — any distance, any bearing. Worse for the promised verb, :4515-4523 re-acquires _nearest_alive_player at the top of EVERY round of the 3-round burst and re-derives a fresh led-aim vector, so there is no stale bearing to get behind either. (Report text truncated at source.)
The HUD plays the "nest is cracking, keep shooting" ping and hint on a nest that provably takes zero damage. WHERE: src/main.gd:2513 (the armor_block event handler), :2521-2530 (the nest_hit proximity test), :2538 (_hint("nest_crack", "THE NEST CRACKS UNDER FIRE — KEEP SHOOTING, OR GRENADE IT")). The string is shipped and localized — locale/strings.es.po:144. The view decides a round "cracked the nest" by testing whether the armor_block event's POSITION is within 14px of a live mg_nest — not by observing that the nest's hp changed. But sim_world.gd:2966-2981 emits armor_block for a bullet eaten by a ROCK too, at the bullet's position. So when a nest is trapped inside a rock, every round that the rock eats fires the "it's cracking" ping.
The Technical's card teaches a dodge window that does not exist — the charge line is not chosen until the tick the truck launches down it. WHERE: src/sim/sim_world.gd:3880-3887 (_step_technical). The rev is e["windup"] = TECHNICAL_REV_TICKS (24) at :3910, but e["aim_lx"] / e["aim_ly"] are written inside if e["windup"] == 0, i.e. on the final tick of the rev, simultaneously with e["lunge_ticks"] = TECHNICAL_CHARGE_TICKS. The Field Manual card is src/view/menu.gd:5522: "TECHNICAL — revs, then charges a LOCKED line. Step off it." The constant's own comment at :83 justifies 24 ticks as "the 24t reaction floor (REAR_WARN/VENT_WARN precedent)". MEASURED over 4 seeds × campaign + endless (.aaa/c4/tech.gd): 117 revs, 113 resulting charges. Distance from the player at the instant the line is actually locked: (report text truncated at source).
The HOSTILES readout names the frogman's bullet immunity and stays silent about the ghillie's — the one that actually gates the wave. WHERE: src/view/hud.gd:1435-1436 sets immune_lurker only for e["kind"] == "frogman" and e.get("submerged", false), which posts a "GRENADES ONLY" chip beside the HOSTILES counter (hud.gd:1457-1470). The sim gives the submerged ghillie the identical immunity: src/sim/sim_world.gd:3023-3024 (bullets skip any submerged enemy) and :3204 (blast explicitly skips a submerged ghillie — so it is MORE immune than the frogman, which grenades still kill). The player sees HOSTILES 1, fires at the shimmer, and every round passes through. For the frogman the HUD says why. For the ghillie it says nothing.
The Field Manual's one economic rate — "Spend it — 6× score" — sits directly under the paragraph that defines "spend" as revive-or-buy, and the revive half pays nothing. WHERE: src/view/menu.gd:5415-5417 ("Spend it to REVIVE yourself or a partner, or BUY supplies…") immediately followed by :5424-5427 ("Spend it — 6× score. What's left when you fall salvages at only 3×. WIN, and what's left banks at 10× — plus a 5,000 bonus.") · src/sim/sim_world.gd:2548 (score += cost * SPEND_SCORE_MULT — the ONLY place a chest debit credits score, inside _try_buy) · :1929 (the priced-crate twin, deliberately kept at parity) · :2028-2090 (_try_revive: war_chest -= cost and no score line anywhere in the function). The page introduces two ways to spend the chest in one sentence, then attaches a single headline rate to "spend it" in the next.
Boss phase labels bypass TranslationServer and Art.fs() — banked in the diff's own comments and re-verified still true on 9bb1cdb. src/main.gd:10628 / :10778 (now :11026 / :11193) draw the English literal directly rather than through TranslationServer.translate(), so the localization suite cannot see the two highest-stakes strings on screen and a longer translation would widen the plate past everything the new chrome matrix measures. Separately neither goes through Art.fs(), so the accessibility TEXT SIZE setting does not scale them — confirmed Art.fs(10) returns 20 at text_scale 2.0 while Art.tw(label, 10) is unchanged. Correctly out of scope for that cycle; neither is fixed. Fix direction is owner decision #17.

Carried from the 2026-07-31 snapshot — not re-verified this run

"PERFECT DODGE!" fires for bullets in the 7–11px band that could never have killed. (audit-lens finding, adversarially CONFIRMED 2026-07-31 — not from the loop's own lenses.) Where: src/main.gd:5192-5207 (_check_near_miss dodge scan — _dist_lte(..., 11 * Fixed.ONE) at :5198 fires show_banner("PERFECT DODGE!") at :5200, pure distance check, no velocity/lethality filter; its own comment at :5182-5185 claims such a bullet "would have killed them") vs src/sim/sim_world.gd:709 (ENEMY_BULLET_HIT_RADIUS := 7 * F_ONE, the ONLY enemy-bullet kill test, :6436-6438). Arithmetic: bullets move 3px/tick, snipers 6px/tick — sampled kill distance ≥ geometric closest approach, so any path missing by >7px can never kill on any tick; the whole (7.6, 11] band during the 18-tick roll window provably credits non-saves. Rate: multiple times per firefight-heavy run. Cosmetic (banner + hitstop + sfx only, no score/currency) — but it hands out skill credit for dodges that never happened.
Riot shield's drawn plate is ~3-4× its sim blocking body — shots at the drawn edge read as blocked but the sim lets them through (or vice versa: the sim blocks air the art claims is open). (audit-lens finding, adversarially CONFIRMED 2026-07-31.) Where: src/view/art.gd:477 (riot_shield SCALE 1.1 — its "~half a p2 specialist's span" comment is false ~2×; the plate draws ~42px on a ~23px carrier) vs the sim's default 10px hit radius — the plate reaches ~32px from a soldier whose block body is 10px. The import-drift class inverted (bunker drew 14px while colliding at 48px; here the art OUTGROWS the body), and no test_hitbox_fairness row pins it. Rate: shield specialists field from campaign sector 2 / endless wave 3, so most runs past the early game show it.
HOW TO PLAY says "Bullets don't" crack "bosses" — measured real gunship kills get 20-87% of their damage from bullets, the weapon the page just told you to stop using. Where: src/view/menu.gd:5045 (_howto_page_controls verb line: "GRENADES crack armor — bunkers, bosses, the Colossus. Bullets don't."); sim truth at src/sim/sim_world.gd:6280-6299 (_bullet_hits_boss — every MG round does 1 damage with a boss_hit event; the gunship has 40 HP base). MEASURED on REAL campaigns (demo bot, god mode, 3 seeds): the gate-3 gunship died at 34.0s/40.8s/66.6s with bullets contributing 32/40, 8/40 and 35/40 of the kill damage (20-87%) — and the bot doesn't even deliberately aim at bosses. A human holding aim on the boss does 7.5 dmg/s (1 dmg per 8-tick cadence), i.e. the MG is the gunship's PRIMARY weapon; grenades (8 dmg each, 12 carried = 96 max) are th… (report text truncated at source)
Rolling in water prints "NEED COINS" — a duplicate match arm silently kills the honest teaching branch. Where: src/main.gd:2256 (the single match kind: in _consume_events) contains TWO "deny": cases — the live one at src/main.gd:2384 and a second, unreachable one at src/main.gd:2771. Sim side: src/sim/sim_world.gd:1129 emits {"t": "deny", "why": "water", ...} when a wading player presses roll, with the comment "Refuse it out loud … so the player learns the RULE (no rolling in water), not just that it failed." Verified: GDScript silently accepts duplicate match patterns and the FIRST arm wins (ran a duplicate-pattern script: prints "FIRST branch wins", no error, no warning). So every deny event routes through :2384, whose reason table {"cap","tank","board","token","fu… (report text truncated at source)
Airstrike hint gates on the 100c base price, not the depth-creeped price. src/main.gd:4977 still gates the teaching hint on sim.war_chest >= SimWorld.SHOP_AIRSTRIKE_COST (base 100c), but the wheel charges the depth-creeped price — measured 150c at wave 6 in this cycle's own capture (c6-wheel-hold.png reads "AIRSTRIKE 150 HOLD"), 125c at wave 3 per the goal's numbers. From wave 3 up the hint can fire while the strike is not actually affordable. Verified the line is unchanged by reading it after the diff. (Reasoned from code + capture, not driven.)

Carried from the 2026-07-29 snapshot — re-flagged 2026-07-31, still open

Five teaching hints stamp hardcoded key letters (E/F/Q) into verbs the rebind system can move — including the one that fires while you're bleeding out. Where: src/main.gd:2625 (FEED THE WAR CHEST TO REVIVE — [E]), :2206 (claymore plant, [F]), :4904 (HOLD [Q] FOR THE SUPPLY WHEEL), :4909 (HOLD [E] TO DEPLOY EARLY), :4914 (airstrike wheel, [Q]). All five read Art.pad_label(...) if Art.use_pad else "<hardcoded letter>". The project ships a full keyboard-rebind system (BIND_DEFAULTS at main.gd:4020-4040, c1-18), and two sibling hints in the SAME file already read the live bind for exactly this reason — :2303 uses GameMenu.key_label(bind("grenade")) and :2888 uses OS.get_keycode_string(bind("revive")), with the comment "a stamped key would have lied about it (and about every rebind) exactly like the old one did." (Reasoned from code, not measured — the rebind path was read, not driven.)
Smoke hint says "BLINDS THEIR AIM" — the MG nest, whose documented identity is suppressing through smoke, keeps firing aimed lethal rounds at you. Where: src/main.gd:2207 — _hint("smoke", "SMOKE — BLINDS THEIR AIM. SHELLS STILL FALL BLIND. KEEP MOVING"). Rule: src/sim/sim_world.gd:4141-4145 — _step_mg_nest is "the one shooter with NO concealment gate (suppressing through smoke is its whole identity)"; its 3-round bursts are aimed, lethal enemy_bullets with lead (mg_nest_led_aim), so they are covered by neither half of the hint's hedge (not AIM-blinded, and not blind AREA shells). Every other aimed shooter — elite (:3443), sniper (:3507), technical (:3599), colossus spray (:6058), gunship spray (:6239) — is hard-gated by _concealed, so the hint's absolute first sentence is true for all of them and false for exactly this one ar… (report text truncated at source)
Claymore sells "IT HURTS BOTH SIDES" — its blast spares players entirely, while the barrel's identical fireball breaks your vest. Where: hint at src/main.gd:2211 ("CLAYMORE — PLANT WITH [%s] AWAY FROM TANKS (IT HURTS BOTH SIDES)") + HOWTO verb line at src/view/menu.gd:5027 ("PLANT a claymore clear of any tank — it hurts BOTH sides.") + sim comment src/sim/sim_world.gd:46 ("it hurts both sides") + the blast itself at src/sim/sim_world.gd:2930 (_explode) and the barrel comparison at :3016-3029. Measured: an enemy-tripped claymore calls _explode(), which has NO player-damage scan — verified by stepping the sim headlessly: player standing 16px from a claymore an enemy trips takes ZERO damage (vest intact), while the identical setup with a barrel detonation breaks the vest (vest_break fired) at the same 16px. The only way… (report text truncated at source)
Trench Gun countdown is hidden exactly when it matters — "same fan" comment is wrong: on Triple it's the 5-fan, and its silent expiry cuts your pellets 5→3 mid-fight. Where: src/view/hud.gd:2071 (chip gate if p["spread_ticks"] > 0 and not p["triple"]: # redundant once Triple is owned (same fan)) and the sibling expiry-urgency gate at src/view/hud.gd:408 (same and not p["triple"] clause); the contradicting sim truth at src/sim/sim_world.gd:1365-1374 (the p["spread_ticks"] > 0 and p["triple"] branch spawns the ±24° outer pair — "a 5-pellet fan, not a no-op" — at +2 ammo per trigger pull); the game's own pickup pitch at src/main.gd:2208 ("TRENCH GUN — 3-ROUND FAN WHILE IT LASTS. ON TRIPLE IT'S A 5-WAY FAN"). Spread on a Triple owner is NOT "the same fan" — it is a 5-pellet fan with a doubled ammo tax, i.e. the single largest timed DPS swing… (report text truncated at source)
SUPPLY CALL spends a Commendation on a hidden random table — the only purchase whose outcome is unstated anywhere. Where: the wheel item at src/main.gd:424 ({"kind": 5, ..., "label": "SUPPLY CALL"} # Commendation spend — costs a token) vs the resolution at src/sim/sim_world.gd:2262-2285 (_try_token_drop: seeded roll over kinds 0-3 filtered by _supply_full). The one spend in the game whose outcome is a dice roll is also the one whose terms appear nowhere — the wheel shows "SUPPLY CALL 1* — N* HELD" (name, cost, stock, no table, no "random"), the HOW TO PLAY pages never mention what a Commendation buys at all (grep over src/view/menu.gd and locale finds zero), and the result is only named AFTER release ("SUPPLY CALL — +30 AMMO", src/main.gd:2791). The roll only filters at the hard cap, so a p… (report text truncated at source)

Resolved from this section

~~WAR CHEST howto teaches spend-6x vs salvage-3x — and hides the 10x victory conversion the whole economy is built around.~~ RESOLVED 2026-07-31 — 7e3d174. (Was: HOWTO printed 6x/3x and "That's the choice."; a WIN converts the unspent chest at 10x plus 5000 — verified live: 137-coin chest at Colossus death produced banked=137 banked_score=1370.)
~~A duplicate TRIPLE SHOT capsule prints "CLAYMORES FULL" — the honest-receipt fix hardcoded the wrong item's name.~~ RESOLVED 2026-07-31 — c957313. (Was: src/main.gd:2185-2192 one hardcoded string for every full-capsule pickup; verified live — triple-owning player + kind-6 pickup produced pickup event: kind=6 full=true cost=0.)
~~TRIPLE SHOT hint says "PERMANENT" — the sim strips it on your very next death~~ RESOLVED 2026-07-29 — 96d8928.

Carried from the 2026-07-26 snapshot — not re-verified this run

HOW TO PLAY mis-sells the Flak Vest — says armor never stops a bullet; it stops one completely. That is a 60-coin purchase decision made on false information.
HOW TO PLAY names grenades as the thing that cracks the Colossus. A full grenade carry delivers 40% of its health, and the MG out-damages grenades 5.6×. (Compounded by this run's "Bullets don't crack bosses" entry above — same page, same shape of lie.)
The taught threat model is wrong. The campaign is 68% melee, and the endless Mast pulse — taught nowhere — is 27% of every hit on the waves it is armed.
Raw token string shown on world pickup UI.

4. UI / visual polish

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

The first two are the visual lens's screenshot findings from cycle 1 (unmeasured — see the standing warning above about screenshot-only findings). The last two are cycle 3's behaviour-lens measurements against the ground-tint / lane-seal attempt, which was REJECTED and reverted — the tree was manually reset afterward, so these describe a diff that is not on HEAD; kept so the same regression is not re-shipped if that work is ever retried.

