# Linux continuation evidence — 2026-09-06

The full presentation objective remains incomplete. This record covers
destination setup and the narration duration portability correction, not final
library acceptance, expanded teaching coverage, Eve integration or visual review.

The worktree was clean at `fc4cd2de4039327b34e99ee27eef4076ab5a59d2`, matching
`origin/main`. The source pin remains
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`. The handoff inventory was reproduced:
380 slides, 213 cited sources, 64 current recordings and 316 pending recordings.
The builder's delivered-artifact freshness check passed before editing, and the
ignored whole-bank fixture was rebuilt and checked successfully.

## Environment

The isolated environment is `/tmp/oshun-presentations-venv.pot9EA` with Python
3.12.3. Both requirements files installed without changing their pins. PyTorch
2.14.0+cpu came from the official CPU wheel index. The environment also has
Kokoro/Misaki 0.9.4, NumPy 2.4.6, SoundFile 0.13.1, huggingface-hub 1.30.0,
spaCy 3.8.16, English model 3.8.0, markdown-it-py 4.2.0 and PyYAML 6.0.3.
FFmpeg/ffprobe 6.1.1 and Poppler are available. The exact pinned Kokoro model
revision and `af_heart` voice downloaded successfully without a credential.

Existing repository dependencies resolve Playwright 1.60.0, axe-core 4.11.4,
ESLint 10.4.0 and Prettier 3.8.3. Playwright's installed Chromium launches with
`CHROME_CHANNEL=chromium` and reports 148.0.7778.96. This launch check is not a
browser-suite pass. The older Chrome 152 report remains historical.

The mandatory frontend skill was read; its established cream/ink/terracotta,
Georgia/Arial composition remains the design direction. Dedicated presentation,
PDF and in-app browser skills are unavailable. Repository Playwright, axe,
Poppler and image inspection provide the applicable workflow. No subagents were
launched.

Resource checks found approximately 15 GiB RAM, 31 GiB swap and 15–17 GiB free
root disk; the repository is on a separate volume with approximately 270 GiB
free. Other worktrees run tests and TypeScript checks; available memory
changes substantially. Only targeted verification and bounded, supervised CPU
narration are admitted, with no broad workspace gates or concurrent owned
browser/narration jobs.

## Duration correction

The unchanged `orientation-boundaries.mp3` has SHA-256
`eb265ed8ad92feb1fff972fdfb3a480832d37be59f69e5d1bfcdf73c98138ec4`.
Linux ffprobe reports a 94.512-second container, while decoding yields
2,266,829 playable samples at 24 kHz: 94.451208 seconds, matching the stored
94.451 seconds. Container padding caused the original 0.06-second comparison to
fail even though every encoded pause and sentence boundary passed.

The renderer now returns duration from the already decoded samples used for
pause verification, for both new recordings and existing-file checks. It retains
the duration tolerance and verifies encoded alignment. No voice, model, script,
pronunciation, pause policy, fingerprint or existing audio bytes were changed by
this correction. The regression exercises the actual transferred file and
rejects a deliberately incorrect stored duration.

## Verified scope

- All 41 presentation-tool unit tests pass, including actual Misaki phoneme
  checks, encoded-pause rejection and the new cross-platform duration regression.
- All 64 handoff-current tracks pass fingerprint, bytes, SHA-256, decoded
  duration, sentence alignment and every encoded pause on Linux.
- Ruff and scoped Prettier pass for `tools/presentations`; Python files use Ruff
  because Prettier has no Python parser.
- The first CPU recording, `member-day`, passes the persisted fingerprint,
  hash, decoded duration and encoded-pause checks. Its duration is 66.105 seconds,
  SHA-256 is `fc25d0f2bdf4aeb49f572c725d06cee849c60b477ad29bd92ceffaca6dc5859c`,
  and generation used one CPU thread. The rebuilt catalog has 65 current tracks
  and 315 pending; the changed guide and ancestor totals include this track.
- The full 12-slide `portfolio-map` standalone run passes desktop, compact,
  mobile, axe, sources/transcripts, navigation/dialogs/history, print/PDF
  sequence, actual playback/seek/speed/end, HTTP failures/races and portable-file
  checks. The tested HTML hash is unchanged from the handoff. See
  [the browser report](portfolio-guide.json) and [log](portfolio-browser.txt).
  The opening desktop diagram and V2 mobile composition were also inspected;
  this limited inspection does not replace the final every-slide visual pass.
- The 38-guide center smoke run passes guide selection, performance budgets,
  search, hierarchy and duration aggregation, prerequisites, history, collapse
  persistence, and the mobile drawer/focus checks with stable HTML hashes.
  See [the report](center-smoke.json) and [log](center-smoke.txt). It used
  `--smoke --draft`; it does not certify all-slide layout or pending audio.
- The remaining 315 recordings, full current-library Playwright/PDF/audio
  gates, subjective listening review and every later checklist phase remain
  outstanding at this verification boundary.

The [unit log](unit-tests.txt), [64-track check](current-audio-check.txt), and
[initial CPU render](narration-first.txt) retain their explicit scopes. Further
temporary logs remain alongside the isolated environment. Update this record
when each further verification scope passes.
