# Tara consumer continuity · working package · 9 September 2026

This package completes the remaining meaningful units in
`V1/features/domain-tara.md`: the honest shipment boundary, customer and depth
surfaces, partial completion, audio controls, accessibility, saved outputs,
Living Scenes integration, library families and related reading. It complements
the existing Tara web-practice, ritual-session and audio-state guides instead of
repeating them.

Implementation baseline: `8f3ee59e5e4accf90f40e8dd68866ed64c27c028`. Primary
citation pin: `1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

Visual thesis: keep the cream, ink and terracotta field-guide language and trace
one member's sit through checkpoints, controls, accessibility, saved outputs and
the next device. Real Tara web captures establish recognizable customer
surfaces. Source-shaped diagrams distinguish visible controls, local models,
service authority and connected execution.

Content plan: open with one interrupted sit and classify the available evidence.
Map the consumer hub, canonical sitting route, generic domain route and
standalone Tara apps. Compare their clocks, audio and playback-rate behavior.
Then follow partial completion into checkpoint selection, local resume,
quick-reset and sleep controls, transcript and haptic support, favorites,
collection/share/download/journal actions, Living Scenes and the library stack.
Close with a recovery matrix and the few connections that would make the member
journey continuous.

Interaction thesis: reuse the canonical deck entrance, keyboard navigation,
narration and source disclosures. Product captures use the existing
keyboard-accessible inspection treatment. Every state or handoff needed for the
argument remains visible in the static slide and PDF; motion is limited to the
reader's established slide transition.

## Source review and corrected boundaries

The handbook is accurate about the broad Tara contract and route families, but
three statements need narrower evidence labels at the pinned revision:

- Cross-device resume is not wholly spec-only. Standalone Tara web and mobile
  build exact-position checkpoint, Iris-memory and Psyche-envelope shapes and
  persist them locally. The web browser can rehydrate a same-browser checkpoint;
  its apparent mobile-to-web E2E case injects a mobile-shaped envelope directly
  into browser local storage. No inspected application path transports those
  shapes between devices.
- Hearing-impaired haptic pacing is not wholly spec-only. The domain library
  compiles deterministic standard and hearing-impaired timelines, the BFF
  exposes an authenticated compile route, and watchOS/Wear OS runtimes consume
  the protocol. The phone bridge can send a timeline, but only tests call
  `pushBreathTimeline`; no inspected member flow requests the BFF timeline and
  forwards it to the watch.
- Watermarked download is more than a planned type. Standalone Tara mobile
  creates a permission-checked manifest, downloads actual audio and preserves
  the manifest IDs in offline metadata. The manifest asserts watermark
  preservation; the inspected path does not transform or verify a watermark in
  the audio bytes.

The source also contains two separate Oshun web players. `/tara/sit/[id]` has a
presentation timer, ambient audio and authenticated start/completion writes; its
voice-speed slider does not alter a guidance track. `/domains/tara` has a real
audio element and changes `playbackRate`, but exposes a generic 0.5×–2× range,
stores favorite state only in React, discards reflection text at its parent
handler and copies the current URL for sharing. Standalone `apps/tara` supplies
a third implementation with local checkpoint and library-action records.

The canonical sitting UI distinguishes completion at 80 percent, while its BFF
projection writes fixed `audio-ended` and `completed` values without linking the
started ritual-session projection. The separate domain continuation model saves
checkpoints from five percent and treats a completion as recent for twelve
hours. Those are three different decisions and must not be presented as one
connected state machine.

## Working story

1. Interrupt one sit and ask what the next surface can actually recover.
2. Separate contract, implementation, local persistence, service authority and
   end-to-end evidence.
3. Map the consumer hub, canonical sit, generic domain player and standalone
   Tara surfaces.
4. Use real captures to identify the canonical product controls and reflection
   outcome.
5. Compare presentation clocks, audio engines and three playback-rate policies.
6. Follow 80-percent completion, five-percent checkpointing and the fixed BFF
   completion projection as distinct layers.
7. Execute the continuation decision table, then distinguish fixture-backed
   practice-plan links from standalone local resume.
8. Inspect quick-reset and sleep timers, accessible controls, transcript paths
   and the implemented haptic compiler.
9. Locate the missing phone request/delivery seam for a watch timeline.
10. Compare local favorite state, Tara service favorites and the unified Library
    response that appends two fixed examples.
11. Follow permission-checked collection, share, download and journal records;
    distinguish local records, native side effects and durable service state.
12. Revisit reflection recovery, Living Scenes and the library families.
13. Resolve the opening case with an explicit evidence and failure matrix.

## Acceptance and recovery

The final guide must preserve all ten exact source-unit hashes and exact slide
fingerprints in its assignment. Focused characterization tests must prove the
player split, speed divergence, completion mismatch, local-only resume evidence,
haptic implementation and missing caller, local saved-action records, real
favorite authority and fixed Library seeds. Existing owning unit and browser
tests remain separate evidence.

The candidate must pass the presentation builder, focused manuscript and source
contracts, assignment validation, narration integrity, standalone Playwright,
Presentation Center review, PDF text/bounds checks and complete optical review.
Mobile source/unit tests and existing Maestro flow definitions are applicable;
device execution is recorded only if an emulator is available. No local or
source-level pass is described as production, cross-device, wearable-device or
human acceptance.

If generation is interrupted, stop only the owned renderer or browser process,
retain its log, and resume missing narration by fingerprint. Canonical content
is promoted from the draft only after the rendered candidate passes. Existing
guide HTML and audio are hash-compared so a failed package cannot silently
replace previously delivered work.
