# A command is not a sound · Tara · practise, chapter 3 · verification receipt · 11 September 2026

Scope: the third chapter of the _Tara · practise_ track. Twenty-three inherited
slides became twenty-nine: a cover, four section dividers, twenty-three content
slides and a close, in sections of six, five, six and six. The
[brief](../../authoring/tara-audio-state-chapter.md) records what the chapter
is for and why each treatment was chosen; the track brief is
[tara-practise-track.md](../../authoring/tara-practise-track.md);
[state.json](state.json) holds the exact counts, tool results and findings.

## What was checked

- **Build freshness.** `build-eve-oshun.py --check` passes on the delivered
  tree: 1,489 slides and 663 sources. The chapter is 29 slides running 38.6
  minutes. It has no table and five diagrams: a state machine, two decision
  trees, a sequence-lanes loop and a layer stack.
- **Every value was executed, not read.** The pinned `audio-session.ts`,
  `audio-session-manager.ts` and `session-launch.ts` were bundled with esbuild
  and run under Node 22, with `Date.now` replaced by a controlled clock where a
  value depends on time. The walk over all eighty phase-and-event combinations,
  the listening trace, the two completions, the late update, the marker samples,
  the snapshot, the throwing listener, replacement and destruction, and the
  closing record are all results of those runs; `state.json` lists each one.
- **Browser rendering.** `eve-edition-review.mjs --guide tara-audio-state`
  sweeps all 29 slides at 1440×900, 1280×720 with audio chrome, 390×844 and in
  print — 116 measurements, no spill, no document overflow, no page errors, no
  findings.
- **Narration.** Twenty-nine tracks rendered on this host's CPU and verified
  with `--check --ids`. The library-wide `--check` passes too, which it had not
  since chapter five of the find-a-practice track — see below. The library is
  1,967.6 minutes over 1,489 tracks, with nothing pending.
- **PDF.** 29 pages, exact authored visual text on every one, in order, read
  back from the downloadable `decks/tara-audio-state.pdf`, and `check-pdf.py`
  passes.
- **Bindings.** No assignment binds a slide of this guide.
  `check-center-coverage.py --check` reports 347 reviewed and 0 stale over the
  whole library.
- **Suites and lint.** Python 473 passed, 0 failed. Node 489 of 489.
  `ruff check` and Prettier clean on every touched file.

## What executing the source changed

Three claims the chapter was going to make did not survive being run, and each
is corrected in the slides, the notes and the brief:

- **The engine update carries no clock.** `TaraPlaybackEngineUpdate` has eleven
  fields and none of them is a timestamp. `activeListenSeconds` adds
  `Date.now()` at this call minus the reading the previous call stored — the
  host's own wall clock between two _calls_, not an interval any report
  describes. So the figure follows the calling cadence: three playing calls two
  seconds apart accumulate four seconds of listening while the track advances
  two. The running example's numbers held — half a second, then one more, then
  nothing, for 1.5 — but only because the calls in it are that far apart.
- **The owning test does not walk the table.** The inherited notes said the
  source fixture checks all eighty combinations; the owning tests assert
  fourteen distinct ones. The chapter
  now says what was actually run: an executed walk over all eighty accepts
  twenty, seventeen of them in the four live phases. The close and the catalog
  summary had turned that into "eighty accepted", which is wrong by sixty and is
  fixed in both.
- **The snapshot has eleven fields, not twelve**, and re-entering a marker is
  announced only after a _different_ marker was active in between — a marker
  sampled, left for empty space and sampled again stays silent, because the
  guard still holds its id.

One behaviour surfaced that the chapter now teaches: because `subscribeState`
replays the current session immediately, a throwing state listener throws out
of `subscribeState` itself whenever a session already exists.

## What was fixed

- **Five overflows at 1280×720.** `tara-audio-status-priority` was planned as a
  card grid grown to five cards; five cards wrap to two rows and ran 97px past
  the narration chrome. It is a five-band layer stack now — the library's
  existing form for precedence — and each band's `boundary` states the
  condition that sends a report to the band below. Three case studies ran 6px
  over with four beats each, because the Eve-edition case list stacks rather
  than wraps and a fourth beat costs a full 88px row; they carry three beats and
  keep the fourth sample in their notes. The fourth, `tara-audio-listening`,
  also had a node detail wrapping to two lines.
- **A label under a line, and exits left undrawn.** In the lifecycle state
  machine the terminal box sat under `active`, so the `complete, abandon or
  error` label landed on the lane the `paused → active` return runs along and
  the line struck through it — the sweep measures spill, not collisions, and a
  screenshot caught it. The box sits under `paused` now and every label is
  clear. The drawing also left out the exits from `preparing` and `ready`, both
  of which accept `session_abandoned` and `session_error`; the specimen label
  now says so, as the state-machine validator asks, and the notes point to the
  command grid that lists them.
- **Fifteen orphaned narration tracks.** Merging the mood, theme and modality
  instalments of `tara-selection` into registries removed fifteen slides from
  the library but left their fifteen tracks — 18.9 minutes of audio — in the
  manifest and on disk. The renderer's full `--check` asserts that the manifest
  inventory matches the script exactly, and had failed on that since the merge.
  The fifteen entries and their files are removed, as earlier merges did.

## What was not checked

- **No capture.** This is the one chapter in the track without one. The shared
  audio model's only consumer is native, no pinned capture of that surface
  exists, and the web player from chapter one does not use this manager, so a
  picture here would show a screen the chapter is not about. The plan item that
  asks for a capture is left unchecked.
- **No device, no audibility, no persistence.** The manager writes no store; a
  snapshot is eleven scalars handed to a consumer that has to exist. The native
  adapter is read in source only, and the linked native guide follows the
  actual mobile caller. A true command return means the transition was legal.
- **No claim about a member.** `activeListenSeconds` is a measurement of
  qualifying wall-clock intervals between calls. It is not attention, benefit or
  completion quality, and the chapter says so where it teaches it.
