GOAL — Ch1 verbatim conformance: the playscript is the script

[juliatetrud]

GOAL — Ch1 Verbatim Conformance: the playscript is the script

Goal issue. Supersedes the partial-apply posture of #111, #129, #130, #131: the playscript
wins everywhere. Ends when every Ch1 dialogue and narration line matches the committed
ch01-playscript.docx verbatim, proven by diff, live on Pages.

Stage 0 — what is actually live (COMPLETE)

0.1 The commits are on the remote. Push is NOT the explanation.

Check	Result
git log origin/main..HEAD	empty
local HEAD	e77fb78d5d62517ec55e7602467dc2a3a9cbeab9
git ls-remote origin main	e77fb78d5d62517ec55e7602467dc2a3a9cbeab9
R01b commits bea4d87 d757fcb 8acfef2 b717f39 e77fb78	all contained in origin/main

0.2 GitHub Pages is serving the current build. Propagation is NOT the explanation.

git show HEAD:game/index.html | shasum  -> 14212587c29e806a325522984a78633c9f9da5b7
worktree game/index.html                -> 14212587c29e806a325522984a78633c9f9da5b7
live https://juliatetrud.github.io/...  -> 14212587c29e806a325522984a78633c9f9da5b7   (HTTP 200, 934,972 bytes)

Byte-identical. The live page is the committed HEAD. Only a client-side browser cache
could still show stale content — unverifiable from here; a hard refresh (⌘⇧R) rules it out.

0.3 So why did Ch1 look unchanged? Because R01b changed almost no words.

This is the real answer, and it is not a deployment problem:

• 14 of R01b's 16 edits changed zero text. They flipped the italic flag — which box a
  line renders in (thought bubble vs. narration box). Every character of every line is the
  same. Reading Ch1 for wording, nothing would look different.
• The 15th edit moved one line's position ('Pip finishes the dish in record time.').
• The 16th was not applied — held as R01b Stage 3 — kitchen bite-beat timing: scoped, NOT applied (needs ruling) #131 pending ruling.

R01b was correctly applied and correctly verified. It simply was never going to change how
Ch1 reads, because the changelog it implemented was about presentation and ordering.

0.4 Fresh conformance baseline vs. the committed canonical playscript

Method: python3 stdlib docx extraction, normalized matching (mixed \uXXXX-escape vs. UTF-8
curly quotes, dash/ellipsis/whitespace folding). No packages, no browser.

108 payloads — 97 match, 11 divergent. Of the 11:

• 8 are the three held items — the only remaining wording divergences in Chapter 1.
• 3 are extraction artifacts, not divergences: dc-cart-inspect's clipboard line (identical
  text, split across two dialogue lines in code); janitor-mutter (whitespace only — one space
  before an ellipsis); PIP (softer now): Hello? (parser artifact — the line "Hello?" is
  present and correct).

Conclusion: Chapter 1 already matches the playscript verbatim except for the three held
items. This goal is therefore, in substance, the decision to apply those three — which is
exactly what the goal states. The remaining work is small in line count and large in risk.

The risk this goal has to decide, before Stage 1 rewrites anything

Two of the three held items are safe text swaps. One is not.

Safe: held item 1 — gp-door-return

Playscript: "Babcia and Dziadek are gone." The Go in. / Not now. choices are separate from
the line text, so swapping the narration does not remove the radio-backtrack entry point. The
objection was continuity (they don't leave until the Ch7 dock-farewell), not breakage.
Applying it is a pure text change. Julia's call, no engineering risk.

Mostly safe: held item 2 — the cart clue

Playscript drops "back in the cabin" and says "call" for "page". The radio still works; what
is lost is the hint that the CP-2b / G-S5 fix added to close a corridor→radio dead-end.
Risk is discoverability, not a hard block. Julia's call.

NOT safe as a text swap: held item 3 — the radio

The playscript describes the pre-Sprint-20 design: a radio in Dziadek's coat pocket that
Pip pulls out and speaks into, with gp-armchair-inspect carrying four lines and setting
radioDiscovered. The shipped design is a windowsill gp-radio object (x=200) with three
gated states.

Applying the playscript's text verbatim produces two failures:

1. Two radios in one room — a narrated coat-pocket radio plus the still-present interactive
   windowsill object.
2. A potentially uncompletable chapter. The playscript replaces gp-radio-page-choice's
   Listen / Page / Not-now choice block with a single spoken line. But chapterState.janitorPaged
   is set inside that block's onSelect. Lose the choices, lose the flag → the janitor never
   walks → the cart never clears → the stairwell gate never opens → Chapter 1 cannot be
   finished.

