# Companion authority and first teaching assignments — 2026-09-06

This follow-up keeps the full [checklist](../../TODOS.md) active. The immutable
teaching revision is `1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

## Source correction

Reviewing the companion unit against `libs/openapi/v2/companion.yaml` and
`apps/v2/web/dev-portal/api-reference.md` exposed an over-broad reading of the
feature overview's phrase “read-only.” The contract contains public-read
alternatives and separate owner-authorized private scopes. It also contains
scoped POST operations for coaching metadata, push subscriptions, tournament
operations, leaderboard ingestion, callbacks and micro-game results.

The corrected slides preserve the actual competitive boundary: companion
assistance does not control deterministic match state or participate in
cross-play. The coach response fixes `gameplayAuthority` to false. The shared
API's other writes do not automatically grant those authorities to a phone or
anonymous caller. The overview's 100/1000/30 traffic-policy figures remain
stated policy, not enforcement measurements; OpenAPI header examples alone do
not establish the complete limit policy. Web Push is also retained alongside
APNS and FCM in the subscription-contract explanation.

The new authority slide is an exact comparison table in the existing reading
workspace. The mandatory frontend skill kept one explanatory job per slide and
placed detailed scope and failure-review cases in notes. The surface guide now
has twelve slides, including its shared context slide. This is not a new guide.
The library has 381 unique slides and 215 cited sources. Audio for the changed
companion slides was already pending; the additional slide also remains
explicitly pending.

## Teaching review

[Two assignments](../../assignments/v2-surface-boundaries.json) now connect the
complete companion and shared-spine units in
`V2/features/product-promise-and-surfaces.md` to their explanatory slides in
`v2-product-surfaces`. Each records the source-unit hash, assembled slide
hashes, guide membership, per-slide explanation and a substantive review
rationale.

The review includes all statements in source lines 229–251 and 252–265. It
qualifies the broad overview with the primary contract instead of asserting that
the overview is runtime proof. These are teaching-fidelity assignments, not
service, mobile, audio or release acceptance. The validator reports two reviewed
units, 103,324 unassigned meaningful units and 10,018 structural units. The
final complete-coverage gate must continue to reject this incomplete state.

The initial survey also read the other feature-page units and the architecture
promise/netcode and documentation truth-state sections. They are not marked
reviewed here. In particular, full implementation mappings in the architecture
table, the netcode diagram's named downstream modules, exact release-state
fallback/repair transitions, reference conventions and related-reading routes
need their own complete explanations and assignments. Broader source citations
do not settle these remaining teaching requirements.

## Evidence boundaries

The pinned-schema regression parses the actual YAML and checks anonymous read
alternatives, distinct required write scopes, the private-player scope and the
false gameplay-authority constant. It also requires the corrected slides to cite
the primary schema. No API implementation, schema or authorization policy was
changed; no live authentication, persistence, rate-limit or provider pass is
claimed.

Current [50-test unit results](unit-tests.txt),
[browser report](guide-browser.json) and [log](guide-browser.txt),
[center draft report](center-draft.json) and [log](center-draft.txt),
[coverage check](coverage.txt), [expected incomplete gate](incomplete-gate.txt)
and [audio check](current-audio-check.txt) retain their explicit scopes.
Standalone and center runs use draft mode because V2 narration remains
incomplete. All twelve surface-guide slides pass desktop/compact/mobile, axe,
sources/transcripts, controls/history/dialogs and PDF page/title/footer checks.
The changed guide PDF is regenerated from that tested HTML.

The center run is the full `--draft` run, not smoke mode: all 407 guide-slide
instances are checked for embedded fit and axe accessibility at desktop and
mobile sizes. All 38 guide PDFs are regenerated and checked, with their exact
hashes recorded in the report. Navigation, source coverage, performance budgets,
history, hierarchy, duration semantics and mobile drawer checks also pass. Draft
mode deliberately excludes unfinished audio; full narrated acceptance and the
final visual/listening audit remain outstanding.

Direct image review covers the new authority comparison at desktop and mobile
sizes and its PDF page. The table uses the existing keyboard-scrollable mobile
layout, including a visible scroll hint. This inspection exposed a QA capture
issue: a preview could retain a shifted table after keyboard-scroll and
accessibility inspection. The harness now waits for stable scroll frames,
restores the position again after axe and asserts a final zero offset before
taking the preview. The contracts draft also exercises that regression across
its table slides. This is a screenshot-state correction, not a global visual
audit.

Seven further recordings are saved: `tara-continuation`, `tara-to-other-rooms`,
`arete-check-in-contract`, `arete-engagement`, `arete-recovery`,
`arete-friction-review` and `nyx-tonight`. The
[render log](narration-progress.txt) records fresh one-track CPU processes and
resource supervision. They retain the stock voice, model, pronunciation and
sentence-pause policy. All 81 current recordings pass persisted fingerprint,
byte/hash, decoded-duration and encoded pause checks: 6,145.579 seconds and 968
sentence segments. The 381-slide library still has 300 pending tracks: 280 stale
and 20 missing. No owned narrator was running during these browser checks.