Victory Screen Stat Logic Misalignment. where: After-Action Report victory overlay (screenshot 11) | what_i_see: The end-of-run overlay shows a prominent 'VICTORY!' header and Rank A badge, while the active run counters behind it and telemetry text explicitly log 'SECTOR 2/6' and only '36m OF SCORCHED DESERT PUSHED'. | why_it_outs_it: A glaring state machine bug where mid-run progress or early deaths spuriously fire complete victory reporting cards, breaking gameplay progression trust. | aaa_version: Rigorous result-screen state validation that ensures campaign victory screens trigger exclusively upon Sector 6 completion and reflect verified run metrics.
Stacking Floating Pickup Tags Obscuring Combat. where: Dense firefights in Sector 2 and 3 (screenshots 8, 13, 18, 20, 22) | what_i_see: Multiple dark rectangular pickup banners ('MAXED', 'TRIPLE', 'SPREAD', 'PERMANENT FORD') stack tightly on top of one another directly over the player character sprite and crosshair. | why_it_outs_it: When status labels pile directly onto the player's position, incoming enemy projectiles and immediate footwork become impossible to read during peak chaos. | aaa_version: Non-intrusive HUD status icons near health meters, paired with spatial label offsets that push world text cleanly away from the player reticle.
Lane-seal SEALED state lost 88% of its contrast and is now indistinguishable from the open SCAR state. src/main.gd::_draw_lane_seals (~:8528-8548), introduced by THIS diff (verified absent on HEAD by A/B GL capture). Measured mean R channel over a 140x80 sample inside the seal rect minus an equal sample of adjacent open ground, same rows, 640x360 GL capture, campaign band 5: SEALED HEAD dR -50 -> this diff dR -6; WARN +30 -> +12; SCAR -3 -> +2. SEALED (a hard collision that reverts your step) and SCAR (walkable) are now 8 R-points apart, separated only by a 1px outline. Root cause: fx_softspot is a radial falloff card, so stretched over a 216x120 rect it deposits nearly all alpha in a small central blob and ~0 at the edges - GROUND_SLAB_MAX_ALPHA is not the binding constraint, the card shape is. This reinstates the exact bug _draw_lane_seals' own docstring says the slab was added to kill.
Trench floor tint lost 61% of its contrast. src/main.gd::_draw_trenches, routed through _ground_slab. Measured on the same GL A/B capture (campaign, trench at screen ~(275,100)-(360,122) vs an equal control patch at (430,100)-(515,122)): dR -36 on HEAD -> dR -14 with this diff. Less severe than the lane seal because the trench keeps its two raw lip rects (light top, dark bottom) which still bracket the zone, so it stays identifiable - but it is the same mechanism and will move the same way if the fix above changes the card.

New 2026-08-23 — from the `f1601b7..3be8149` window

⚠️ All seven below are from the multimodal screenshot reviewer and are UNMEASURED — they name what a frame looks like, not what the code does. That is exactly the class the standing warning at the top of this file is about; several of them (the banner stacking especially) re-flag ground the label-arbiter and band-arbiter work already moved, so enumerate the arbiter's call sites before believing the frame. Kept because the reviewer photographed real frames.

Overlapping text banner stack during combat. WHERE: mid-combat during boss phases and downed states (screenshots 8 & 9). Multiple banners — BRIDGE GUNSHIP - MORTAR VOLLEY, DESTROY THE GUNSHIP TO ADVANCE, DOWNED - PUSHER, [MORTAR INCOMING] — print directly on top of each other across the center third of the screen, "creating a tangled, unreadable block of overlapping pixels". Wanted: a prioritised event notification stack that offsets active banners, pins boss health to dedicated screen edges, and queues secondary warnings without collision. ⚠️ main.band_rows() is the existing arbiter for stacked HUD message bands and boss bars are outside it — that seam is the likely locus, not a new queue.
Mid-screen UI text stacking and boss-bar clipping. WHERE: Sector 3 & Sector 6 combat HUD (frames 10, 15). In frame 10 the boss health bar (BRIDGE GUNSHIP - MORTAR VOLLEY), white muzzle flashes, floating MAXED crate labels and a [MORTAR INCOMING] caption are stacked directly over one another, with the flash physically obscuring the boss text. In frame 15, five separate text plates clutter the bottom-center playfield. (Same class as the entry above, different frames; both kept because they are independent captures of it.)
Monochromatic visual hierarchy and muddy sprites. WHERE: in-game combat across desert and trench sectors (screenshots 2, 3, 12, 14). Player soldiers, hostile riflemen, sandbags, crates and desert scrub share almost identical desaturated tan/olive/brown tones with flat top-down silhouettes, so active combat units blend into ground cover. Wanted: high-contrast unit sprites with dynamic drop shadows, distinct directional silhouettes, and saturated uniform accents. ⚠️ Cross-ref §1 #12 — the warm-light separator rim already exists and is an open owner decision on its width (2.2 px); and §1 #11 records that ghillie and sapper have no working rim at all. This is the same problem the rim was built for, reported from the screen rather than the code.
Unmasked tile-stamp boundaries and geometry artifacts. WHERE: trench walls and fort ground tiles in Sectors 3 and 4 (screenshots 3, 12, 14). Dark grey rectangular tiles and hard ground blocks sit on the desert sand with harsh, unblended rectangular edges and floating tile stamps along trench perimeters, exposing the tile grid. Wanted: seamless auto-tiling with organic dirt transitions and edge-smoothing tiles. ⚠️ Cross-ref §1 #30/#34 — the ground base lattice was measured at only ~4.4% luminance amplitude and two lenses agreed the original tell was over-claimed. This report is about structure tiles, not the sand base, so it is a different surface; do not close it against the ground-repeat measurements.
Draft-quality screen layout in the Field Manual. WHERE: Field Manual / How To Play (frame 17). Page 1 of 4 shows "a massive empty dark green modal containing a single sparse bullet point (MOVE with W/A/S/D.) surrounded by vast unused black space" — over 85% of a full-screen menu panel blank. ⚠️ This is the known cost of the entry-granular pager, already measured: §5 records the sparse leaves at 125% WAR CHEST 73%/99%/14% and 200% WAR CHEST 85/34/90/49/90/52, and §1 #26 is the standing owner decision on page-count vs page-density. Do not fix this without ruling #26 — the blank space is the price of never breaking a sentence, and 9538eca ratcheted the sentence-integrity half.
Inconsistent UI frame colour schemes across sub-menus. WHERE: Pause, Field Manual, Chapter Select and Options (frames 16, 17, 19, 20). Each submenu has a different frame palette — Pause olive-tan, How To Play bright toxic green, Chapter Select pale greenish-tan, Options cool blue-grey. Reads as sub-menus built by different hands. Wanted: one chrome palette with intentional per-tab accents rather than four independent frames.
Stark text-only Chapter Select menu. WHERE: Main Menu → Campaign → Chapter Select (screenshot 17). A plain vertical text list of six zone names inside a green CRT frame — no thumbnails, biome artwork, sector stats or completion badges. Wanted: a tactical map or carousel with sector previews, objective briefings, high-score badges and biome art. Scope warning: this is a feature request, not a defect — it is the largest single piece of new art/UI work anywhere in this file.

New 2026-08-22 — from the `9bb1cdb..f1601b7` window

Anchored fork signage bypasses the label arbiter entirely, so it is not rail-bounded and the new ratchet does not scrape it — banked because the CLASS is unpinned, not because it currently misbehaves. src/main.gd:9582-9583 append crect/brect straight into _label_slots without ever calling claim_label_slot (verified on f1601b7: both _label_slots.append(crect) and _label_slots.append(brect) are there, and neither goes through the arbiter). That is by design — anchored signage reserves its pixels first and never moves — but the new ratchet's producer scrape only covers _world_label*, floattext "text": literals and BUY_FLOAT, so signage sits outside the swept class. Verified NOT a live defect today: fork_sign_relevance (main.gd:9417) reaches alpha 0 at screen y = SIGN_FADE_GONE 210 (main.gd:10615), 102px above the caption scrim top at y 312, and _fork_sign_fade lerps toward it at 0.12/frame — so the sign is invisible long before the rail. Fix the scrape's coverage, not the sign.
Three simultaneous toasts in the bottom 15px now yields 2 legible + 1 suppressed, where before it was 3 drawn with 2 illegible. A visible behaviour change worth knowing about, not a defect. Measured on a posed capture (REVIVE 40¢ @ screen y348, CLICK @ y352, AIM AWAY @ y344, all three at once): BEFORE, all three drew — REVIVE and CLICK smothered under the scrim, AIM AWAY clear. AFTER, REVIVE and CLICK are lifted clear and legible and AIM AWAY is dropped by the droppable-suppression path (fx["sup"] = true). That is the codebase's own blessed outcome for a saturated ladder and it is a net win. Note the scope: this affects floattext toasts only — the actionable persistent labels (REVIVE cost, AIM AWAY as drawn by _world_label in production) clamp above the rail rather than drop.

New this run (2026-08-21)

Regressions introduced by this window's own commits — fix these first.

REGRESSION 9bb1cdb — Ground band stack lost 32px of bottom overscan; a 1-4px grey strip can show at the frame's bottom edge under max shake. Measured, not predicted. ground_base_bands (src/main.gd:1139) emits int(ceil(360.0 / p)) + 1 = 5 bands x 96px = 480px of vertical coverage; HEAD painted 8 rows x 64px = 512px. _bg_root is a child of main (main.gd:965) and main.position carries shake/kick/roll (main.gd:6224), so the ground rides the judder. Measured by rendering shot 0 in both trees (real GL, screenshots.gd, gl_compatibility) at forced camera offsets and finding the last warm-ground row: dy=-120 gives HEAD 359 vs working 335; dy=-200 gives 287 vs 255. Reachability: band bottom = floor(-fposmod(cam_y,96)) + 480, spanning 384..480, so the worst 1-in-96 camera phase puts it at 384+offset.
REGRESSION 9bb1cdb — The off-subject label gate is wired into 3 of 15 call sites; 12 offset-anchor callers can suppress a label whose subject IS visible. (HEAD had no gate, so these always drew.) main.gd:10827-10831 states the contract — offset callers pass subject — but only 9642 (capsule name), 9648 (MAXED) and the 9655 crate-price claim do. These pass an anchor 8-26px above their subject with no subject, so WORLD_LABEL_FRAME.has_point(pos) suppresses on a VISIBLE subject in the top 8-26px: 9158 (-8), 9209 (-8), 9746 LOW FUEL (-26), 10219 RESCUE (-18), 10482 SILENCE THE SPOTTER (-20), 11531 REVIVE %d (-16), 11548 revive prompt (-16), 11691 AIM AWAY (-14), 11834 and 11850 GET UP (-26). (9142 FORD OPEN is +10, below its subject, safe; 10215 ESCAPING! hand-clamps already; 12823 CALLOUT_OVERFLOW anchors at y=40/band_bo… (truncated at source).)
REGRESSION-adjacent 9bb1cdb — ESCAPING! is the one world label whose off-frame gate reads the TEXT's left edge, not its subject — and the ratchet exempts exactly that site. src/main.gd:10220 _world_label_centered("ESCAPING!", epos.x, maxf(epos.y - 18.0, float(Art.fs(8)) + 2.0), pi_col) is the only producer that passes no subject. Its gate point therefore becomes Vector2(epos.x - Art.tw("ESCAPING!")/2, …), i.e. the left edge of the plate, so the red escape warning is suppressed whenever epos.x < tw/2 — 29.5px at 100% text size, 59px at 200%. At HEAD it always drew. The ratchet does not cover it: tests/test_main.gd's _wl_anchor_dy returns INF for any arg containing maxf(/clampf(, and the sweep continues on INF. That exemption's stated justification ("a caller that clamps its own baseline into the frame is safe by construction") is a Y-axis argument applied to an X-axis failure.
A visible priced crate in the top ~21px of the frame now shows no price. src/main.gd:9566 gates on the label's want-rect (Rect2(ppos.x-15, ppos.y-34, w, 13)), so for ppos.y <= 21 the rect is entirely above y=0 and the price is skipped even though the crate itself is on screen. Same shape at main.gd:10733 for _world_label: the capsule name is suppressed for ppos.y <= ~11. Left of x=262 that band is behind the HUD plate, but the right half is bare playfield. This is what the plan asked for verbatim, and it is strictly better than the old behaviour (which printed the price on the HUD), but the gate should test the SUBJECT's position, not the plate's.

Ground / desert repetition — what actually survived 9bb1cdb.

The ground base still repeats verbatim — the peak MOVED 64px → 96px, it did not vanish. Rendered on real GL (gl_compatibility, M4 Max) with tools/ground_lag.py, own before (src/main.gd stashed to HEAD) and after captures. lag-96 excess over the neighbouring-lag floor: 01-jungle-firefight x +0.128 / y +0.117 (HEAD -0.002 / -0.010); 02-tank-assault x +0.136 / y +0.138 (HEAD -0.007 / -0.001). The old lag-64 peak was +0.148/+0.144, so amplitude is essentially unchanged. src/main.gd:7191-7208 and ground_lag.py's docstring both state this; nobody is lying about it, but the goal named ground-tile repetition and 6.7 verbatim repeats across the frame is still repetition. Needs a second source card (hash-selected per world BAND, which is scroll-stable).
The earlier dressing-only pass moved the verbatim 64px repeat by −5%, not −43%. Rendered before/after with tools/screenshots.gd under --rendering-method gl_compatibility. Ground-masked (|lum - row median| < 10, eroded 8px) and high-passed (subtract a 32px box blur) horizontal autocorrelation over rows 96..320: lag-64 0.648→0.617 (01-jungle-firefight), 0.733→0.696 (02-tank-assault), 0.672→0.637 (04-bridge-gunship); lag-128 0.598→0.570, 0.702→0.656, 0.642→0.600. Off-lags (48/56/60/68/72) sit at −0.05..0.005 in both, so the lag-64 peak is unambiguous and still 0.62–0.70 after the fix. Median reduction 5.0%. The plan's target was "below ~0.6 at the desert stop" from a modelled 0.436→0.247 (−43%) — the model over-promised.
lag-192 y residual survives on the two desert shots (the 64px dressing × 96px base lcm). tools/ground_lag.py --lag 192: 01-jungle-firefight y +0.058 (x -0.005), 02-tank-assault y +0.055 (x -0.013). HEAD was 01 y +0.248 / x +0.136 and 02 y +0.253 / x +0.132, so this is ~4.3x better on the axis that still shows and gone on x — but not gone. ground_lag.py's header states this openly and attributes it to the two dressing lattices rather than the base, which matches: the base is flat at both 64 and 96. Nothing gates it. See owner decision #34.

Boss / fly-in presentation.

_boss_flash never reaches the fly-in hull — the body does not react to 32 HP of damage. src/main.gd:10535 vs :2639 and :10606. boss_hit sets _boss_flash = minf(1.0, _boss_flash + 0.35) with the comment "the big body reacts, not just a spark", but the fly-in draw is a bare _spr(body_tex, apos, PI, asc, Color(0.92, 0.94, 1.05, 0.35 + eta_f * 0.65)) and the arrived path's hull_mod.lerp(Color(2.2, 2.2, 2.2), _boss_flash) is at :10606, past the branch's return at :10544. MEASURED: on seed 7 the shipped bot removes 32 of the boss's 40 HP during the 420-tick approach and the hull never flashes once. Sparks (fx_impactdark), the hitmarker, ping_shell and the newly-drawn HP bar all do fire, so the reported "zero feedback" tell is genuinely gone — this is the last missing beat.
Hull scorch/smoke damage cues never appear during the approach. _boss_wounds(bpos, 1.0 - bfrac, 34.0) is called only from the arrived path (src/main.gd:10641), after the fly-in branch's return at :10555. MEASURED on a driven run (demo_input, god_mode, seed 7): the bot takes the miniboss from 40 to 8 HP during the approach — an 80% wound — and the hull shows zero scars, zero smoke, zero sputter, because BOSS_WOUND's scar_start threshold is only consulted in the arrived branch. Only the bar drains. Not a lie, but the window lets a boss be shot to near-death with no hull-read at all.
The whole approach renders behind the opaque top HUD panel (pre-existing framing, NOT a regression — verified identical on HEAD by reading the diff hunks; BOSS_Y_OFFSET, the camera_top + 90 gate_y offset and the 150/55 ramp magnitudes are byte-identical, just renamed to BOSS_FLYIN_DX/DY). MEASURED from captures and the ramp: the hull's screen y runs from -5 (phase_t -420) to 50 (arrival), while the HUD panel's bottom edge sits at about y 57. Combined with scale 0.5 → 1.3 and alpha 0.35 → 1.0, the hull is dim and small AND occluded for the first several seconds — at phase_t -420 I could not locate it in the 640x360 frame at all; it first becomes findable around phase_t -180 as a rotor cross showing through the panel at (384, 26). Caps how much of the fly-in fix a player can actually see.
The boss HP bar's max is learned from the first frame the view sees, so a boss first observed at partial HP shows a full bar (PRE-EXISTING; the line is carried over unchanged by the _boss_bar extraction). main.gd: _boss_hpmax[bkey] = maxf(_boss_hpmax.get(bkey, 1.0), float(boss['hp'])). Observed directly: a capture harness posed the miniboss at hp 31/40 and the bar rendered as a full bar, because 31 became the learned max on the first drawn frame. Benign today (the first drawn frame is now the spawn tick at full HP); a latent trap for any future path that reveals a boss mid-fight (save/resume, spectator join, pre-damaged boss). The honest fix is to read BOSS_HP / the sim's own max.
Wreck hulk from a mid-approach kill is hard-coded to x = 320 (pre-existing pattern, newly reachable; re-verified present on 9bb1cdb at src/main.gd:10511). _hulks.append({"x": 320 * Fixed.ONE, …}) with a rotating y. Before 016b4cd the miniboss could only die at x == SCREEN_CX == 320, so the constant was accidentally correct. It can now die anywhere along the ramp up to screen x 470 (measured: seeds 11 / 3 / 23 kill it at phase_t -170 / -227 / -33). Cosmetic only — the hulks are a stylised scar pattern, not a crash-site record — but the constant is now a coincidence rather than a fact.

