chore(reports): default the report recipes to the committable form

Plan 59's PR #51 incident: the obvious recipe name (report-agent)
produced the --baseline delta form, which bakes a machine path into the
markdown and fails CI's freshness gate — and a git diff against your own
stale commit reads clean, so the bad form looked verified locally. Fix
the mechanism, not the instance:

- report-agent / report-ios now emit the no-baseline form CI compares;
  delta prints move to new -diff recipes that omit --out, so the
  uncommittable form never touches reports/ at all.
- code-quality-cycle skill: prove a report-touching commit green by
  running the gate's own check script locally, never by eyeballing a
  diff (follow-up pushes don't re-trigger PR CI).
- check-report.mjs stale message: recommend report-ios, not -save
  (re-snapping the baseline mid-PR would absorb a ratchet regression).
- quality-status seeds marked priority-ordered (#1 = next cycle's
  default scope) after the cycle-5 seed deflection.

Plan: 59

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: vinidlidoo

.claude/skills/code-quality-cycle/SKILL.md

@@ -57,7 +57,7 @@ surface inherits its *shape* (report → baseline → assessments → CI gate),
    (summon → `session.start()`, unit-unreachable), that exercise is the **dev-runner selector** (plan 46):
    `just coach-dev`, flip the app's debug pane **Coach worker → Dev**, summon ✨, confirm a response.
    Deliverable = before/after deltas, zero broken tests. Also the **stopping line** (no churn-for-churn).
-   `npm run report -- --baseline reports/viewer.baseline.json`
+   `npm run report -- --baseline reports/viewer.baseline.json` · `just report-agent-diff` · `just report-ios-diff`
 5. **Evolve.** Write `reports/assessments/cycle-NN-plan-XX.md` (table + "where the numbers mislead" + next
    recommendation) and add its index row; **update `plans/quality-status.md`** (status row, verdict, worklist
    seeds); **graduate knowingly-deferred findings into `plans/tech-debt-tracker.md`** (debt lives there
@@ -110,8 +110,9 @@ Coverage records lines **executed, not asserted**, and can't tell unit from inte
 Majors pinned in `TOOLS` (bump deliberately; resolved versions recorded in the JSON so drift shows).
 **Surface-agnostic:** `SURFACES` holds per-surface scope + thresholds; the runners take the binding.
 `slopWaivers` = reviewed false positives, subtracted *and* printed in a "waived" section. **Commit the
-*no-baseline* form** (`report:save` / `report.py --out`) — the `--baseline` delta print embeds a
-machine-specific path and reads as stale on CI. Tool runners **fail loud** (a crashed scanner is an error,
+*no-baseline* form** (`report:save` / `just report-agent` / `just report-ios`) — the `--baseline` delta print
+embeds a machine-specific path and reads as stale on CI; the `-diff` recipes print it without writing to
+`reports/`, so the uncommittable form never lands on disk. Tool runners **fail loud** (a crashed scanner is an error,
 never a green 0), and behavior claims about deps are **enforced in the manifest + lockfile**
 (`uv sync --locked`), not in comments — the freshness gate guards the report, not its inputs.
 
@@ -126,6 +127,10 @@ never a green 0), and behavior claims about deps are **enforced in the manifest
 toolchain-split surface, a **projected** freshness diff over the toolchain-independent fields). **Fails only on
 a target regression**, never a guide. Once gated, the baseline = current hardened state, so feature PRs are
 gated against "don't regress." Minify-on-deploy (`viewer/vercel.json`) realizes the bundle target on the wire.
+Before pushing a commit that touches a committed report, prove the gate green locally by running its own
+check script (`scripts/check-report.*`) against what you're about to commit — never by eyeballing a diff: a
+fresh regen diffed against your own stale commit reads clean. (CI also won't re-run on follow-up pushes to an
+open PR, so local proof is the only proof.)
 
 ## Per-cycle autonomy
 

agent/reports/assessment.md

@@ -12,7 +12,8 @@ does two jobs:
 
 - At cycle start, snapshot `reports/agent.json` → `reports/agent.baseline.json`; at cycle end, re-snapshot to
   the hardened floor (the CI gate enforces it going forward).
-- Run `just report-agent` for the deltas vs the committed baseline.
+- Run `just report-agent-diff` for the deltas vs the committed baseline (local print only); `just report-agent`
+  regenerates `reports/` in the committable no-baseline form.
 - Write the cycle up as `assessments/cycle-NN-plan-XX.md` (full table + "where the numbers mislead" +
   recommendation), and add one row below. **Targets** may be driven; **guides** are read-only (Goodhart) — a
   red guide delta is a prompt to *look*, never an automatic fail.

ios/reports/assessment.md

@@ -13,7 +13,8 @@ two jobs:
 
