fix(foyer): round-12 true captions — NodeSlide two-layer evidence + honest publish gates

[HomenShum]

Summary

Round-11 review (Steward critical) flagged FYwall beat 02: the caption
claimed "this product's front door and back door were both checked live,
and they reported the same version" while it was shown over NodeVoice,
whose own pill on that exact frame reads "1 layer" — a caption false of its
own frame.

• Repoints beat 02 at NodeSlide (verified frontend+backend on production
  ec2a44e), hovers its card to reveal the apparatus, and asserts the
  apparatus text contains matches backend — the real two-layer evidence
  node-foyer's e2e/rules.ts two-layer caption rule requires before it will
  accept a caption claiming both layers were checked.
• Rewrites the caption in plain words (round-11's judge failed the "mom
  test" on the jargon "front door and back door"): "Green means verified.
  For this product both the website people see and the service behind it
  were checked live a moment ago, and both reported the same version."
• Beat 00: "every few minutes" -> "every half hour" (LEDGER_SWEEP_MINUTES = 30).
• Two further comprehension cycles on FYwall (judge is noisy, documented in
  this repo's own README) removed remaining jargon ("sweep", "deploy
  workflow") and named a persona up front, until the judge passed clean:
  publish, 0 P0, comprehension passed (worse of 2 runs kept).
• FYagent: one attempt at plain-language captions for the two JSON-endpoint
  beats, per the round-12 charter ("state the everyday purpose before naming
  the technical address"). Comprehension still blocks on non_expert_sense
  after re-judging twice (worse of 2 runs: rework, 2 P0) -- kept
  unpublished with its honest record rather than stretched further.
• Fixed a real capturer bug found while wiring the NodeSlide assert:
  assertHolds()'s text-matches branch recorded the generic phrase
  "<sel> text matches" for every capture, discarding what was actually
  checked, so no caption could ever satisfy a rule that greps asserted for
  specific matched text. Now records the matched pattern (or the equals
  value) -- a more honest receipt on its own, independent of this bug.
• FYphone: not re-captured (no caption issue was found in review). Re-ran
  the "run the judge twice, keep the worse" rule honestly against the two
  existing judge runs on its current MP4 -- the worse run has comprehension: blocked, so it is correctly demoted from published to unpublished
  versus the round-11 draft, which had cherry-picked the better run.

Vendored judge.json for all three captures and the node-foyer manifest
draft (data/walkthroughs.manifest.json) under the companion node-foyer
worktree; the planner's seal script now prints zero problems.

Test plan

• node check.mjs -- 60/60 JS files parse, 36/36 tour steps, 34/34 prose citations
• Fresh capture -> render -> judge cycle for FYwall (3 comprehension cycles) and FYagent (1 cycle) against production / a fresh built preview
• node-foyer's seal script (data/walkthroughs.manifest.json -> public/media/walkthroughs.json): 0 problems, media budget 3.47 MiB / 4 MiB
• node-foyer e2e/foyer.spec.ts file-mode stand-in for FYagent: 13/14 (rebuilt fresh; round-11's report could not be located)

🤖 Generated with Claude Code