Rooted-unit arrival.

mg_nest sandbag base draws at full opacity on its birth frame — 41% of the emplacement still pops. src/main.gd:10071: _spr("sandbag_beige", epos, 0.0, 0.5, Color(0.82 - n_dmg*0.13, 0.8 - n_dmg*0.15, 0.62 - n_dmg*0.12)) — implicit alpha 1.0, never multiplied by n_arr. Only mg_stand (:10081) and the 3 armour pips (:10084-10087) ramp. Measured via opaque-alpha bbox × art SCALE × call-site spr_scale: sandbag.png 80x40 opaque → 33.2x16.6 px on screen, 325 px²; mg_stand.png 80x118 opaque → 24.0x35.4 px, 477 px². So 325/802 = 41% of opaque area, and the widest element of the silhouette, appears from nothing at 100% opacity on frame 1. Verified by capture: a single nest posed at screen (320,160) rendered at age -1 (solid) and age 0 — the sandbag ring is unchanged between the two.
_rooted_arrive is registered for the broadcast mast but nothing ever reads it. src/main.gd:3300 registers _rooted_arrive["x,y"] for every rooted_spawn event including kind=="broadcast", but the broadcast draw branch (:10047-10062: radio_tower sprite, the two 48px base arcs, the BROADCAST_HP pips) never calls rooted_arrival_alpha. Verified by capture: the masts are pixel-identical at ramp ages 0/6/12 and at no age. The shipped test asserts "'broadcast' arrival registers its dig-in ramp under the unit's stable x,y key" and therefore passes on write-only state. Either stop registering the mast's key or make the branch read it. (Design half is owner decision #27.)
_rooted_arrive is not cleared on restart. The dict is populated in _consume_events and aged in _update_feel, but no restart/mode-change path clears it alongside _fx.clear(). Entries evict themselves within ROOTED_ARRIVE_FRAMES (24 frames, ~0.4 s), so practical exposure is one stale ramp key on a run restarted mid-fade at a colliding fixed-point x,y. Small, but unbounded by luck rather than by construction.

Field Manual layout.

The 200% Field Manual now leaves a big void on leaf 1 — entry-granular packing never splits an entry that fits a leaf, even at a sentence boundary. NEW, introduced by 7c16776, visible in the release capture set. Re-rendered both sides through tools/screenshots.gd shot 10 (200% / CONTROLS): HEAD's leaf 1/4 is full (heading + 2 bullets, last line "gun fires on its own —"); the fixed leaf 1/4 holds a 2-line heading + the single bullet "MOVE with W/A/S/D." and then ~150 px of empty black — roughly 42% fill of the 166 px capacity. Measured fill ratios at 200% over _howto_large_pages: WAR CHEST 84/33/90/48/90/52% across SIX leaves (HEAD produced five), SPECIALS 80/48/63/97/97% across FIVE (HEAD produced four). So 2 of the 20 (tab × scale) combos gained a leaf, and four leaves sit under 50% full. (Density trade is owner decision #26.)

Visual-lens entries — reasoned from screenshots, not measured.

Menu Header Overlapping Tab Bar Border. Ordnance Rig / Controls Rebind screen: the section header text "CONTROLS" is positioned too high on the panel, directly intersecting and double-exposing over the bottom border line of the tab navigation bar. Text colliding into structural UI borders shows rigid absolute positioning that breaks visual hierarchy. AAA version: flexible layout containers with defined margins so headers and tab bars never overlap regardless of resolution.
Overlapping On-Screen Text Banners. Upper-center screen space during active combat objectives (Image 10): the orange banner "GOD MODE — DEBUG ONLY — RUN CANNOT END" renders directly over the objective text "DESTROY THE GUNSHIP TO ADVANCE", creating a messy collision of overlapping pixel fonts. Static screen Y-coordinates for event text without a layout queue or stack manager cause notification prompts to collide whenever multiple conditions trigger at once. (Note for the fixer: band_rows() is the existing arbiter — the god-mode banner is presumably drawn outside it. Enumerate its call sites first.)
Inconsistent Player Health HUD Display. Top-left player HUD strip, comparing frames 1-3 vs frames 4-5: the P1 health readout constantly changes style between screens. Frames 1 through 3 display a clean, green segmented bar next to the player label, while frames 4, 5, 18 and 29 abruptly replace the bar with a cryptic row of ASCII text symbols ("P1 E F"). Switching core HUD elements back and forth between graphical bars and raw debug character strings breaks presentation standards and confuses vital player feedback.
Sparse and Unfinished How To Play Screen Layout. Field Manual / How To Play modal, Controls tab (frame 19): the body card contains only a single header line "MOVE AND AIM FIRST — THE REST IS EXTRA:" and one line of text "MOVE with W/A/S/D.", leaving over 80% of the screen as completely empty dark space. Vast empty cards in help menus look like placeholder content caught mid-development. (Same root as the 42%-fill measurement above — the two entries are the visual and the measured halves of one defect.)
Unmasked Overlay UI Clipping. End-of-run Victory report screen (Image 23): targeting reticle icons and UI crosshairs on the bottom-left render directly over and under the border of the modal Victory card popup, clipping through the frame instead of being hidden or properly layered behind a full-screen menu mask. Modal pause screens and post-game report cards should cleanly freeze and isolate world-space UI. (Duplicate of the same-titled 2026-07-31 entry below — kept because it was re-observed on a later build.)

Carried from the 2026-07-31 snapshot — not re-verified this run

Repetitive Grid-Based Ground Tile Seams. Where: Sector 1 and river crossing stretches (Screens 1, 4, 6, 20). The desert ground exhibits obvious repeating grid patterns, featuring harsh rectangular transitions where land tiles meet water channels or road strips without visual edge smoothing, foliage overhangs, or organic blending. Visible tile repetition and abrupt straight-line terrain boundaries give the environment an unpolished tilemap-editor look rather than an immersive battlefield feel. AAA version: rich autotiling transition sets, layered foliage fringes, ground decals, and subtle noise variation to mask the grid. (Visual-lens entry — reasoned from screenshots, not measured.)
Floattext down-stack can overprint the commander bark row at the bottom rail (pre-existing, not from this diff). Visible in both HEAD (6-plate) and fixed (4-plate) captures: the floattext stacking block bumps toasts DOWN from the anchor, and at the bottom rail the lowest plate overprints the COMMANDER bark row. Verified by capture (/tmp/cic_shots/shot1_BEFORE.png vs shot1_AFTER.png: HEAD stacks 6 plates into the bark, the fix stacks 4 and the top edge of the lowest still touches the bark plate). The cap strictly reduces the pile but the downward-bump rule has no rail guard against the bark zone. Right answer is a bottom-rail reservation the way _band_floor() reserves the top; measured: 4 plates × 11px can still reach the bark row when the anchor is the player standing … (report text truncated at source)
Intrusive Knockdown Screen Smear Obscuring Gameplay. Where: knockdown and revive sequence transitions (Screenshots 5, 13, 14). When the player is knocked down, a violent horizontal motion-blur glitch filter and heavy crimson tint flood 100% of the viewable screen; enemy positions, incoming bullets, and terrain obstacles blur into near-complete illegibility. Sacrificing gameplay clarity during critical rescue/revival moments for heavy post-processing undermines player feedback and co-op tactical awareness. AAA version: localized feedback — peripheral vignette darkening, subtle screen shake, distinct audio ducking — while … (report text truncated at source. Visual-lens entry.)
Unmasked Overlay UI Clipping. Where: end-of-run Victory report screen (Image 23). Targeting reticle icons and UI crosshairs on the bottom-left render directly over and under the border of the modal Victory card popup, clipping through the frame instead of being hidden or properly layered behind a full-screen menu mask. Modal report cards should cleanly freeze and isolate world-space UI; active targeting cursors bleeding through modal menu cards reads as an unpolished camera/UI rendering pass. AAA version: isolate modal report screens onto a clean UI render pass that suppresses, hides, or darkens all active world-spa… (report text truncated at source. Visual-lens entry.)
Vertical-only label dodge + deck gunner has no player exclusion. claim_label_slot dodges vertically only (x is clamped, never moved), and the player-label reservation added this cycle skips any player with in_tank >= 0 — but a non-driver tank occupant (deck gunner) is drawn standing on the field at _to_screen + (0,-7) (main.gd:9866-9868) and gets no exclusion rect, so labels can still ink over the deck gunner. Both limitations are named in the plan as banked; verified the gunner draw path is unchanged.
Four entity-anchored world marks bypass the label arbiter. The "!" mark (:7850), ford label (:8048), gate numeral (:8169), and "?" (:8882) remain drawn outside the claim_label_slot arbiter. Small entity-anchored marks, none part of this cycle's tell; banked by the plan and still unarbitrated in the diff.

Carried from the 2026-07-29 snapshot — still open

Unmasked Radial Shop Menu Overlay. Screenshot 12 (Wave Cleared shop screen): the circular shop icon wheel drops directly onto the active map terrain without a background dark scrim, darken vignette, or focal blur, causing ground sandbags and rocks to clip visually into the item icons. AAA titles apply a soft full-screen dark overlay (e.g. 60% black scrim with background blur) whenever a shop opens. (ee25263 shipped a fix titled exactly this on 2026-07-29 — the lens still flags a remainder. Re-measure against HEAD before acting; the open item is what survived ee25263, not the original complaint.)
Full-Screen Low-Pass Distortion Blurring the HUD. Heavy damage hit-stop frame (Screenshot 5): taking near-lethal damage applies a radial screen-space chromatic blur across the entire frame, smearing HUD numbers, objective banners, and bottom button legends into double-vision pixel noise. Applying heavy screen-space shaders blindly over top-layer HUD elements reveals a flat single-pass camera setup rather than rendering UI on a dedicated overlay camera buffer. Best-in-class titles restrict high-impact combat VFX to the game-world layer while keeping HUD bars, scores, and text crisp on an un-distorted top pass. (A related fix shipped — "full screen low pass distortion blurring the hud" is on the shipped list, sha not located (5cfea32 covered menus only). As above: the open item is the remainder.)
Unaligned Leaderboard Column Formatting. Hall of Fame table screen (Screenshot 16): the leaderboard table header ('# RANK SCORE MODE REACHED STREAK') features uneven horizontal spacing where column titles crowd together irregularly, above plain, unadorned sub-panel framing. AAA arcade menus use pixel-precise grid alignments, crisp column dividers, distinct visual hierarchy between headers and entries, and glowing selector states. (2bedaf1 shipped a fix titled exactly this; the lens still flags a remainder. Verify against HEAD before acting.)
Raw Blocky Text World Signage. Images 3, 15, and 17 (Bounty and Cache screen markers): huge, flat cyan and yellow pixel-font text strings reading 'BOUNTY >' and '< CACHE' sit drawn flat across the lower playfield directly over rocks, sandbags, and active combat sprites without depth or diegetic framing. Rendering massive raw text overlays flat across environmental art looks like developer debug flags rather than polished UI navigational signposting. AAA version: anchor directional markers to screen edges or render them as diegetic, styled UI badges with subtle floating animations and drop shadows. (A fix titled "raw blocky text world signage" shipped — sha not located. The open item is the remainder.)
Static Scanline Overlay Obscuring Menu Typography. Options menu and Control Rebind sub-menus (Images 18, 19): a heavy, dark horizontal scanline filter is drawn over the entire menu container, cutting directly across fine pixel letters and rendering small button callouts and option labels dim and grainy. AAA version: top-tier retro-styled titles place UI text overlays on an unfiltered canvas or fine-tune scanline strength so interface typography remains sharp, bright, and effortlessly readable. (Re-flagged 2026-07-31, unchanged.)

Resolved from this section

~~Floattext spawn punch can poke ~4.5px above band_floor for ~2 frames~~ RESOLVED 2026-07-31 — 1ea6508. (Was: fpivot.y clamped to band_floor + fsz but the 1.5x spawn punch scales the ascent about the pivot, reaching band_floor - 4.5px for fsz=9 during the first ~2 frames; self-correcting because the band draws after _draw_fx. The down-stack bark overprint above is the remaining floattext-rail gap.)
~~Screen-cluttering world-space text overlays~~ RESOLVED — f255cab part 2 (claim_label_slot() body written to spec AND wired: _draw() clears the claim list each frame, _world_label routes every world-space string through the arbiter, plate and ink move together; ratchet: 15,320 colliding pairs and 361 off-frame escapes → 0 and 0).

Carried from the 2026-07-26 snapshot — not re-verified this run

Inconsistent pixel typography and menu padding (How To Play, Options).
Generic box-border menu framing — reads as prototype chrome. (6ecbf0d "programmer-style UI modals and plain menu framing" is on this run's shipped list — likely this entry; verify before striking.)
WARN and LABEL wheel plates overlap by 2px, leaving a double-blended dark band.
HALL recency band collides with the filter tab row (introduced by the cycle-1 diff).
MODES tab row pitch is too tight for the new 2-line tips.
HOWTO "N / 5" tab counter crowds the ENDLESS tab label.
Wheel text is the only view text that ignores the accessibility text scale (pre-existing).

5. Test & tooling debt

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

Four untracked scratch probes (probe_two.gd, probe_ammo_econ.gd, probe_bash_rate.gd, probe_offscreen_fire.gd) sit in tools/ with no .gd.uid sidecars, filed repeatedly across this run's cycles. (cycle 1) tools/probe_ammo_econ.gd, tools/probe_bash_rate.gd, tools/probe_offscreen_fire.gd, tools/probe_two.gd are present as untracked files (git status ?? on all four) with no accompanying .gd.uid, which this repo commits as a convention for script files. Grepped all four for concussion|_blast_warp|screen_fx: zero hits, so they are unrelated to this goal — almost certainly litter from a parallel session. I did NOT run them and make no claim about whether their output agrees with the tests they touch. They are a hygiene risk because the next git add tools/ sweeps four unvalidated measuring rigs into a commit.

(cycle 1) git status --porcelain shows A tools/probe_ammo_econ.gd, A tools/probe_bash_rate.gd, A tools/probe_offscreen_fire.gd, A tools/probe_two.gd — staged and off-goal (shotgun fan pellet counts, archetype telegraph vs the 24-tick reaction floor, ammo economy, bash rate). Each passes godot --headless --path . --check-only. None has a .gd.uid sidecar, while ls tools/*.gd.uid | wc -l = 30 covering the existing tools — the repo convention is that these sidecars are committed.

