# Tara audio state and native playback · delivered increment

The next linked guides explain the shared audio state machine and its actual
native caller. They extend the ritual-session guide without equating its
validated ritual lifecycle with the separate eight-phase audio lifecycle.
Offline storage, reflection, quick reset, sleep, companions and the remaining V1
subjects remain required follow-on teaching; discovery below is not delivery.

Implementation baseline: `4d252910f85edd5009f23dead20220f62513d763`. Primary
source snapshot: `1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

## Design and sequence

Visual thesis: retain the cream field guide and use terracotta relationships to
make a playback command, an observed status and a saved record distinguishable.
Content plan: begin with an interrupted practice, explain the two state records,
follow commands and status updates, then show the native launch and feedback
loop. Finish with a concrete recovery and evidence check. Interaction thesis:
reuse slide entrance, source reveal and diagram inspection with reduced-motion
support; the complete explanation must remain readable in the exported PDF.

The shared-state guide will cover handoff fields, initial state and identity;
track/metadata/marker ownership; every legal transition including rejected
controls; interruption; the two completion paths; update precedence; duration,
position and measured time; marker intervals, ordering and repeat behavior;
event/subscription/snapshot behavior; and a worked recovery trace.

The native guide will cover the actual domain-screen metadata lookup and
constructed track records; route/lesson/session identity; asynchronous module
attachment; native audio-mode configuration; subscription before launch; track
creation and lock-screen metadata; command translation; status feedback; resume
seek ordering; visible controls and retry; error and teardown boundaries; and
the evidence required to establish actual device playback. Native delivery
belongs to V1.1 even though the shared library is used in Tara's V1 model.

Use state transitions, a directed feedback loop, record specimens, a timing
trace, interval/marker comparisons and concrete command/result examples. Tables
are appropriate for reference fields and legal controls. Peer rows do not get an
automatic highlight. Show exact observed case values rather than invented
playback, reliability or engagement metrics.

## Source findings to exercise before final teaching

- `audio-session.ts` and `audio-session-manager.ts` were read in full. Command
  transitions and engine updates are different paths. A manual completion fills
  progress to the duration; an engine finish retains the reported position.
  Active listening accumulates wall time only when the previous playback phase
  and the incoming playing flag both indicate playing.
- `meditationPlaybackEngine.ts` and its owning test file were read in full.
  Attachment configures the module before subscribing. It does not replay an
  already loaded track. Attachment failure returns false and clears references;
  the method comment's promised manager error report is absent from the body.
  The domain effect discards the returned readiness boolean.
- Native play/pause calls are synchronous listener calls; seek rejection is
  handled asynchronously. Player status reports `error: null`; load exceptions
  and missing URI have explicit report paths. Teardown wraps pause, lock-screen
  cleanup and player removal in one try block. Earlier exceptions can prevent
  later cleanup calls. Test the actual consequences before diagramming them.
- The source handbook's audio-controls paragraph attributes ambient mixing and
  sleep fade to the shared audio modules. Those modules expose state fields,
  play/pause/seek and interruption behavior; a field or event type alone does
  not implement a native setter, mix or fade. Trace the separate caller before
  repeating the broader claim.
- The actual native screen has a fixed twelve-entry audio metadata library. Its
  standard/high/preview URLs are constructed, not discovered through a media
  availability check. Resume positions include 344 seconds for the evening
  practice and 420 seconds for a course lesson. The caller starts the manager,
  resolves tracks, invokes play and then seeks from positive metadata; it does
  not await a native loaded callback before announcing readiness.

## Wider native discoveries retained for the following guides

The complete `taraOfflineCache.ts`, `download-manager.ts`, transfer test helper,
`mobileTaraReflection.ts` and reflection card component were read. Owning cache
and reflection-model tests were also read. These findings still need dedicated
source cases, teaching and delivery:

- Tara recents default to memory storage, three entries and fourteen-day
  retention. The download manager has a real native transfer port. Its hash
  comparison and Tara's descriptive checksum string need an actual comparison;
  the healthy test port returns whichever checksum the queued item expects.
- A completed download exposes `localPath`, while Tara's playback-track resolver
  creates an `offline://` URI. A complete manifest entry alone does not prove
  that the native player can resolve that URI. Trace any registered resolver.
- Cache selection, quality fallback, expiry, pruning, preference disabling,
  presence-media treatment and manifest removal have distinct consequences. Do
  not imply that manifest removal proves physical-file deletion.
- The reflection model labels the greater of measured listening and total
  duration as minutes listened. Its continuation builder supplies completed,
  100% and end-of-duration fields without checking the input phase. The actual
  completed-session effect supplies its caller guard.
- The native reflection handlers update component state and show saved/skipped
  labels. Locate any persistence consumer before claiming a saved return after
  remount. Telemetry's configured sink logs to the console; it is not a
  checkpoint store. The real reflection component supports 420 characters,
  starter suggestions, disabled empty save, saved preview, edit, skip and Done.

The source fixtures now pass 120 unmodified owning tests, 19 added audio cases
and nine exact native-caller cases. The latter extracts complete named functions
and metadata declarations through the TypeScript AST, using the actual shared
manager with explicit cache, resolver, analytics, haptic and UI state-setter
seams. It does not mount the native screen. The first expectation confused the
manager's session_created event with the screen's session_preparing funnel
stage; the source comparison corrected the fixture and manuscript.

The authored pair has 23 shared-state slides and 19 native-playback slides.
Draft optical review replaced repeated headlines and takeaways with concrete
case/specimen names and distinct lessons. The semantic check rejected lifecycle
rows whose starting states existed only in attributes; the visible descriptions
now name every source state. Final publication, export and optical evidence is
recorded in the delivery receipt when those checks settle. No current device
playback, human listening or durable storage acceptance is claimed.

## Delivery evidence

The [receipt](../verification/v1-tara-audio-2026-09-08/README.md) binds the
42-slide pair and 2910.381 seconds of new narration. Final guide review passes
302 checks with 168 captures, 126 embedded axe runs, eight enlarged-text cases
and five rejected content removals. Focused center review covers all 42 slides
in two embedded contexts, four audio boundaries and 42 PDF pages. All 70 prior
guide HTML files, four shared UI files and 792 earlier recordings remain exact.
Source fixtures and their hashes retain the actual evidence scope. Final PDF
optical review is recorded separately from draft checks. The first final review
found both closing slides extending beyond the 720-pixel desktop frame; shorter
case titles, detail and captions restore fit without changing narration.

The native guide is under V1.1 with both shared audio and native orientation as
prerequisites. The shared-state guide follows ritual sessions under V1.0. No
automatic peer-row emphasis is used; ownership diagrams focus the command owner
explicitly. In-app browser control was unavailable because its required
JavaScript tool was not exposed; repository Playwright and rendered-image review
provided the applicable fallback. Full V1 and human/device acceptance stay open.