 - At cycle start, snapshot `reports/ios.json` → `reports/ios.baseline.json`; at cycle end, re-snapshot to the
   hardened floor (the CI gate enforces it going forward). `just report-ios-save`.
-- Run `just report-ios` for the deltas vs the committed baseline (slow — a real `xcodebuild clean test`).
+- Run `just report-ios-diff` for the deltas vs the committed baseline (local print only; slow — a real
+  `xcodebuild clean test`); `just report-ios` regenerates `reports/` in the committable no-baseline form.
 - Write the cycle up as `assessments/cycle-NN-plan-XX.md` (full table + "where the numbers mislead" +
   recommendation), and add one row below. **Targets** may be driven; **guides** are read-only (Goodhart) — a
   red guide delta is a prompt to *look*, never an automatic fail.

ios/scripts/check-report.mjs

@@ -69,7 +69,7 @@ if (checkFreshness) {
   try {
     const committed = JSON.parse(execFileSync('git', ['show', 'HEAD:ios/reports/ios.json'], { encoding: 'utf8' }));
     if (freshnessProjection(cur) !== freshnessProjection(committed)) {
-      failures.push('report stale: committed reports/ios.json disagrees with a fresh run on the toolchain-independent fields (LOC / duplication / complexity / slop / scenarios / testability / structure / dead-code) — run `just report-ios-save` and commit the result');
+      failures.push('report stale: committed reports/ios.json disagrees with a fresh run on the toolchain-independent fields (LOC / duplication / complexity / slop / scenarios / testability / structure / dead-code) — run `just report-ios` and commit the result');
     }
   } catch (e) {
     console.log(`  (freshness check skipped: ${(e.message || '').split('\n')[0]})`);

justfile

@@ -132,21 +132,31 @@ test-ios-ui:
 test-agent:
     cd agent && uv run --group dev pytest -q
 
-# Deterministic code-quality report for the coach-agent surface (plan 43): prints
-# the markdown (via the shared renderer) with deltas vs the committed baseline.
+# Deterministic code-quality report for the coach-agent surface (plan 43).
+# Regenerates agent/reports/ in the committable no-baseline form (what CI's freshness gate compares).
 report-agent:
-    cd agent && uv run scripts/report.py --out reports --baseline reports/agent.baseline.json
+    cd agent && uv run scripts/report.py --out reports
+
+# The delta form bakes a machine path into the markdown — that's why it never touches reports/.
+# Print agent deltas vs the committed baseline (local-only; writes nothing to reports/).
+report-agent-diff:
+    cd agent && uv run scripts/report.py --baseline reports/agent.baseline.json
 
 # Re-snapshot the agent baseline to the current hardened state (after a cycle).
 report-agent-save:
     cd agent && uv run scripts/report.py --out reports && cp reports/agent.json reports/agent.baseline.json
 
 # Deterministic code-quality report for the iOS surface (plan 48): runs a real
-# `xcodebuild clean test` (coverage + warnings) + periphery/swiftlint, rendered
-# via the shared renderer with deltas vs the committed baseline. Slow (~5 min);
+# `xcodebuild clean test` (coverage + warnings) + periphery/swiftlint. Slow (~5 min);
 # needs Xcode == report.mjs's PINNED_XCODE for the warning ratchet to be comparable.
+# Regenerates ios/reports/ in the committable no-baseline form (what CI's freshness gate compares).
 report-ios:
-    cd ios && node scripts/report.mjs --out reports --baseline reports/ios.baseline.json
+    cd ios && node scripts/report.mjs --out reports
+
+# The delta form bakes a machine path into the markdown — that's why it never touches reports/.
+# Print iOS deltas vs the committed baseline (local-only; slow — a real xcodebuild clean test).
+report-ios-diff:
+    cd ios && node scripts/report.mjs --baseline reports/ios.baseline.json
 
 # Re-snapshot the iOS baseline to the current hardened state (after a cycle).
 report-ios-save:

plans/quality-status.md

@@ -17,7 +17,7 @@ describes the *process*. Known defects/debts live canonically in the tracker; wo
 | **coach agent** (`agent/`) | provider keys + LiveKit secrets; prompt/trust boundary; summons are billed | 4 (plans 43, 45, 51, 55) | 01–03 | 82.75% | **closed** — same triggers |
 | **iOS app** (`ios/`) | HMAC seeds baked in binary; `GlassesSource.publish` frame path ~6.6% covered (#197-blocked, on-device-verified); live vendor fold-crash | 5 (plans 48, 53, 54, 56, 59) | 01 | 22.0% | **active** — next cycles land here |
 
-## iOS worklist seeds (next cycles)
+## iOS worklist seeds (priority-ordered — #1 is the next cycle's default scope; argue deviations to the user)
 
 - **File the [fold crash](tech-debt/dat-fold-channel-open-race.md) upstream with Meta** — 3 cycles overdue (3→4→5); the single highest-leverage parked item, re-confirmed on-device in cycle 5 (the only thing between the glasses path and a clean fold).
 - [RoomConnection recording dual-writer](tech-debt/roomconnection-recording-dual-writer.md) restructure (correctness; off the glasses path, fully off-device-verifiable).