(cycle 1) tools/probe_ammo_econ.gd, tools/probe_bash_rate.gd, tools/probe_offscreen_fire.gd, tools/probe_two.gd — untracked, no .gd.uid (the other 30 scripts in tools/ each have one, repo convention), zero references to concussion/screen_fx/periph, mtimes 22:34-22:57 on 2026-08-23 vs 00:47-00:48 on 2026-08-24 for this attempt's files. Leftovers from an earlier cycle in this shared worktree. Verified unrelated by grep; they will be swept into the fix commit by a git add -A.

(cycle 1) tools/probe_two.gd, probe_ammo_econ.gd, probe_bash_rate.gd, probe_offscreen_fire.gd are untracked (??), have no .gd.uid sidecars, and are absent from CLAUDE.md's tool inventory. They are NOT this attempt's work: mtimes 2026-08-23 22:34–22:57 vs this fix's 2026-08-24 01:18–01:20. Contents are unrelated to the concussion goal (ammo economy under always-fire, bash rate, offscreen fire, pellet-fan value + telegraph lengths). Left untouched deliberately — they belong to whoever wrote them, and deleting another cycle's evidence is worse than leaving it. Someone should either commit them with docs+sidecars or bin them.

(cycle 3) tools/probe_ammo_econ.gd, tools/probe_bash_rate.gd, tools/probe_offscreen_fire.gd, tools/probe_two.gd - untracked, no .gd.uid sidecars (repo convention requires them), and about ammo economy / bash rate / offscreen fire, i.e. nothing to do with this cycle. Almost certainly another session's scratch in this shared worktree. Stage this cycle's files explicitly.
The +9.5% draw-call / +10% render-CPU cost is unguarded. tools/perf_probe.gd on real GL, 121 frames, driven by demo_input: draw_calls_avg 259.760 -> 284.512, draw_calls_max 272 -> 309, cpu_ms_avg 0.801 -> 0.883, cpu_ms_max 2.557 -> 2.681. Fine against a 16.7 ms budget, but +66% desert dressing cards x 3 draw passes is the kind of thing that gets bumped again next cycle with nothing to notice. test_perf.gd is sim-tick only and does not see this at all.
No automated check anywhere asserts concussion centre readability, so a fix and a no-op are indistinguishable to the suite. Grepped tests/ for concussion|screen_fx|blast_warp: only three hits exist. tests/test_main.gd:410 pins the source text _concussion = maxf(_concussion, down_scale); tests/test_view_honesty.gd:3045-3049 asserts the shader has a uniform float clock and does not animate off GPU TIME; tests/test_hud.gd:3420 is a comment. tools/probe_concussion_hud.gd measures HUD edge SURVIVAL (EDGE_KEEP_RATIO 0.85 at :35, enforced :137) but censuses only the two HUD bands and a mid-screen world band as a liveness control (WORLD_Y0/Y1_LOGICAL 180/260 at :30-31) — nothing measures whether the world at screen centre stays legible. Verified by running the full suite on an unmodified tree: 1162 methods / 37620 assertions / 0 failures, i.e. the suite is green on the defect. PREDATES THIS DIFF.
probe_concussion_hud.gd calls an undefined _box_edges() — the probe no longer loads, and a load failure exits 0. tools/probe_concussion_hud.gd:164-165 call _box_edges(a, cx0, cx1, cy0, cy1); the file defines only _band_edges(img, y0, y1) at :206. Running godot --path . --rendering-method gl_compatibility -s res://tools/probe_concussion_hud.gd prints SCRIPT ERROR: Parse Error: Function "_box_edges()" not found in base self. (×2) plus two dependent 'Cannot infer the type' errors, then ERROR: Failed to load script ... with error "Parse error" and EXITS 0 with no PROBE OK / PROBE FAILED line. Verified this is a REGRESSION introduced by this diff: I stashed only that file and re-ran, and HEAD's version printed PROBE OK — HUD edges survive concussion (top 1677->1599, bottom 601->582), world blurred (51200 px differ). Measuring rig now silently reports success by exit code.
The probe's new centre box contains no player soldier and falls under its own signal floor. CENTRE_{X0,X1,Y0,Y1}_LOGICAL = 240/400/135/225 in tools/probe_concussion_hud.gd. I supplied the missing _box_edges and ran the probe: centre 99->1 (keep 0.01, floor 0.85) and the probe's own guard fired — probe_concussion_hud: the centre box only carried 99 strong edges at baseline — nothing legible was there to protect, so the centre check proves nothing (ea_mid 99 < the 200 floor at :167). The doc comment claims the box is 'the patch the player soldier occupies'; in the probe's own posed frame (dumped via PROBE_SHOT_DIR) the soldier is at roughly (278, 292) logical, below cy1=225. Measured on the dumped 640x360 PNGs at EDGE_LUMA_JUMP 0.19: box 240-400 x 135-225 = 100->6 (keep 0.060); a soldier-containing box 200-440 x 240-330 = 408->68 (keep 0.167); 160-480 x 120-280 = 598->22 (0.037); 120-520 x 100-300 = 868->48 (0.055).
Fabricated measurements committed as a source comment. tools/probe_concussion_hud.gd, above CENTRE_KEEP_RATIO: 'Measured on this tree (see the commit body): pre-fix 0.62, post-fix 0.93.' Neither number is producible from this tree. The probe does not compile as shipped; with the missing helper supplied the real pre-fix reading is keep 0.01 in that box (99->1) and keep 0.167 in a box that actually contains the soldier (408->68). There is no post-fix state at all because the shader was never changed.
README test-method count not updated alongside the new test — 2 suite assertions fail. The diff adds one func test_ to tests/test_view_honesty.gd, taking the live count from 1162 to 1163, but README.md still reads tests-1162%20methods (line 20, badge) and **1,162 test methods / 37,400+ assertions** (line 338). Full-suite output: ✗ README quotes the live method count ("1,163 test methods") — recount says 1163 and ✗ the README test badge carries the live count ("tests-1163%20methods"). Recount confirmed by hand: grep -hE '^func test_' tests/test_*.gd | wc -l = 1166, minus 3 in test_perf.gd = 1163. CLAUDE.md states adding a test requires updating README.md in the same commit.
Full test suite wall time was 10m53s at 35% CPU, ~28x the ~23s documented in CLAUDE.md. time tools/run_tests.sh reported 221.49s user 12.04s system 35% cpu 10:53.13 total while producing a coherent report (FAIL — 10 of 37631 assertions failed) with a single Godot process observed at 82% CPU mid-run. Not attributable to this diff (the failures are all assertion-level, not hangs), and not the documented low-CPU abort class since it completed. Flagging as an unexplained discrepancy against the ~23s figure in CLAUDE.md so a later cycle can decide whether that documented number has rotted or whether something in the tree got slow. SUITE=view_honesty alone ran in 35s.
The committed check passes only because the probe poses the soldier at the one x that works. tools/probe_concussion_hud.gd never sets the player's position — it captures wherever the 150-frame boot settle leaves him, printed as soldier at (280.0, 300.0). That is 40px from mid-field, the single position at which the screen-centred ramp spares him. Every other x in the play field fails the probe's own 0.75 floor (see previous finding). The probe is loaded, not wrong-headed: it needs an x sweep (x=40 / 320 / 600) at the shipped peak.
Probe pins amplitude 0.7 while the shipped solo-knockdown peak is 1.0, and the floor fails there. tools/probe_concussion_hud.gd:121 sets _main._concussion = 0.7; tests/test_view_honesty.gd:3057 calls that "the shipped concussion peak". src/main.gd:3155 sets _concussion = maxf(_concussion, down_scale) where down_self_scale(partner_standing=false) returns 1.0 (src/main.gd:3735) — a solo knockdown, the case this goal names, peaks at 1.0. 0.7 is the flak-vest break (src/main.gd:4372). Ran the probe with the constant changed to 1.0, everything else identical, fix in place: centre 348->254 (keep 0.73, floor 0.75) and PROBE FAILED, 3/3 runs (254/257/253 of 348/351/347). Pre-fix at 1.0: centre 351->26 (keep 0.07). So the ratchet's amplitude window is 70% of the real thing, and correcting it to the real thing turns the committed floor red.
probe_concussion_hud.gd 'keep' ratio can exceed 1.0, so it is not a bounded retention fraction. _box_edges(..., gain) multiplies every luma jump by gain = mean_luma(A)/mean_luma(B) (measured 1.12 at x=320, 1.37 at the edge poses). That promotes sub-threshold gradients into counted edges, so the concussed frame reports MORE strong edges than the clean one: measured keep 1.16/1.10/1.06 (GL) and 1.13/1.09/1.06 (Metal). The ungraded values printed alongside are the honest ones (0.62/0.99/0.70). Not currently masking anything — both hard controls land at 0.00–0.09, ~25x below the 0.85 floor — but the metric would partially absorb a moderate future smear at an edge pose. Fix shape: scale the THRESHOLD by 1/gain instead of scaling every jump by gain.
MIN_CENTRE_EDGES=150 has only 36 edges of headroom at the thinnest swept pose. _player_box_logical clamps to the frame, so x=40 and x=600 get a 160px-wide box instead of 240. Measured baselines this session: 186 (x=40), 349 (x=320), 265 (x=600). The signal floor of 150 sits 36 under the thinnest. Any change to the boot pose, the desert dressing under the soldier, or the box half-width will push x=40 to INCONCLUSIVE before it pushes anything else. Measured, not speculated — the probe prints centre N->M every run.
probe_concussion_hud.gd leaks at exit (pre-existing, confirmed identical on HEAD). Both the new and the HEAD version print Texture with GL ID ... leaked, RenderingServer::get_singleton() is null, and 1 RID allocations ... leaked at exit (GL) / RendererRD Texture leaked (Metal). I ran HEAD's probe verbatim from git show HEAD:tools/probe_concussion_hud.gd to confirm this predates the diff. Cause is the documented one: the probe never frees _main and never calls await Quiesce.teardown(self, main). It is a GL-only hand-run tool outside CI and outside run_tests.sh, so nothing gates it — the same known gap CLAUDE.md already records for screenshots.gd / biome_capture.gd / gif_capture.gd.
ground_slab_ops caps the 1px STROKE alpha at the PLANE's ceiling, contradicting its own rationale. src/main.gd::ground_slab_ops. The stroke op is built as Color(edge.r, edge.g, edge.b, minf(edge.a, GROUND_SLAB_MAX_ALPHA)), so the lane-seal boundary requested at alpha 0.9 draws at 0.55 and the warn lip at 0.5 draws at 0.5. The seam's own docstring argues 'a boundary LINE is authored signage, a filled PLANE is a debug volume' - capping the line at the plane's ceiling directly costs the zone read for no stated reason and is the cheapest half of the fix for the item above. tests/test_view_honesty.gd arm 2 asserts the cap on EVERY op, so it must be scoped to k=='soft' at the same time (and the non-vacuity control extended with a soft op at 0.92 so it still fails two arms).
The new ground-tint ratchet is structurally blind to the regression it just shipped. tests/test_view_honesty.gd::test_world_ground_tints_are_soft_masked_not_raw_quads arm 2 asserts only: no op has k=='fill', every alpha <= GROUND_SLAB_MAX_ALPHA, the soft box strictly encloses r, and the stroke box equals r at width 1.0. None of those can observe pixel contrast, so the SEALED-vs-SCAR collapse measured above passes the ratchet green. There is no assertion anywhere that the three lane-seal states stay distinguishable. Verified by running the suite on the current tree: PASS, 0 failures, with the regression present.
Two docstrings in the diff cite row counts that contradict the ratchet they name. src/sim/sim_world.gd::choke_band_span says 'all 115 enumerated choking rows (test_world_ground_tints_...)'; src/main.gd::choke_slab_rect says 'Measured over all 115 enumerated choking rows of segments 2-12' and 'drew 35 of those rows fully off screen'. I re-ran that test's own enumeration (segments CHOKE_START_SEG..12, off_u 0..1000 step 10) against the current tree: rows = 285, head_blank = 285, head_off = 285. The other figures in the same comment are correct (head_dx = 24.00 px, head_dy = 128.50 px, head_unmarked = 69,660 px2). 115 and 35 are wrong. The branch's own most recent commit is 'stop the docs' own numbers from rotting'.

New 2026-08-23 — from the `f1601b7..3be8149` window

The R2b ratchet false-reds on an innocent COMMENT that happens to contain a draw-call token — and the failure message accuses the parser. tests/test_view_honesty.gd::test_no_string_is_drawn_under_a_scaled_canvas_transform (:3848 on 3be8149) builds expected from the RAW file (raw.count(tok) over Art.text( / Art.text_center( / draw_string( / draw_string_outline() but builds the walk from a comment-BLANKED copy, then asserts checked == expected. REPRODUCED by mutation: inserting the single whole-line comment # NOTE: prose mentioning Art.text(self, ...) above _update_hud() in src/main.gd turns the suite red with "the transform/text walk reached every string draw in the view (77 of 78) (got 77, want 78)". Restored, verified green. The message points at _xform_args, so the next person will go hunting in the parser instead of deleting a comment. Worse than a nuisance: it is a gate that goes red on a correct tree for a documentation edit, which is precisely the shape CLAUDE.md warns turns a gate into something nobody reads (test_perf.gd → OPT_IN_SUITES). Fix: count expected off the same blanked copy the walk uses.
tools/screenshots.gd still leaks at exit — and this window's diff touches that file. MEASURED on this run's real-GL capture (14 live / 14 unique shots, all SAVED): ERROR: Texture with GL ID of 329: leaked 5460 bytes and ERROR: 1 RID allocations of type N5GLES37TextureE were leaked at exit. It never frees its main and never calls await Quiesce.teardown(self, main), unlike the three gated tools. PRE-EXISTING and NOT caused by the m._result_t = 1.0 line this diff adds — verified by capturing the same shot on the stashed (unmodified) tree and finding the leak lines identical. Banked because this is the first thing in a while to touch that file. Re-verified on 3be8149: grep -ln 'main.tscn' tools/*.gd lists eight files and grep -n 'Quiesce.teardown' tools/*.gd covers only e2e_playthrough, perf_probe and smoke — so four GL/manual tools (screenshots.gd, biome_capture.gd, gif_capture.gd, probe_concussion_hud.gd) are outside the gate, not the three CLAUDE.md's older text names. Cross-ref §1 #32 and the §8 row for a39c1f3.
README's assertion badge reads low — and the correct fix is NOT to retype it, because the number is unfalsifiable. The badge reads tests-1162 methods · 37.4k asserts (README.md:20), and the prose at README.md:338 says "1,162 test methods / 37,400+ assertions". Re-verified on 3be8149: the method half is exact and gated — grep -hE '^func test_' tests/test_*.gd | wc -l = 1165, minus test_perf.gd's 3 opt-in methods = 1162, matching the badge, and d3bc2a1 added a test_assets.gd gate that recounts it (including the URL-encoded copy inside the shields.io URL, which is exactly why it rotted unnoticed while the prose beside it was being corrected). The assertion half is ungated on purpose: bf27b52's build log records that one commit (4c7ae78) measured 37,436 assertions on Linux and macOS, 37,451 on Windows, and 37,473 on a local macOS run of the same tree — so no single figure is correct everywhere, and d3bc2a1 deliberately made assertion counts a FLOOR rather than a pin. 37,400+ therefore remains true; only the badge's flat 37.4k reads as a stale pin. ⚠️ File this as "make the badge say 37.4k+", and do NOT re-open it as a stale number — two previous passes "fixed" it by retyping, which is the failure this entry exists to prevent. (This is the first entry in this file whose counter-evidence is the finding.)