Verbatim conformance for this beat is achievable, but it is a feature rewrite — move the
radio back into the coat, re-home the radioDiscovered / janitorPaged chain, and keep the
gate satisfied — not a line substitution. It needs to be scoped as such.

Blocked — awaiting the rest of the goal prompt

The goal instruction terminates mid-sentence in Stage 1:

"1. Re-extract the committed ch01-playscript.docx to a plain-text working transcript,
scene by scene,"

and it refers to "the hard stops" ("stop only at the hard stops") which are never defined.
There is no Stage 2 or beyond. Stage 0 is complete and reported; Stages 1+ cannot be run to
completion from what was given.

What is needed to proceed

1. The remainder of the goal prompt — the rest of Stage 1, subsequent stages, and the
   definition of the hard stops.
2. A ruling on held item 3. Verbatim text conformance and a completable Chapter 1 are not
   simultaneously achievable by substitution here. Either:
   • (i) authorize the radio feature rewrite (pocket radio restored, ability chain re-homed,
     stairwell gate preserved) — verbatim conformance, larger sprint; or
   • (ii) apply items 1 and 2 verbatim and carve item 3 out as the single documented
     exception until the rewrite is scoped; or
   • (iii) something else Julia specifies.

Nothing has been changed. Stage 0 was read-only.

[juliatetrud]

GOAL COMPLETE — Ch1 verbatim conformance shipped and live

Result: 108 structural comparisons, ZERO divergence outside the item-3 carve-out.

52e7ff0 on origin/main; tag r-checkpoint-ch1-verbatim; git log origin/main..HEAD
empty; live page hash f81c20c5… == committed HEAD (verified after propagation, cache-
busted). DEBUG_WARP = false. node --check PARSE OK.

Why the earlier test saw no change (Stage 0)

Not deployment. Push was clean and Pages was serving committed HEAD byte-identically. R01b
changed almost no words — 14 of its 16 edits flipped the italic flag (which box a line
renders in), one moved a line, one was held.

The real root cause: R01 only ever performed text SWAPS, and this playscript's dominant
edit is DELETION. Lines Julia had cut were still in the build. Every prior audit missed it
because they asked "is this line present?" — which a deleted line passes trivially — instead
of "does this node hold exactly these lines, in this order, with these speakers and italics?"
That is why #129 could honestly report 97/107 conformance while Ch1 still didn't read as her
script.

Applied — one commit per scene

ff9a570 S2 · 7d507bb S3 · 1a4b0bd S4 · 6270eef S5 · a650f3f S6–7 · 78e9ce2 fixes
· 52e7ff0 docs. Each node's lines: array regenerated wholesale from the committed docx —
wording, order, speaker attribution, italic/roman, choice labels. choices: and onEnd
handlers were never touched.

Line-items for Julia's in-browser judgment

These are real content losses, all applied because the playscript wins. Flagging each so
the decision is visible rather than buried in a diff:

Where	What is now gone
Cinematic: bed-reveal	"The boy in the bed does not move." — R01 deliberately kept this as the reveal's setup; the reveal is now a single line
Cinematic: wire-shock	"The wire bites. White light punches through Pip…" — the shock itself is no longer described; only "Oh woah, I can float even higher now." remains
Cinematic: henrik-sits-down	3 lines: "Your grandmother. She told me, last night, in the dining room…", "Eat. Tell me if it's as good as she promised.", "Henrik does not answer immediately…"
Cinematic: kitchen-meeting	"Maybe he never has, now. Maybe this is the first time."
cabin-window-inspect	the fjord/Norway description + "He watches until the lights are gone."
hallway-sconce-inspect	the "it reacts to him" beat (2 lines)
gp-babcia-inspect	2 narration lines before "Babciu, jestem tutaj."
gp-photo / gp-suitcase	the second line of each
dc-cart	"It's back in the cabin." (the CP-2b / G-S5 dead-end clue) and "page" → "call"
dc-entry-nudge	the line naming the sparking sconce
dc-sconce-pre/post, dc-post-janitor	"He reaches out." / "The corridor breathes back into light." / "(can they?)"
gp-door-return	now "Babcia and Dziadek are gone." — note they do not actually leave until the Ch7 dock-farewell

Traversal (Stage 3.2) — no hard stop hit

Statically re-traced. cartFound, nokkenStoryHeard, janitorPaged, radioDiscovered,
bedRevealed all intact; every choice block intact (Try the handle. / Go in. / Not now.
/ Go down. / Not yet. / Listen. / Page someone over the intercom.); gates unchanged.
Two clue-strength reductions (the cart's "back in the cabin", the sconce nudge) leave the
player pointed by implication rather than instruction — weaker, but not stranded and not
misdirected, so under hard stop 1 they ship. Worth watching in play.

