v3.7.1

"Artifact Truth": what is tested, what is declared and what is shipped are the same files (the review of 3.7.0, LOG
#91). inventory(dist) = code + locked assets + named data + fonts + icons + licences, checked on every build.

Fixed

• The archive held what Vite copies from public/, shared by the games: the sample game's shipped the reference
  chapter's score and bridges, files its provenance lock never named. Every build now ends with tools/dist.ts seal:
  the files that are not the game's are removed and dist/licenses/ is written (the engine's LICENSE,
  LICENSE-ASSETS, CREDITS.md from the provenance, THIRD_PARTY_NOTICES.txt with the licence of every bundled
  package and the fonts' OFL, assets-manifest.json). npm run verify:dist, a step of build and of
  verify:commercial, refuses any file of dist/ that is not code, a locked asset with its reviewed bytes, a data
  file the game names, a font, an icon or a notice.
• A stinger that does not fit in the decoded audio beside the playing score is no longer played over the cap: the
  director lets it go (stinger() returns null, lastStinger.skipped) and it is streamed instead. The cap holds after
  every operation, which tests/director.test.ts now checks after every lifecycle case.
• playtests --strict passed with no session at all. It still only refuses a diverged session, and says when it
  checked none; the field quotas are new options (below).
• The release workflow builds the commit CI tested (workflow_run.head_sha), checks that the tag still points there,
  takes any SemVer tag (not only v3.*), and refuses a release that already has files (no --clobber): a published
  release is never replaced. The archive's assets manifest is attached beside it.
• CI runs on feature/**, fix/** and release/** branches, whatever the version (it stopped at v37-*).
• The nightly's shards split any total exactly (501 seeds ran 504) and the merge fails on a gap or an overlap
  (audit:corpus --shard=i/n --total=N, --merge … --total=N). The four-shard nightly ran on main (501 seeds of
  each kind: 910 compared, 593 partial, no divergence).
• The README's "Releases" section named v3.5 two releases later; tests/readme-release.test.ts now holds it to
  package.json's version. Python bytecode (__pycache__/, *.pyc) is no longer tracked.

Added

• Field quotas for playtests: --require=N, --require-completed=N, --require-devices=N (a diverged session counts
  for none), and npm run verify:field (verify:commercial, then 5 sessions, 3 played to the end, 2 device families;
  not a release gate, D12). A shared session says its device family (ios, android, desktop), nothing finer.

Changed

• The player's first visit runs 122 KB of gzipped JavaScript (153 before): the save envelope's schemas use
  zod/mini (same checks, tree-shaken), and each built-in minigame loads when one starts
  (src/engine/minigames/meta.ts keeps what the tools read). New budget assetBudgets.initialJsKB, checked on the
  build by npm run verify:dist. The engine's layers are written down (src/engine/BOUNDARIES.md) and checked by
  tests/boundaries.test.ts: the core imports no DOM or tool, the player never the solver, the validator or the Studio.

Manual passes (D12: reported, not blocking)

0 of 7 done.

Pass	Status	Who, when	Device, OS, browser, versions	What failed
Screen reader (docs/dev/SCREEN-READER.md, VoiceOver iOS or TalkBack Android)	not done
Safari offline (docs/dev/SAFARI-OFFLINE.md: first visit online, airplane mode, a room never visited, a reload)	not done
Real phone, frame rate of the heaviest scene (≥ 30 FPS)	not done
Playtesters who do not know the puzzles (npm run verify:field: sessions, finished, device families)	not done
Recorded voices (lines approved / lines with an id)	not done
Listening (each score on a phone speaker and on headphones, each bridge, a save loaded mid-transition)	not done
Signed tag (git tag -s)	not done