New 2026-08-22 — from the `9bb1cdb..f1601b7` window

The final clampf in claim_label_slot's least-overlap return is DEAD CODE, and its comment overstates it. src/main.gd:10893 on f1601b7: return Rect2(x, clampf(best_y, min_y, maxf(min_y, max_y - h)), w, h). MEASURED by mutation: reverting ONLY that clamp back to Rect2(x, best_y, w, h) leaves the rail ratchet reporting 0 collisions (SUITE=view_honesty PASS). The bound is already enforced twice over — by k_hi := floor((max_y - h - py) / 11.0) and by best_y's initialiser, which is itself clamped to max_y - h. Only when ALL FOUR max_y uses are removed does the ratchet go red (343). The eight-line comment above that return claims the clamp is "what makes the reservation stick", which is not what the measurement says. Harmless belt-and-braces — fix the comment, or drop the clamp and keep the comment on whichever bound actually carries the load.
Stale collision number inside the rail ratchet's own comment (116,564 vs the measured 113,839). tests/test_view_honesty.gd:3605 on f1601b7, in test_no_world_label_ever_lands_on_the_bottom_hud_rail: the comment above the counter-factual block reads "116,564 / 343 collisions is ample proof of detectability". The test itself prints A_no_reservation = 113,839 (read out by forcing the assertion red: "COUNTER-FACTUAL A: with no rail reserved the sweep still sees the defect (113839 collisions)"). 343 is right. Comment-only — the assertion is > 0, so nothing depends on the literal. Pre-existing only in the sense that it arrived with the diff that added the ratchet.

New this run (2026-08-21)

Tests that measure the wrong thing, or cannot fail.

The R2 fly-in ratchet cannot see a sprite/hitbox divorce — it measures a test-local copy of the ramp, not the view. tests/test_view_honesty.gd:2960-2961 hard-codes FLYIN_VIEW_DX := 150.0 / FLYIN_VIEW_DY := -55.0 and _flyin_drawn_px (:2978-2984) computes the "drawn" point from those, never from main.gd. The source half only greps for boss_flyin_offset plus absence of the 150.0/55.0/420 literals. MEASURED: changing src/main.gd:10532-10533 to int(foff[0]) * 2 / int(foff[1]) * 2 — which re-creates a 159.8 px divorce at phase_t -420, the exact defect — leaves SUITE=view_honesty at PASS, 86 methods, 1584 assertions, 0 failures, RC=0. The plan's R2 promised a measured "<= 20 px at every sample"; the test as written structurally cannot produce that number.
…and the only thing that DOES catch the divorce is a whitespace-exact source pin. Measured against the same mutation: exactly ONE assertion went red, and it was the character-exact substring pin on the var apos := _to_screen(...) expression. The "worst drawn-hull vs hit-disc separation" assertion stayed green. So the ratchet is fail-closed, but its protection rests entirely on a two-line string match that a harmless reindent will break (red on a correct tree). Two ways to harden it: derive the drawn point from main.gd itself, or capture the real draw call.
A test docstring ships a measurement its own seed refutes. tests/test_view_honesty.gd, test_flyin_rewards_tracking_not_the_arrival_point docstring: "in the arena the sim actually builds, it lands 0 during the whole approach because cover eats the lane (armor_block from tick 7)". REFUTED on the seed the file's own _flyin_boss_sim helper uses. Driving main.demo_input on SimWorld.new(7, 1, "endless") from tick 0, the miniboss reaches phase_t == 0 with hp 8/40 — 32 damage landed during the approach — and on seeds 11 and 99 it is destroyed mid-approach (phase_t -95 and -16). The assertions in the method are correct and load-bearing; only the prose is false, and it will mislead the next reader about the balance the fix creates.
tests/test_event_coverage.gd does NOT actually cover rooted_spawn — the plan's "free ratchet E" is not free. The plan claimed adding a sim event without a _consume_events arm would go red automatically via tests/test_event_coverage.gd:37. MEASURED: with all three events.append({"t": "rooted_spawn"…}) lines commented out (mutation B), SUITE=event_coverage stayed GREEN (PASS — 1 test method, 17 assertions). That gate only sees event types the determinism torture actually emits, and the torture produces zero rooted births. The diff's own comment states this honestly, and test_archetypes.gd::test_every_rooted_kind_announces_its_arrival covers it instead — so nothing is unguarded. Banked so nobody later leans on event_coverage for a type the torture never fields.
Entry-granular pagination costs leaves and leaves sparse pages — no ratchet pins the density. Measured with the real GameMenu (Art.text_scale swept 125/150/175/200), fill = summed row height / capacity per leaf. Regressions: 125% WAR CHEST 2 leaves (93%/93%) → 3 leaves (73%/99%/14%); 200% WAR CHEST 5 (100/94/94/94/19) → 6 (85/34/90/49/90/52); 200% SPECIALS 4 (95/98/98/98) → 5 (80/49/64/98/98). Several others rebalance for the better (175% MODES 90/90/17 → 77/90/30; 125% ENEMIES 96/14 → 75/34), and 200% CONTROLS holds at 4 leaves. A defensible trade for readability, but nothing in the suite pins leaf-count growth or a minimum fill, so a future copy edit can quietly double the manual's page count.

Test/tool comments that document numbers which do not reproduce.

tests/test_menu_layout.gd c4-19 header: 26 offending leaves across 16 combos, not "20 of 20". The header above test_manual_never_ends_a_leaf_mid_sentence says "Measured at HEAD: 20 of the 20 (tab × enlarged scale) combos, 20 offending leaves, across all 5 tabs." Reverting _howto_large_pages to the line-greedy loop (leaving _ends_sentence and the provenance keys in place, i.e. exactly HEAD's pager) and running SUITE=menu_layout reports 26 offenders spanning 16 combos — 125% WAR CHEST/MODES/ENEMIES (3), 150% WAR CHEST(×2)/ENEMIES/SPECIALS (3 combos), 175% all five tabs (2 leaves each), 200% all five tabs. CONTROLS and SPECIALS are clean at 125%, so "all 5 tabs" at every rung is not what HEAD does.
tests/test_assets.gd:2632-2633 documents HEAD period-64 fundamentals that were never true of this code. Header states "desert (march 0.0): fundamental 0.795 (x) / 0.798 (y)" and "foundry (march 1.0): 0.819 / 0.792". Re-measured by mutating src/main.gd's generator back to the % 33 jitter with the 96px pass deleted and running the shipped test: 0.617/0.620 (desert), 0.623/0.613 (foundry) — printed by the test's own failure messages. src/main.gd's inline comment on the same edit already carries the correct 0.617/0.620, so the two files disagree with each other. The card-count consts (_HEAD_DESERT_CARDS 3224, _HEAD_FOUNDRY_CARDS 7548) ARE correct — only the fundamentals are wrong.
tests/test_main.gd:2778 documents HEAD arbiter violation counts that do not reproduce. Header states "MEASURED ON HEAD (7c16776): 1,260 violations at 100%, 1,534 at 200%". Restoring the old return Rect2(x, clampf(rect.position.y, min_y, maxf(min_y, 360.0 - h)), w, h) exhaustion path and running the shipped test gives 1,506 @100% and 1,742 @200%. src/main.gd's comment on the same fix already says 1,506/1,742. The 1,260/1,534 figures came from the plan's scratchpad prototype, not from this tree.
tests/test_assets.gd:2637-2639 banks "this environment has no GL context" — FALSE. The comment states the rendered-composite autocorrelation "CANNOT be ratcheted here" because "this environment has no GL context". Running SHOT_DIR=… /Applications/Godot.app/Contents/MacOS/Godot --path . --rendering-method gl_compatibility -s res://tools/screenshots.gd in this exact working directory printed "OpenGL API 4.1 Metal - 91.7 - Compatibility - Using Device: Apple - Apple M4 Max" and saved all 14 shots (colors 949–5696, stddev 22.5–54.6, "ALL SHOTS DONE — 14 live, 14 unique"). tools/perf_probe.gd also rendered 121 frames under forward_plus. The correct bank is "CI's Linux runner would need xvfb", not "no GL here". As written it instructs the next cycle not to look at the screen.
tools/ground_lag.py's documented BEFORE numbers are ~0.01 low against a true HEAD render. Docstring: 01 +0.138/+0.128, 02 +0.147/+0.138, 04 +0.085/+0.094. Re-measured with src/main.gd stashed back to 7c16776 and re-rendered: 01 +0.148/+0.144, 02 +0.155/+0.151, 04 +0.094/+0.098. The AFTER column is exact to the thousandth and 05-foundry is right, which suggests the BEFORE column was taken with the dressing fix already applied rather than at clean HEAD. The table is presented as the before/after of the pitch change alone.
tools/ground_lag.py's docstring mislabels the lag-192 residual axes. It reads "lag-192 … still carries +0.058 x / +0.055 y on 01". Measured: 01 is x=-0.004, y=+0.058 and 02 is x=-0.013, y=+0.055 — both quoted figures are y-values from two DIFFERENT shots, presented as an x/y pair for a single shot. The immediately preceding "down from HEAD's +0.136 / +0.247" IS a genuine x/y pair for 01 (measured +0.136/+0.248), so the sentence switches convention mid-clause. Every other figure in that docstring reproduced exactly on independent renders, which is what makes this one worth fixing rather than ignoring.
tools/ground_lag.py's control lags are frozen at (52,56,72,76) regardless of --lag. CTRL_LAGS is a module constant, so --lag 96 and --lag 192 measure their excess against a floor sampled near lag 64. Residual autocorrelation decays with lag, so the floor is slightly HIGH for lags above 64 and the reported excess is understated — conservative for every claim the diff makes (it makes the after-numbers look worse, not better), but the instrument does not track its own argument and a future reader will assume it does. Either derive the control lags from lag (e.g. lag ± 8, 12) or say so in the docstring.

Uncalled, unwired or broken instruments.

tools/ground_lag.py joins tools/ground_profile.py as a SECOND GL-only pixel probe with zero callers and no CI job. Re-verified on 9bb1cdb: grep -rn ground_lag . --exclude-dir=.git matches only the file itself and two test comments; ground_profile has zero references anywhere. .github/workflows/ci.yml runs only --import, lint_sim, lint_assets, i18n_check, smoke, run_tests and the export/soak jobs — grep -n 'gl_compatibility\|xvfb' .github/workflows/ci.yml returns nothing. Both are honest instruments whose numbers reproduced, but the repo now has two uncalled rendered-pixel measuring tapes and no xvfb job to run either. (Owner decision #32.)
~~tools/probe_mg_lane.gd is still committed and still broken~~ RESOLVED: deleted (with its .gd.uid sidecar). tests/test_view_honesty.gd already carries the correct, position-based attribution this probe was trying (and failing) to do via enemy_bullets[-1], so it was a redundant instrument whose numbers a green test refuted — nothing referenced it outside its own history comment.
Three GL-only capture tools still print leak diagnostics and no gate sees them (PRE-EXISTING; CLAUDE.md's updated run_tests.sh bullet names it as a known gap rather than claiming it fixed — this is a bank, not a lie). Measured: godot --path . --rendering-method gl_compatibility -s res://tools/screenshots.gd prints ERROR: Texture with GL ID of 321: leaked 5460 bytes. and ERROR: 1 RID allocations of type 'N5GLES37TextureE' were leaked at exit.; the implementer's identical HEAD and post-fix runs both additionally show WARNING: 2 ObjectDB instances were leaked at exit (one reviewer's run did not — it is racy). tools/probe_concussion_hud.gd prints the same two GL lines (verified alongside PROBE OK). tools/biome_capture.gd is unmeasured. Corroborated structurally: grep -ln 'main.tscn' tools/*.gd → biome_capture, e2e_playthrough, perf_probe, probe_concussion_hud, screenshots, smoke, quiesce; grep -ln Quiesce tools/*.gd → e2e_playthrough, perf_probe, smoke, quiesce. screenshots.gd / biome_capture.gd / probe_concussion_hud.gd contain no free() or queue_free() at all and quit holding their main. Neither is in CI (both need a GL context) and neither can be run through tools/run_tests.sh headless, so no gate sees them.
tools/run_tests.sh HANGS FOREVER (>10 min, no output, low CPU) on an incompatible method-override return type instead of erroring. Hit while gate-proving a ratchet the prescribed way (stash the diff and run it): with src/view/menu.gd stashed to HEAD, tests/test_menu_layout.gd's _CtrlLayoutMenu._verb_line(segs, base_y, col) -> float overrides a base _verb_line(...) -> void, and SUITE=menu_layout tools/run_tests.sh produced ZERO output and had to be killed at the 10-minute timeout. NOT a defect of the shipped tree — on the real tree that same suite runs in 1.3s and PASSes 215 methods / 8,615 assertions. But it is a live footgun: the symptom (one Godot process, no output, indefinite runtime) is indistinguishable at a glance from the cold-import and CPU-contention stalls CLAUDE.md documents. This is a FOURTH stall class and belongs in CLAUDE.md's stall triage list.

Cosmetic / style.

Single blank line between an extracted function and the next. src/main.gd:10690 — _boss_bar ends and static func _boss_wound_scars begins after ONE blank line; the file's convention throughout is two. Introduced by the 016b4cd extraction. No parse or test impact.
main.gd:10625 comment still says the boss bars sit "below the corner HUD panel's max height (~60px)". The comment above _draw_one_gunship's bar placement predates 66d146a and describes the geometry the fix corrected. It is now incidentally true (panel_bottom max 60, BOSS_BLOCK_TOP 66) but for the wrong reason — it reasons about the BAR line clearing 60, which is exactly the mistake that let the panel bury 7px of the boss name. The next pass should rewrite it to reference BOSS_BLOCK_TOP. (Not re-verified against 9bb1cdb line numbers — the surrounding code moved.)
CLAUDE.md's Commands block still advertises the ungated raw suite invocation. The run_tests.sh bullet deep in Conventions describes the shutdown-leak gate, but the Commands block near the top still lists godot --headless --path . -s res://tests/run_tests.gd as THE "Full test suite" command, with tools/run_tests.sh presented only as the parallel-safe variant. A reader following the first command gets the suite with no leak gate at all — the same structural blindness the gate exists to remove, since _gate_engine_errors reads user://logs and the leak lines are emitted after the file logger is torn down. One sentence in that block closes it.