Item-3 carve-out — kept as shipped, per ruling

gp-armchair-inspect and gp-radio-page-choice untouched. The playscript's coat-pocket radio
is the pre-Sprint-20 design, and chapterState.janitorPaged is set inside the choice block a
verbatim apply would delete — losing it leaves the janitor unpaged, the cart unmoved and the
stairwell shut, i.e. an uncompletable chapter. Scoping note for a future radio-rewrite
sprint: move the radio from the windowsill gp-radio object (x=200) back into the coat on
the armchair; re-home radioDiscovered (currently the windowsill's first-inspect discovery)
and janitorPaged (currently gp-radio-paging's onSelect) onto the pocket-radio beat;
preserve the dc-stairwell gate on janitorWalked. That is a feature change, not a text swap.

Two things caught mid-flight, both fixed

1. A regression I introduced and then found: the array regenerator matched only
   {speaker, text, italic} and silently dropped pauseAfter: 1.5 from a Nøkken line. Caught
   by diffing property counts against the pre-goal commit e77fb78, restored in 78e9ce2.
   Audited the whole file afterwards — that was the only real property loss.
2. R01b Stage 3 — kitchen bite-beat timing: scoped, NOT applied (needs ruling) #131 is resolved by this work. One of the three deleted henrik-sits-down lines is
   "Eat. Tell me if it's as good as she promised." — the shipped line whose existence
   falsified the changelog's premise that Pip ate before being told to. With it gone the
   script is internally consistent and R01b Stage 3 — kitchen bite-beat timing: scoped, NOT applied (needs ruling) #131's blocking objection no longer holds.

Also landed

design-docs/ch01-playscript-extracted.md — the machine-readable transcript the conformance
diff runs against, regenerated from the docx. Dead 13-line legacy grandparents comment block
deleted; no commented-out dialogue remains anywhere in the file.

What to walk in-browser

Full Ch1, start to finish: hallway → grandparents' cabin → wall-phase into 646 → mirror →
bed → corridor (sconce, cart, radio backtrack, janitor) → kitchen (Henrik, Nøkken) →
observation deck → handoff. Watch for: the deletions above reading as too bare; whether the
cart still leads you back to the cabin without the explicit clue; and whether the corridor
nudge still points you at the sconce.

Ch1 now waits on that pass. R17 is not started and will not be until it is confirmed.

[juliatetrud]

RULING REVERSAL — the item-3 carve-out is closed

Julia reversed the carve-out ruling recorded above: rather than keep the shipped windowsill radio as a documented exception, she wants her hand-held radio version. Built as its own sprint, #133, shipped in e837256 (tag r-checkpoint-ch1-radio, live hash 8ba3ea61… verified).

Ch1 conformance is now 108 structural comparisons, ZERO divergence — with NO exemptions. The scoping note this issue left for a future radio-rewrite sprint has been discharged in full; there is no longer any part of Chapter 1 that departs from the playscript.

What changed, in brief: gp-armchair (world-x 280) now hosts all three radio states, with cartFound tested first so the CP-2b guarantee (page reachable on the first post-cart inspect) survives, plus the ability backstop so paging works even if the discovery beat was skipped. The windowsill gp-radio object is gone, gp-dziadek-inspect loses the now-false windowsill line, and the radio sprite is redrawn as a hand-held set in the coat's lapel with its proximity shimmer re-keyed to the new position.

Traversal was proved by simulation rather than reading — the real gp-armchair object was lifted out of the built file and its getter executed across all 8 combinations of cartFound × radioDiscovered × janitorPaged; every state returns a node, and paging is reachable from every post-cart state. Both orderings (radio-first and cart-first) complete.

Full detail in #133. The one thing that changed for the player beyond the radio's location: the paging beat is now a single spoken line with no Listen / Not-now choices, because that is what the script gives it.

This issue's remaining open thread is unchanged: Julia's in-browser Ch1 walk.