# Nyx narration and contracts visual review — 2026-09-06

The full [continuation checklist](../../TODOS.md) remains active. This increment
does not finish narration, teaching coverage, Eve integration or the library's
final editorial/listening review.

## Narration

Eight regenerated recordings bring the current total to **89 tracks, 6,764.814
seconds and 1,062 sentence segments**. There are still **292 pending tracks: 272
stale and 20 missing**. The persisted [audio check](current-audio-check.json)
records the manifest hash and exact ID lists; every current MP3 passed
fingerprint, byte/hash, decoded-duration and encoded-pause checks. The voice,
model, pronunciation and timing policy are unchanged.

The new recordings are Nyx's computation, observation quality, calendar loop,
almanac, opening boundary, worked evening and failure analysis, plus
`nisaba-reading-desk`. The [resume log](narration-resume.txt) saved the first
five. Its controller session ended with code 143 without a resource-threshold
message; the remaining bounded worker was separately confirmed live, then
confirmed finished before any replacement job. The
[three-track Nyx run](narration-nyx.txt) finished normally. No narrator was left
running during browser verification.

All nine slides of Nyx now have current audio. Its
[full standalone report](nyx-browser.json) and [log](nyx-browser.txt) cover
desktop/compact/mobile layout, accessibility, sources/scripts, controls,
navigation, PDF sequence and actual audio playback, pause/resume, seeking,
speed, end transitions, HTTP recovery and playback races. This is not a
normal-speed human listening review or acceptance of the Nyx product's live
providers.

The new [Nyx learning map](../../drafts/nyx-learning-map.md) records remaining
teaching outcomes after reading the entire pinned feature page. It also
reconciles the depth tree: 41 files means 21 implementation files and 20 test
files, not 41 computation modules. The map creates no reviewed teaching
assignments and does not publish placeholder guides.

## Contracts manuscript: complete isolated visual pass

The [review ledger](contracts-review.json) binds all sixteen ordered slides to
their assembled content hashes and the inspected artifacts. Direct inspection
covered every desktop and compact-desktop image, every mobile initial state,
every mobile table's far-end state, and all sixteen rendered PDF pages. All
notes and questions were read in sequence. The saved
[browser report](contracts/result.json), [log](contracts-browser.txt),
[PDF](contracts/presentation.pdf) and image folders retain the actual evidence.

The review found one conceptual defect: the cache slide's arrows could make
different command routes look like a single uniform sequence. Following the
mandatory frontend skill, the correction uses an exact comparison table in the
established cream/ink/terracotta workspace. The table distinguishes CI setup,
explicit BuildGraph Cook entries, BuildGraph BuildCookRun commands and direct
wrappers. The notes now explain that BuildGraph itself also uses `-cook` without
an explicit DDC argument on those commands. Inherited environment can still
affect the cache, so this absence is not proof of a different effective cache or
a failed cook.

The pinned XML regression checks all seven cook-bearing nodes, the order of
their explicit Cook and BuildCookRun entries, and their different arguments. It
does not execute Unreal or claim effective runtime cache verification. The
harness now saves compact views and mobile table-end views for every applicable
slide, supplementing its keyboard-scroll and fit assertions.

Fifteen unchanged slides' desktop/mobile captures matched the previously
inspected captures byte for byte; the corrected twelfth slide was re-inspected
at both sizes. All compact and PDF renders were inspected from the new run.
Mobile tables intentionally retain horizontal scrolling: the first and last
states together expose the content, with visible focus and the scroll hint.

This completes the isolated draft's current visual/editorial review, not its
publication. Catalog integration, embedded-context review, source-unit
assignments, recorded narration and full listening remain pending. Future
content, order, asset or renderer changes invalidate the corresponding review.

## Verification scope

[All 51 unit tests](unit-tests.txt) passed, including the real Misaki frontend
and MP3 decoding regressions. Ruff, scoped ESLint, Prettier, builder freshness
and teaching-report freshness also passed. The teaching report still contains
only two reviewed units and 103,324 unassigned meaningful units; neither this
visual ledger nor the Nyx learning map is a substitute for those assignments.
The [post-sync artifact check](artifact-validation.txt) confirms the saved
slide, visual, HTML, PDF and manifest hashes against the delivered files.

The [center report](center-draft.json) and [log](center-draft.txt) describe the
full 38-guide draft run, not a smoke run. It checks all 407 embedded guide-slide
instances on desktop and mobile, accessibility, navigation, search, history,
hierarchy, duration semantics, performance and all 38 regenerated PDFs. Draft
mode excludes unfinished audio. The separate Nyx report above supplies its
actual-audio evidence; no full-library narrated release pass is claimed.

The [first center attempt](center-first-attempt.txt) wrote a passing report but
its supervising session returned 143. No owned process remained when checked.
That attempt was not accepted as a clean gate; the full run was repeated, and
the saved `center-draft` report above is from the repeat whose supervising
session exited normally with code 0. No resource-threshold termination was
reported; the cause of the first session's exit remains unestablished.

Reproduce with the pinned Python environment, resource checks and
`CHROME_CHANNEL=chromium`:

```sh
python tools/presentations/build-eve-oshun.py --draft v2-contracts
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --draft-id v2-contracts --draft
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide nyx
python tools/presentations/run-supervised.py node tools/presentations/tests/presentation-center.mjs --draft
```

For the saved draft PDF, `pdftoppm -png -scale-to 1200` produces the inspected
page images. Each run needs its own output directory so later artifacts do not
silently replace earlier evidence.