Resolved in the `390c12d..9bb1cdb` window — kept with how they were re-verified

~~e2e_playthrough leaks 30 ObjectDB / 1 resource at exit (PRE-EXISTING on HEAD) — and the new gate hard-fails it.~~ RESOLVED — 66d146a. (Was: 3 of 4 runs printed WARNING: 30 ObjectDB instances were leaked at exit and ERROR: 1 resources still in use at exit AFTER [E2E] === 80 checks, 0 FAIL ===; verified pre-existing by stashing the whole diff and running HEAD 5× — all 5 printed the identical 30/1. --verbose named all 31: 16 AudioStreamPlaybackWAV, 10 AudioStreamWAV, 2 AudioStreamPlaybackPolyphonic, 1 AudioStreamMP3, 1 AudioStreamPlaybackMP3, plus Resource still in use: res://assets/vo/cmd/cmd_levelstart_6.mp3.) The shipped fix is tools/quiesce.gd: measured 30 → 0, 6-of-6 runs.
~~tools/run_tests.sh -s res://tools/smoke.gd exits 1 on 5 of 6 runs without LEAK_FLOOR=2.~~ RESOLVED — 66d146a. (Was: 6 consecutive runs — run1 exit 0, runs 2-6 exit 1 with 2 ObjectDB instances leaked at exit (floor 0); the 2 objects are a genuine engine artifact, reproduced 1-per-polyphonic-player in 14 lines of vanilla Godot 4.7.2 with no repo code.) The shipped answer was not a floor but a teardown: smoke.gd routes through Quiesce.teardown and measures 2 → 0, 6-of-6. Re-verified on 9bb1cdb: LEAK_FLOOR does not exist in tools/run_tests.sh (grep -n LEAK_FLOOR tools/run_tests.sh → no match), so the escape hatch the finding named never shipped.
~~CLAUDE.md not updated for the new LEAK_FLOOR lever.~~ RESOLVED — 66d146a. Re-verified: CLAUDE.md's run_tests.sh bullet now documents the shutdown-leak gate, both leak causes, the --verbose discriminator, Quiesce.teardown, and names the three ungated GL tools as a known gap. There is no LEAK_FLOOR to document.
~~The new leak gate misses Godot's RID_Owner leak string and the GL byte-leak string.~~ RESOLVED — 66d146a. (Was: all three copies matched only RIDs of type .* were leaked, missing ERROR: 1 RID allocations of type 'N5GLES37TextureE' were leaked at exit. and ERROR: Texture with GL ID of 321: leaked 5460 bytes.) Re-verified on 9bb1cdb: tools/run_tests.sh:86 matches 'were leaked|resources still in use at exit|leaked [0-9]+ bytes', and the header comment at :81-83 states outright that "Godot words RID leaks two different ways … so the pattern matches the shared substrings, not one sentence."
~~Vacuous assertion in the new test_hud ratchet — x == x cannot fail.~~ RESOLVED — 66d146a. (Was: test_every_boss_plate_owns_its_reserved_band asserted Art.tw(label, SIZE) == Art.tw(label, SIZE) at a single scale, so it passed on any implementation — including one where TEXT SIZE does reach the labels.) Re-verified by reading tests/test_hud.gd:3513-3523: it now captures label_w1 at text_scale 1.0, label_w2 at 2.0, restores the scale before asserting, and asserts eq(label_w2, label_w1) alongside the live Art.fs(10) == 20 arm.
~~hud.gd:1706 # 38.0 in the finale is now 55.0.~~ RESOLVED — 66d146a. (Was: bottom_band_lift measured 55.0 on that tree — 344 + 8 - (COLOSSUS_BLOCK_TOP 305 - BOTTOM_RESERVE_GAP 8) = 297 → 55.0 — while annotated 38.0; it was 48.0 on HEAD, so the comment was already wrong by 10.) Re-verified: grep -n 'in the finale' src/view/hud.gd → no match.
~~main.gd:12390/12393 mirrors LAST_STAND_Y as a bare 350.0.~~ RESOLVED — 66d146a. (Was: var oy := 40.0 if is_top else minf(350.0, HudIcons.band_bottom(sim)) under a comment claiming "350 is LAST_STAND_Y", after LAST_STAND_Y moved to 353.0. Blast radius measured as none: the +N more overflow label right-flushes near x=606 while the LAST STAND plate is centered at x 170..470, so no rect intersection at either 350 or 353.) Re-verified: grep -n 'minf(350.0' src/main.gd → no match.
~~claim_label_slot's own docstring still documents the behaviour the fix removed.~~ RESOLVED — 9bb1cdb. (Was: the leading ## block still claimed "a PERSISTENT label … keeps its place, x-clamped, rather than being shoved off-screen" — exactly the return the fix 60 lines lower deletes.) Re-verified at src/main.gd:10684, which now reads settle for the row of LEAST total overlap. It no longer "keeps its place, …. Still open: tests/test_main.gd:1969 repeats the stale claim inside test_world_text_saturation_drops_instead_of_overprinting's docstring — that test still passes and is still non-vacuous (drops > 0 asserted and green), only its prose is wrong.
~~Dangling doc reference: GROUND_BASE_SHUFFLE does not exist.~~ RESOLVED — 9bb1cdb. (Was: src/main.gd:1153 said "eight 128px slots, each a different dihedral transform of sand.png (see GROUND_BASE_SHUFFLE)" while the constant is GROUND_BASE_SLOT_DIHEDRAL.) Re-verified: grep -rn GROUND_BASE_SHUFFLE . --exclude-dir=.git → no matches anywhere.
~~Stale doc: strip builder says "4 × 1024x128 blits", it is 64 blits across 8 variants.~~ RESOLVED — 9bb1cdb. Re-verified at src/main.gd:1156, which now reads "Builds one 1024x128 base strip: eight 128px slots, each a different dihedral…". The useful number from the original finding, kept: the whole build measures 11.71 ms for 4.00 MB of raw pixels, once per process.
~~Fly-in boss bar names an act the sim refuses to perform.~~ RESOLVED — 016b4cd. (Was: main.gd:10554 passed a NEGATIVE phase_t into _boss_bar; :10668's gphase := 1 if pt < SimWorld.BOSS_STRAFE_TICKS else 2 selected GUNSHIP_PHASE_NAMES[0] == 'STRAFING RUN' for the whole approach. MEASURED by posing phase_t -420 / -300 / -180 / -60 through the real view (main.tscn, gl_compatibility, 12 settle frames) and reading the plate in all four PNGs: top center read "GUNSHIP — STRAFING RUN" for the full 420 ticks = 7.00 s, while _step_one_boss returns before any firing at negative phase_t.) Re-verified: src/main.gd now defines const GUNSHIP_INBOUND_NAME := "INBOUND" with a comment naming the exact defect.
~~Mid-approach kill puts the explosion, coin toast and ejecting pilot on the landing pad instead of on the wreck.~~ RESOLVED — 016b4cd. (Was: _damage_boss's death block computed var by: int = boss["gate_y"] - BOSS_Y_OFFSET and emitted explosion, kill and pilot_down at the ARRIVAL point with no boss_flyin_offset. MEASURED by direct _damage_boss probe across the window: phase_t=-420 hull=(470.0,-365.0) explosion@(320.0,-310.0) d=159.8px, pilot d=214.3px; -300 → 114.1px; -158 → 60.1px; 0 → 0.0px.) Re-verified at src/sim/sim_world.gd:6910-6927, which now computes var doff: Array = boss_flyin_offset(boss.get("phase_t", 0)) and offsets bx/by before every emit.

Carried from earlier snapshots

Engine shutdown leak ERRORs on every suite run (pre-existing, identical on HEAD; re-flagged 2026-07-31). Every suite run prints ERROR: 3 resources still in use at exit plus WARNING: 4383 ObjectDB instances leaked at exit, after the PASS line. Verified pre-existing: stashed the entire diff (tracked + untracked), re-ran the suite on HEAD, and got the identical leak counts (4383 ObjectDB, 3 resources). Not from the new CanvasLayer. These fall outside the engine-error gate's window (the gate reads the log up to its own marker; these print at shutdown), so they never fail anything — but they are ERROR lines every run trains people to ignore, which is exactly how a real one gets missed.
NEW 2026-07-31 — tools/screenshots.gd victory pose fakes the WAR CHEST row. _shot_victoly/_dress_victoly never sets _victory_banked/_victory_banked_score, so the harness's 06-victoly.png renders "0¢ WAR CHEST BANKED → +0" — the narrowest possible version of the card's widest row. Confirmed in this cycle's own capture (/tmp/cic-cycle5/06-victoly.png). This is the exact "harness lying about the game" pattern _dress_victoly's own comment guards against for kills/streak, and it means the signature victory shot can never exercise the row the trophy used to cover. The plan listed this as optional hardening ("take it or leave it") and the implementer left it. Pre-existing, verified by capture. Two lines in _dress_victoly would fix it.
NEW 2026-07-31 — Toast-string scrape's _coin_pop branch is dead code — coin toasts (BOUNTY/RANSOM/+¢) never captured. tests/test_main.gd::_shipped_floattext_strings: for lines containing _coin_pop(, line.find("\"", at) lands on the quote in ev["x"] (the first argument), producing fmt="x", which the fmt.length() < 2 filter then drops — so the line is consumed without ever reaching the actual text literal. Verified by simulating the scrape in Python against src/main.gd: 21 strings captured, all from the "rate" branch; BOUNTY +%d¢, RANSOM +%d¢, +%d¢ absent. The helper's docstring ("the _coin_pop text arg") overclaims. Not blocking: coverage is still wide (21 shipped strings incl. wider ones than BOUNTY's 94.5px — "PILOT DOWN — REACH HIM" is 247.5px at scale 1.0) and the size>=4 guard plus the drops>0 a… (report text truncated at source)
tools/probe_concussion_hud.gd is gated nowhere — see Owner decision #1; the tooling half is that the probe exists and nothing runs it. (Also: tools/probe_salvage.gd and tools/probe_colossus.gd are untracked orphans in the tree — they produced this run's measurements, so either commit them or delete them; don't leave them to rot.)
tools/probe_frame_bounds.gd reports 44 false [OVER] lines — every one the footer legend strip, which the real test correctly excludes and the probe does not. Its output contradicts a green test, so it will mislead whoever runs it next. Fix or delete.
test_wheel_plate_masks_the_terrain really pins alpha ≥ ~0.72, not 0.93 — the assertion is weaker than it reads.
The "two-scale" half of check 1b is vacuous — it tests the helper, not the call sites.
The content-well ratchet cannot see intra-content collisions.
test_menu_layout.gd:1840 pins HALL_RECENCY_Y against a hardcoded 66.0.
src/view/menu.gd:4852 comment is now stale and untested.
src/main.gd mg_nest draw-block comments still say the lane is LOCKED — no longer true.
No mg_nest pose exists in the capture harness, so the visual reviewer can never see that fix.
New probe scripts have been committed without their .gd.uid sidecars (repo convention).

6. Process / infrastructure

NEW 2026-08-23 — Turning god_mode OFF was the single highest-yield change this run made, and it should be the default for any behaviour lens from here on. The 2026-08-21 snapshot discovered that _god_restore refills the magazine every 60 ticks against an 8-tick fire loop (entry below) but did not re-run anything under the corrected instrument. This window did, and the ammo-economy findings that fell out — the Triple Shot surcharge pricing a choice the game deleted, the riot shield's flank, the MG nest's off-screen quarter — were invisible to every previous pass, not missed by them. Standing instruction: an economy, attrition or weapon-pressure claim measured under god_mode is measuring a different game. State the flag's value in the entry.
NEW 2026-08-23 — Cite symbols, not line numbers. 3be8149 alone moved sim_world.gd by ~+130 lines below the revive block and main.gd by ~+450 below the HUD block, so a meaningful fraction of the citations merged this run were already stale on the day they were written. Three spot-checks are in the header warning. The cost is not cosmetic: a line-number citation that lands on plausible-looking neighbouring code is the same failure mode as reading the wrong worktree, and it is silent. Prefer grep -n '<symbol>' output in an entry over a bare file:line.
NEW 2026-08-21 — The debug auto-restore refills ammo AND grenades to cap every 60 ticks, so every review pass this project has run has been blind to its most common failure state. WHERE: src/sim/sim_world.gd:1988-2004 (_god_restore: on every tick_count % GOD_RESTORE_TICKS == 0 it clears wiped, respawns the down, and then unconditionally sets p["mg_ammo"] = MG_AMMO_MAX and p["grenade_ammo"] = GRENADE_AMMO_MAX for every player) · :257 (GOD_RESTORE_TICKS := 60). The flag's own docstring is careful about this — "Ammo is topped up on the same heartbeat and NOT continuously: between beats you still run dry and still hear the empty-mag click (the pressure stays observable)." Measured, that is not what a 60-tick heartbeat does against an 8-tick fire loop: the gun burns 7.5 rounds between refills and the refill returns the magazine to cap. Consequence for this backlog: every god_mode measurement in it — and there are many — was taken in a world where the player never runs out of ammo. Re-read the ammo-economy entries in §2 with that in mind, and prefer god_mode OFF drives for any economy or attrition claim.
NEW 2026-08-21 — A FOURTH headless-stall class exists and is not in CLAUDE.md. An incompatible method-override return type makes tools/run_tests.sh hang >10 min with zero output at low CPU, indistinguishable at a glance from the three documented stalls (CPU contention, cold import, aborted _init). See §5 for the full measurement. CLAUDE.md's stall triage paragraph should grow a fourth bullet, because the wrong remedy (pkill -f Godot) is the documented answer to a different one.
NEW 2026-08-22 — Three hung headless Godot probes from a previous session were still spinning when this run's gate started. Hygiene note, not a diff defect, but it is the third time the parallel-session failure modes in CLAUDE.md have bitten. ps showed PIDs 72354 / 73917 running -s res://tools/probe_bottomrail… for 67 and 65 minutes, and 90405 running -s res://tools/_probe_geo.gd for 50 minutes, all at 0.6-0.7% CPU — the aborted-_init signature CLAUDE.md documents, not slowness. The probe scripts themselves are no longer in the tree (git status clean), so these were orphaned by a delete. All three were killed before anything was measured; leaving them would have starved the suite and rotated the engine log out from under the error gate. Standing lesson: ps aux | grep -c '[G]odot' before a measuring run, and deleting a probe script does not reap the process running it.
NEW 2026-08-21 — 96 findings banked against ~20 shipped in this window. The ratio is still going the wrong way, and this snapshot is the largest single merge the file has taken. The drain-mode proposal below is now three snapshots old and still not built.
A 20-cycle /triple-a-game run costs ~50 hours at the current all-Opus/high settings (measured: cycle 1 = 2h29m; the skill's own estimate of 12–20h predates the model change). Decide whether that is the right trade or whether fewer cycles is better value.
The endless bot plateaus around wave 11 even with the wedge fix — deep-wave endless remains partly unobserved, which caps how deep the difficulty telemetry can see.
Backlog grows faster than it drains. One cycle drains ~1 item while the reviewers bank several; this is a triage queue, not a work queue, and it will not converge on its own. A dedicated drain mode (skip the visual reviewer, spend every slot on banked findings) is the obvious answer and is not built. (2026-07-31 was a drain-flavoured run: 10 shipped, 31 new banked — the ratio moved the wrong way anyway.)

7. Blocked / can't be fixed right now

Real, understood, and deliberately not actionable today. Each says WHY, so nobody re-derives the reasoning and nobody mistakes them for oversights.

The "in-rock rooted enemies stall endless waves for 30,000 ticks" causal claim is REFUTED — LEAVE THE STALL ALONE, the premise was wrong. The 4d704cc goal statement said in-rock rooted enemies are "what stalls endless waves for 30,000 ticks". MEASURED, 8 seeds × 30,000 ticks, endless, god_mode, main.demo_input, longest single-wave duration per seed: BEFORE (HEAD sim) median wave reached 10, WORST wave-duration 26,151 ticks; AFTER (fixed sim) median wave reached 10, WORST wave-duration 26,151 ticks. Per-seed it is noise in both directions (seed 5: 3,893 → 19,963 worse; seed 0xC0FFEE: 13,785 → 4,692 better) — the distribution does not move. Diagnosed the worst case directly: seed 1 holds wave 7 open for 26,150 ticks with exactly two ghillies at (75,-308) and (142,-308), and on the FIXED tree both measure CLEAR of all four cover families. The in-rock defect was real and worth fixing on its own terms; the stall it was blamed for is the BOT, and the fix did not touch it. Blocked on: a better bot, or a human playing deep endless. (This is the canonical example of the standing warning at the top of this file — a real smell, consequence wrong by a whole class.)
Miniboss tier escalation stops at wave 20 — LEAVE ALONE, premise was wrong. Escalation does not stop: miniboss HP is uncapped and linear (+20 HP per 5 waves, forever). Two cadence curves floor — mortar shells at tier 3 (wave 15, not 20) and spray interval at tier 4 (wave 20) — and both are explicit maxi() floors, i.e. deliberate. README.md:133 already scopes its promise to "extra mortars by w15", which is exactly what the code does, and tests/test_boss.gd:119 pins the cap. Adding a T4 means reversing a tested design decision, and the mortar act is already spacing-compressed (T3 is +40/+40/+40/+20, so a 6th shell fits only at ~295 with 4 ticks of act left). Blocked on: wanting it at all.
BOSS_HIT_RADIUS 20 vs the gunship silhouette — OWNER CALL, not a bug. The reported "47px half-extent" is rotor blades: four 7px blades reaching the canvas edge on m_heli_attack2, which gunship_body does not have. Eroded, the two hulls measure 30.92 vs 31.14 drawn px — 0.7% apart — so there is nothing to split with a per-boss radius. Raising the radius makes both bosses easier to hit; that is balance, not a fix. Also note the view already draws its hit bloom and damage art at r=34 against a 20px hit disc. Blocked on: a difficulty decision.
Deep-wave Endless is still largely unobserved. The scripted bot plateaus around wave 11 even after the wedge fix, so the difficulty telemetry cannot see past it and wave-20 behaviour has never actually been watched by anything. Blocked on: a better bot, or a human playing it.
The /triple-a-game ledger prompt is 12,272 chars and grows with the backlog. Measured 2026-07-26: ledger.json is 45,996 chars, of which 36,547 are backlog detail (mean 1,142, max 3,807). That whole payload is embedded in a prompt every cycle and handed to a model whose only job is to reproduce it byte-for-byte — which is exactly the setup that produced the documented title-drift corruption. Capping stored detail at ~400 chars would cut it 69% and now costs nothing, since full text lives in this file. Blocked on: nothing — just not done yet. (Do not "fix" it by paraphrasing entries; the titles are dedupe keys.)
tools/probe_frame_bounds.gd ships knowing it emits 44 false [OVER] lines. The gate measured this, filed it, and committed anyway. The gate now blocks on it going forward, but the already-committed probe is still in the tree. Blocked on: someone fixing or deleting it.
A 20-cycle run now costs ~40-50 hours at all-Opus/high (measured: cycle 1 = 2h29m, cycle 2 > 2h25m and still going). The skill's own 12-20h estimate predates the model change. Blocked on: deciding whether that is the right trade.
Miniboss fly-in airstrike window is a partial whiff — banked by the plan. The intermission half is provably empty and shipped in 8425bd7; the fly-in half is not provably empty (the trickle keeps spawning) and _fire_mission spares bosses, so a strike bought during the fly-in still kills only trickle infantry. Blocked on: a design call on whether fly-in strikes should hit the boss. (Mirrored in §2 because the lens keeps re-flagging it.)

8. Shipped — resolved, kept for the record

Shipped in the 2026-08-23 window (`f1601b7..3be8149`, 6 commits)

Every sha below was read directly out of git log f1601b7..3be8149 in /Users/shoemoney/Projects/commander-in-chief on branch main.

Finding	Commit
Endless 2P: a broke death respawns FREE after 5 s whenever a partner is up — the mode's only death brake (the compounding revive price) is waived exactly when you can't pay, so spend-to-zero is the dominant strategy	`3be8149` — the WAIT now compounds instead: `broke_wait_ticks()` rides the same `deaths` curve the price does, capped at `BROKE_WAIT_MAX_MULT` (20.0 s). Measured pre-fix, 2P endless, chest 0: the free rally took 300 ticks and the paid self-revive took 1, both landing at `_checkpoint_y()` with the same stripped body — the coin bought 299 ticks and nothing else, at wave 1 and wave 60 alike, while the bill it dodged climbed 50 → 7800. Keyed on `deaths` only, never on wave. The rally itself stays (a downed player is never a spectator with zero actions) and the WIPE clock is re-checked live and never compounds. CAMPAIGN untouched, which is also what keeps GOLDEN still: the 60 s campaign torture arms this 11 times at deaths 4..12 and every one stays at `BROKE_RESPAWN_TICKS`. Both how-to surfaces derive both ends of the clock off the consts. Three balance questions it deliberately left open are §1 #37–#39
~~`tools/versions.lock` pins CI to 4.7.1 while local dev runs 4.7.2 — the lock's own justification is false~~ (owner decision #36)	`cebc3b7` (bump) + `4c7ae78` (close the backlog item). Verified before moving the pin: full suite PASS — 1157 methods, 37473 assertions, 0 failures; determinism 7 methods, GOLDEN and ENDLESS_GOLDEN reproduce UNCHANGED; `lint_sim` OK; `lint_assets` OK (322 PNGs); `i18n_check` in sync (63 keys, fr + ja). The sha256 was authenticated against the release's published SHA512-SUMS.txt before being hashed. Commit body commits to reverting rather than re-recording if any of the three CI OSes moves a golden
README/CONTRIBUTING test-count line had gone stale a THIRD time (1154 → 1159) and two prior passes "fixed" it by retyping	`d3bc2a1` — three gates added to `test_assets.gd`, each verified to FAIL on a planted wrong value: README's method count recounted from `tests/`; the shields.io badge's URL-encoded copy (`1159%20methods`) — the reason it rotted unnoticed, since `grep "methods"` never matched it; and lockstep still has ZERO production callers, the single load-bearing fact behind "online co-op is not playable" in three docs. Assertion counts became a FLOOR, not a pin. Also corrected two claims against code rather than against the previous doc: CLAUDE.md's "latency jitter only: no packet loss, no duplication" was false (`FakeWire` schedules `drop_every` and `dup_every`, and `test_survives_packet_loss` carries a CONTROL proving it can fail), and `lockstep.gd` is ~260 lines, not the 237 pinned in three files. README now teaches `tools/run_tests.sh` over raw `godot`. Suite green at 1159 methods
A doc that UNDERSELLS its own tests — the finding nobody audits for	`d3bc2a1` + `bf27b52`. CLAUDE.md claimed the netcode harness had no packet-loss coverage while `docs/PLAN.md` said it did; the code sided with PLAN.md. A pessimistic doc invites an agent to rebuild coverage that already exists, and it fails silently — there is no gate anywhere for "this doc is too modest". Now recorded as a class in the build log
A README number that could not be stated correctly at all	`bf27b52` — first entry in `blog/`, and the reason §5's badge entry is filed as counter-evidence rather than as a stale number. Reading the `PASS` line out of all three CI jobs for ONE commit (`4c7ae78`) gave 37,436 assertions on Linux and macOS, 37,451 on Windows, 37,473 on a local macOS run of the same tree. The number was never stale — it was unfalsifiable, and no edit to that line could have been correct. Method count IS platform-stable, so it is pinned by a test. Also banked there, because the wrong turns are the substance: zsh `noclobber` silently refusing a heredoc so a "bump" nearly shipped with the pin untouched; a `git push -q` chained to a self-authored `&& echo "pushed $(git rev-parse HEAD)"` that printed a real sha for a push which moved nothing; and CI verified green against the previous commit's run because `gh run list --limit 1` had not caught up

Shipped in the 2026-08-22 window (`9bb1cdb..f1601b7`, 21 commits)

A short drain window. Every sha below was read directly out of git log 9bb1cdb..f1601b7.

Finding	Commit
Owner rulings §1.2 (hazards pause under the wheel), §1.9 (salvage refuses a no-op) and the off-frame label tolerance	`0f3f480` — tolerance landed at 12px, not 24, because 12.0 is the smallest above-anchor extent of any labelled subject's own art; 0-overlap result re-measured and HOLDS across 7,200 frames at 100% and 200% text scale. See the ruling box in §1 for the wired-to-nothing `BUY_WHEEL_OPEN` catch
`screenshots.gd` victory pose fakes the WAR CHEST row	`8005b80` ("stuffs the war chest before the screenshot") + `042a3c7` ("derives the chest instead of inventing a second number") — this is the row the 2026-08-21 merge listed as NOT corroborated; it is now located. Moved out of that table below
Boss flash never reaches the fly-in hull — the body does not react to 32 HP of damage	`43de2f2` ("gunship fly-in hull finally shows its own damage")
`tools/probe_mg_lane.gd` is still committed and still broken	`40ada89` ("retire the mg-lane probe that shot the wrong shooter") — retired rather than repaired
`tools/run_tests.sh` hangs forever (>10 min, no output, low CPU) on an incompatible method-override return type — the FOURTH stall class, missing from CLAUDE.md	`0d4b243` + `208da7b` — the second commit corrects the first: the parse error does shout, it just gets skimmed past
CLAUDE.md not updated for the new leak-floor lever / still advertises the ungated raw suite invocation	`0d4b243` ("prioritize leak-gated test command")
ESCAPING! is the one world label whose off-frame gate reads the TEXT's left edge, not its SUBJECT — and the ratchet exempts exactly that site	`63dddf8` ("finally names the pilot instead of its own left margin") / merge `30f17fe`, whose subject notes the ratchet stops hiding it
Floattext down-stack can overprint the commander bark row at the bottom rail (2026-07-31 remainder under `1ea6508`)	`f1601b7` — HEAD
CONTROLS page sentence-integrity ratchet (owner decision #26's "should get a ratchet")	`9538eca` ("has to keep its sentences whole, not just in bounds") — pins one side of #26; the density side is still unpinned
Verb-chip lift ratchet reaching real pixels (guards owner decision #15's `MIN_HUD_CHANNEL`)	`2f0b3bc`
Tests badge lying by 154 methods and 14.5k assertions; CONTRIBUTING.md stale twice over	`6f8770c`, then `a39c1f3` — CONTRIBUTING.md was rewritten to pin no figure at all ("a date stamp does not stop a number rotting; removing the number does"). Measured at that point: 1154 methods / 37,418 assertions, and `test_perf.gd` has 3 methods, not the 2 the old parenthetical claimed
CLAUDE.md's un-gated `main.tscn` bootstrap count was a closed-form claim that went false	`a39c1f3` — `d743f4a` added a fourth (`tools/gif_capture.gd`, same no-`Quiesce.teardown` class) three hours after the "exactly THREE" line shipped. Now four, and named. Gating them is still banked as owner-decision territory (§1.1 / §1 #32)
`.gitattributes` never declared `gif` binary, so `d743f4a`'s four promo GIFs rode git's NUL-byte heuristic	`a39c1f3` — latent hole, not damage: `git check-attr binary docs/media/gifs/staging.gif` goes "binary: unspecified" → "binary: set", verified both sides, no re-normalisation churn

Shipped in the 2026-08-21 window (`390c12d..9bb1cdb`, 60 commits)

Shas below were located by git log -S<identifier> 390c12d..HEAD against the symbol each fix had to touch. Where that returned nothing, the row says sha not located rather than guessing — and where the predicate still reads unchanged on 9bb1cdb, the row says so, because a title on a run's shipped list is not evidence that the code moved.

Finding	Commit
Engine shutdown leak ERRORs on every suite run (pre-existing, identical on HEAD)	`66d146a` — baseline was 265 RIDs / 5192 ObjectDB / 8 resources on `7bed222`; gated at zero in `run_tests.sh` and twice in `ci.yml`. Owner decision #18 (3-OS exposure) open
Boss HUD overlap with control hints	`66d146a` — folded `BOSS_BLOCK_TOP`/`COLOSSUS_BLOCK_TOP`/`LAST_STAND_TOP` onto `label_plate_rect`; the hand-typed 9px ascent mirror was 2px optimistic and buried the top of the boss name in 3 of 4 configs. Layout cost is owner decisions #15/#16
Rooted enemies spawn inside rocks, become permanently bulletproof and keep shooting	`4d704cc` — the "stalls endless waves for 30,000 ticks" half of the title is REFUTED, see §7
Endless miniboss is drawn and shootable for a 7s fly-in while bullets pass through it	`016b4cd` — closes the 2026-07-31 remainder under `a512bee`. Deletable-approach balance is owner decision #23
Every turret, mast and hidden sniper materialises inside the frame; in Endless the same pixel row every wave	`7c16776` — `ROOTED_SPAWN_Y_SPAN` + `rooted_arrival_alpha`. Mast exclusion is owner decision #27; sandbag-base pop is open (§4)
Truncated tutorial text in Field Manual	`7c16776` — `_howto_large_pages` entry-granular pager. Density cost is owner decision #26
Truncated body copy in Controls manual	`7c16776` — same pager
Repetitive grid-based ground tile seams	`9bb1cdb` — peak moved 64px → 96px at essentially unchanged amplitude, it did not vanish (§4). Chase-further call is owner decision #30
Overlapping floating MAXED world labels	`9bb1cdb` — `WORLD_LABEL_FRAME` off-subject gate; 0 overlaps across 14,400 sampled bot frames vs 46.32%/47.01% at HEAD. Suppression trade is owner decision #31; 12 unwired call sites are a live regression (§4)
CI lint red — two hint strings shipped without catalogue entries	`7bed222`
Trench Gun countdown hidden exactly when it matters; "same fan" wrong on Triple	`5c0f574`
Claymore sells "IT HURTS BOTH SIDES" but its blast spares players	`b5f7608`
Colossus closed core: MG tracers phase through with zero contact feedback	`27d47e9`
Deep-endless veteran armor silently kills the empty-clip bash	`3fd729c` — closes the 2026-07-31 remainder under `f6666b8`
Grenades and every blast kill the cloaked, bullet-immune ghillie	`d5c7931`
The revive's go-to-the-body grammar signposts a run the sim does not require	`34dd43f` ("a rescue finally lands at your side")
A tank driver's/gunner's revive key is a fully swallowed input	`6fa0aaa` ("a swallowed key")
The supply receipt lies on partial stocks (+30 AMMO / +4 GRENADES at the cap edge)	`6fa0aaa` (`BUY_FLOAT`)
Smoke hint says "BLINDS THEIR AIM" while the MG nest keeps firing aimed rounds	`1542b4b`
End-of-run modal card overlays leak active HUD elements (K.I.A. card leaking the boss bar)	`1542b4b`
NG+ HARD lands on the same board with no marker	`08ccda4`
A tank parked on a free supply crate collects nothing and says nothing	`08ccda4`
Grenade-family hits on bosses emit no `boss_hit` feedback event	`08ccda4`
Airstrike hint gates on the 100¢ base price, not the depth-creeped price	`08ccda4` (`SHOP_AIRSTRIKE_COST`)
SUPPLY CALL spends a Commendation on a hidden random table	`1ac09a6` ("first mint teaches spend")
Intrusive knockdown screen smear obscuring gameplay	`34a4037` ("concussion freezes") — likely; verify against the capture before striking
Closed gates pin the camera while rooted nests pile above the viewport	`7c16776` — likely (same spawn-row work); verify against a gate fight before striking
The Spanish commander got cut off mid-brag and CI never noticed	`dc0a4ac`
`.todo-worktrees` unignored — 316MB one `clean -fd` from deletion	`a71c367` (chore)

On this run's shipped list but NOT corroborated in the code — treat as still open until someone finds the commit. For each of these, the symbol the fix would have to touch is unchanged since 390c12d (git log -S… 390c12d..HEAD returns only the two docs commits):

Title claimed shipped	Evidence against
2P ready-up is a unanimous party vote with no tally	`ready_hold` unchanged since `390c12d`; `grep -n 'ready_hold\|READY_HOLD' src/main.gd src/view/hud.gd` on `9bb1cdb` returns NOTHING — still zero view reads. Entry stays open in §2
Daily Run's one attempt refunded by R / QUIT TO TITLE (both phrasings)	`_daily_done_seed` unchanged in this window; `d45132a`'s arm+demote is visible at `main.gd:1724-1726`, so this is the already-tabled 2026-07-31 fix re-listed, not a new one. Remainders stay open
A salvaged or expired tank hulk keeps burning and drawing as cover	`HULK_TICKS` unchanged since `390c12d`
"WAVE CLEARED — SHOP OPEN" fires while a live Spotter is still shelling	`observer_alive` unchanged since `390c12d`
Endless 2P: a broke death respawns FREE after 5s	`rally_is_free` unchanged since `390c12d`
~~`screenshots.gd` victory pose fakes the WAR CHEST row~~	RESOLVED 2026-08-22 — `8005b80` + `042a3c7`. The 2026-08-21 pass was right that nothing had moved in that window; the fix landed in the next one. See the table above
Toast-string scrape's `_coin_pop` branch is dead code	`_coin_pop` unchanged since `390c12d`
The Colossus siege restocks grenades in total silence	sha not located
Five teaching hints stamp hardcoded key letters (E/F/Q)	sha not located — `40087f6` ("rebind labels now show TAP vs HOLD for shared E key") is adjacent but is not this
Rolling in water prints "NEED COINS" (duplicate `match` arm)	sha not located
HUD keybind string clipping in downed state	sha not located
The shop bills full price for air it partially delivers (charge half)	sha not located — `6fa0aaa` fixed the receipt half only
Miniboss fly-in airstrike window remains a partial whiff	still open by the plan's own bank (§2/§7)

Shipped in the 2026-07-31 window (`2bedaf1..390c12d`)

Finding	Commit
Downed-pilot ransom is geometry-locked	`66e57e5`
Endless miniboss's 7s fly-in — grenade-soft half	`a512bee` — bullet pass-through remainder open (§2)
Grenade landing telegraph lies on held throws	`1ea8d25` — telegraph now tracks the real release trajectory
Endless's unbounded difficulty term invisible from wave 13	`f6666b8` — banner/chip shipped; empty-clip-bash consequence open (§2)
Floattext spawn punch pokes ~4.5px above band_floor	`1ea6508` — down-stack bark overprint remainder open (§4)
Endless shop sells a 100¢ airstrike in the window it cannot hit + token-roll whiff	`8425bd7` — fly-in whiff (§2/§7) and hint-gate (§3) remainders open; pre-call trade is owner decision #10
WAR CHEST howto hides the 10x victory conversion	`7e3d174`
DAILY RUN one-attempt refunded by R / RESTART / QUIT TO TITLE	`d45132a` — scouting remainders still flagged (§2); banked-quit and same-board calls are owner decisions #13/#14
Duplicate TRIPLE SHOT capsule prints "CLAYMORES FULL"	`c957313`
Endless milestone shop window is a kill zone	`390c12d` — Spotter-shelling remainder open (§2)

Also on this run's shipped list — titled fixes whose commits predate the 2026-07-29 snapshot (the run's diff window was wider than the last backlog's):

Finding	Commit
Copy-pasted barricade tiling	`5bbfff1` ("vary barricade tile stacking to kill repetitive sandbag look")
Generic modal frames for menus and victory cards	`6ecbf0d` ("programmer-style UI modals and plain menu framing")
UI text banner stacking / UI label stacking / overlapping HUD text popups	`f255cab` + `b439779` (label arbiter + band arbiter)
Menu UI bounds truncation and clipping	`82373e7` (likely — "the price, the recovery word and the recency band all stayed in the box")
Unlit and floating heavy equipment sprites	`f29d7ca` (likely — gunship ground shadow + technical deny pip)
Muddy low-contrast unit silhouette readability	this run's `m_soldier2` rim fix — sha not located this pass (see owner decision #11 for the ghillie/sapper deferral)
Flat unintegrated world UI banners	sha not located this pass (`b18c889` "back the in-world callouts" is adjacent)
Overbearing in-world objective UI labels	sha not located this pass

Shipped 2026-07-29 (previous snapshot)

Finding	Commit
Unmasked Radial Shop Menu Overlay	`ee25263` — lens still flags a remainder (§4)
TRIPLE SHOT hint says "PERMANENT" — the sim strips it on your very next death	`96d8928`
Unaligned Leaderboard Column Formatting	`2bedaf1` — lens still flags a remainder (§4)
Subtitle speaker name duplication	`8e4ea34`
Incorrect biome text string on results card (victory card said JUNGLE, screen showed DESERT)	`0708419`
Mismatched environment stat string on victory screen	`0708419` (likely the same defect as the row above — dedupe on next pass)
Abrupt water tile transitions without shore blending	`a81a8cc`
MG nest's 30-tick aim lane drawn but never fired down, HUD says dodge it	`37e8287`
Planting a claymore in the direction you're aiming kills you 4 ticks later	`2fe8b00`
Gate 4 route fork is a fake choice — one lane holds every reward, the other none	`23019c2` / `0e93888`
40-coin sandbag plants on your firing axis and eats 100 of your own outgoing rounds	`170c8d9`
Buying the same item off the ground pays 67% more score than the wheel	`254b2a9` (crate `cost10` vs wheel `cost6` → both now `SPEND_SCORE_MULT`; measured: 30-coin ammo crate +300 / wheel +180 → both +180)
Endless unspent War Chest worth exactly zero score	`d6c2aa9`
Dying silently deletes an earned Commendation and every carried claymore	sha not located — still not located 2026-07-31
Screen-cluttering world-space text overlays	`f255cab`
Severe UI layering clutter during boss battles	`f255cab` (likely — verify on next pass)
Full-screen low-pass distortion blurring the HUD	sha not located this pass — lens still flags a remainder (§4)
Raw blocky text world signage	sha not located this pass — lens still flags a remainder (§4)
Subtitle box fails to clear or fade out	sha not located this pass

Resolved 2026-07-26 (for context)

bc75b0f sandbag no longer eats the buyer's own rounds · ca3ed3f MG nest leads its target · 3f1c11e MG nest telegraph made honest · the four unclamped gated cooldowns · the VP parade · ten retired entity bakers · the endless bot's aim · fork-gate island coords (0421d8a) · the claymore self-kill (2fe8b00).

9. Notes on how to read this file

A title on a run's "shipped" list is not evidence the code moved. New this snapshot: of the ~120 titles handed over as shipped, 13 could not be corroborated — the symbol each fix would have had to touch is unchanged since 390c12d, and in one case (ready_hold) the defect's headline claim, "zero view reads", still greps to nothing on 9bb1cdb. Those titles are tabled in §8 with the evidence against them and their entries stay OPEN. Locate the sha with git log -S<identifier> <base>..HEAD before striking anything off.
A finding measured against a work-in-progress diff is not a finding against HEAD. Twelve entries this snapshot were filed by lenses reading a branch mid-flight and were fixed before the commit landed (vacuous test_hud assertion, the leak-gate regex, the LAST_STAND_Y/38.0 comment mirrors, GROUND_BASE_SHUFFLE, the fly-in "STRAFING RUN" plate, the mid-approach death coords…). They are marked RESOLVED with the command that re-verified them, not deleted.
Independent measurements of the same number are the evidence that the number is real. Where several lenses reported one decision, this file keeps all of their numbers side by side rather than picking the tidiest — see owner decisions #15, #23, #30, #31. Where they disagree (the miniboss skip rate: 2-of-5, 3-of-7 and 2-of-6 across three lenses), the disagreement is the finding.
god_mode measurements are blind to the ammo economy. _god_restore refills the magazine AND grenades to cap every 60 ticks against an 8-tick fire loop (§6), so any attrition or economy claim in this file taken under god mode is measuring a world where you never run dry. Prefer god_mode OFF drives for those.
A guard nothing can reach is not a guard — new to this list 2026-08-22. 0f3f480 found its own §1.2 hazard-pause sim guard shipped correct, tested and unreachable: BUY_WHEEL_OPEN appeared only in sim_world.gd because the slice that wrote it did not own main.gd, so _update_wheel's held-open path returned a bare 0 and the tested rule could never fire. Same failure shape as the claim_label_slot signature-and-return rect stub CLAUDE.md already documents. When a fix spans the sim/view seam, prove it end to end — the test passing on the sim half is not evidence the view ever calls it.
A number in a comment rots faster than a number in an assertion. Two entries this merge are stale comments beside correct code (the rail ratchet's 116,564 vs the printed 113,839; the clampf comment claiming load it does not carry), and a39c1f3 fixed a third by deleting the number rather than re-stamping it. Prefer printing the measurement from the test over restating it in prose.
Entries marked RESOLVED stay in place so a decision and its outcome stay adjacent. Keep doing that.
The loop writes this file itself at the end of each run and MERGES rather than overwrites. Hand-edits are safe, but expect it to reorganise entries into the same section headings.
The backlog grows faster than it drains (~4 banked vs ~1 fixed per cycle; the 2026-07-31 drain-flavoured run shipped 10 and banked 31). That is what an honest audit of a real codebase looks like — but it means this file is a triage queue, not a work queue, and it will not converge on its own.
Where an entry has no measured number, it says so. Treat those as suspicions, not findings — this project has repeatedly had a real smell reported with the consequence wrong by a whole severity class.
Several report bodies arrived truncated at source (marked "report text truncated at source"). The surviving measurements are verbatim; the missing tails were reasoning, not numbers.
A number can be STALE or it can be UNFALSIFIABLE, and the fixes are opposites — new 2026-08-23. bf27b52 established the distinction by reading one commit's PASS line out of all three CI jobs: 37,436 assertions on Linux and macOS, 37,451 on Windows, 37,473 on a local macOS run of the same tree. Two previous passes had "fixed" that line by retyping it, which could not have been correct at any value. The test is cheap: measure the claim on more than one machine before deciding it is stale. If it varies, the fix is to stop stating it (or state a floor), not to restate it — d3bc2a1 made assertion counts a FLOOR and pinned only the platform-stable method count. §5's README-badge entry is filed under this rule and should not be re-opened as a drift.
A doc that UNDERSELLS its own tests fails silently — new 2026-08-23. CLAUDE.md claimed the lockstep harness had "latency jitter only: no packet loss, no duplication"; FakeWire in fact schedules both, and test_survives_packet_loss ships a control proving it can fail. Nothing gates a pessimistic doc, and the cost is an agent rebuilding coverage that already exists. Audit docs in both directions.
Held, checked, no finding (2026-08-23) — four suspicions chased and REFUTED. Recorded so nobody re-spends the cycle. All measured in the same worktree at bf27b52, with probes that step the sim directly (never awaiting frames), 4 campaign seeds × 12,000 ticks unless noted:
1. NO BLIND DEATHS. Instrumented every player_down and asked whether anything on screen explained it: 133 knockdowns, and ZERO had no visible cause within 200 px (a live enemy inside the viewport), 90 px (an enemy bullet inside the viewport), 60 px (an active strike telegraph) or 40 px (an armed mine). The hit-resolution order in step() (offense before contact death, contact death before enemies move, cited sim_world.gd:1105-1110) is doing exactly what its comment claims.
2. THE CAMERA NEVER SCROLLS BACKWARD. 0 backward ticks (report text truncated at source).
3. (and 4.) (report text truncated at source — the surviving measurements above are verbatim; the missing tail was reasoning, not numbers.) ⚠️ Note the instrument caveat that makes these trustworthy in one direction only: these were taken before the god_mode-OFF discipline in §6 was applied everywhere, so treat #1's "nothing was invisible" as a claim about telegraphing, not about attrition.
Held, checked, no finding (2026-07-31): a full pass over the remaining roster — src/sim/sim_world.gd steppers (elite/sniper/ghillie/drone/technical/mg_nest/grenadier/ observer/colossus/gunship) plus the view telegraph paths (src/main.gd:8708 telegraph_dir, :11920-11981 off-screen threat pips and strike wedges) — found no concealed-difficulty or telegraph lies: every lethal windup is >= the 24t reaction floor, aim-locked shooters fire down the drawn vector (the elite's re-aim cheat is already fixed at src/sim/sim_world.gd:3474-3479), the mg_nest's drawn lane is computed through the sim's own lead function so the painted lane is the lane the round takes, predict_grenade_landing (:2898-2928) replays _step_grenades' exact integrator including the marsh drift and fuse-hand airburst, and off-… (report text truncated at source). Recorded so the next lens doesn't re-audit the same roster cold.
The 2026-07-26 snapshot this file's ancestor replaced was already marked STALE against FINDINGS.md (62 of 115 entries already fixed at re-triage on 2026-07-28). Carried entries above are marked "not re-verified this run" for the same reason: trust the measurement, re-verify the predicate before fixing.

FilesExpand file tree

Backlog.md

Latest commit

History

Backlog.md

File metadata and controls

Backlog

TODO — opening combat and readability follow-up (2026-08-01)

1. Owner decisions — these need a human, not a fix

✅ OWNER RULINGS — 2026-08-21, three of them SHIPPED as code in 0f3f480 on 2026-08-22

New 2026-08-21 — from the 390c12d..9bb1cdb window

New 2026-08-22 — from the 9bb1cdb..f1601b7 window

New 2026-08-23 — from the f1601b7..3be8149 window

New 2026-08-24 — from the wf_a7d3399b-122 run (recovered from .aaa/ledger.json after the run crashed)

2. Sim / gameplay defects — player-facing

RECOVERY (2026-08-24) — accepted work stranded off HEAD, not yet landed

New 2026-08-24 — from the wf_a7d3399b-122 run (recovered from .aaa/ledger.json)

New 2026-08-23 — from the f1601b7..3be8149 window

New this run (2026-08-21), measured unless noted

Carried from the 2026-07-31 snapshot — not re-verified this run, measured unless noted

Carried from the 2026-07-29 snapshot — re-flagged 2026-07-31, still open

Carried from the 2026-07-26 snapshot — not re-verified this run

3. Teaching / content honesty — the game says one thing, the sim does another

New 2026-08-24 — from the wf_a7d3399b-122 run (recovered from .aaa/ledger.json)

New 2026-08-23 — from the f1601b7..3be8149 window

New this run (2026-08-21)

Carried from the 2026-07-31 snapshot — not re-verified this run

Carried from the 2026-07-29 snapshot — re-flagged 2026-07-31, still open

Resolved from this section

Carried from the 2026-07-26 snapshot — not re-verified this run

4. UI / visual polish

New 2026-08-24 — from the wf_a7d3399b-122 run (recovered from .aaa/ledger.json)

New 2026-08-23 — from the f1601b7..3be8149 window

New 2026-08-22 — from the 9bb1cdb..f1601b7 window

New this run (2026-08-21)

Carried from the 2026-07-31 snapshot — not re-verified this run

Carried from the 2026-07-29 snapshot — still open

Resolved from this section

Carried from the 2026-07-26 snapshot — not re-verified this run

5. Test & tooling debt

New 2026-08-24 — from the wf_a7d3399b-122 run (recovered from .aaa/ledger.json)

New 2026-08-23 — from the f1601b7..3be8149 window

New 2026-08-22 — from the 9bb1cdb..f1601b7 window

New this run (2026-08-21)

Resolved in the 390c12d..9bb1cdb window — kept with how they were re-verified

Carried from earlier snapshots

6. Process / infrastructure

7. Blocked / can't be fixed right now

8. Shipped — resolved, kept for the record

Shipped in the 2026-08-23 window (f1601b7..3be8149, 6 commits)

Shipped in the 2026-08-22 window (9bb1cdb..f1601b7, 21 commits)

Shipped in the 2026-08-21 window (390c12d..9bb1cdb, 60 commits)

Shipped in the 2026-07-31 window (2bedaf1..390c12d)

Shipped 2026-07-29 (previous snapshot)

Resolved 2026-07-26 (for context)

9. Notes on how to read this file

✅ OWNER RULINGS — 2026-08-21, three of them SHIPPED as code in `0f3f480` on 2026-08-22

New 2026-08-21 — from the `390c12d..9bb1cdb` window

New 2026-08-22 — from the `9bb1cdb..f1601b7` window

New 2026-08-23 — from the `f1601b7..3be8149` window

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json` after the run crashed)

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

New 2026-08-23 — from the `f1601b7..3be8149` window

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

New 2026-08-23 — from the `f1601b7..3be8149` window

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

New 2026-08-23 — from the `f1601b7..3be8149` window

New 2026-08-22 — from the `9bb1cdb..f1601b7` window

New 2026-08-24 — from the `wf_a7d3399b-122` run (recovered from `.aaa/ledger.json`)

New 2026-08-23 — from the `f1601b7..3be8149` window

New 2026-08-22 — from the `9bb1cdb..f1601b7` window

Resolved in the `390c12d..9bb1cdb` window — kept with how they were re-verified

Shipped in the 2026-08-23 window (`f1601b7..3be8149`, 6 commits)

Shipped in the 2026-08-22 window (`9bb1cdb..f1601b7`, 21 commits)

Shipped in the 2026-08-21 window (`390c12d..9bb1cdb`, 60 commits)

Shipped in the 2026-07-31 window (`2bedaf1..390c12d`)