Mount the Brain chat-wide and absorb the inline brain strip #59

Acceptance Criteria

• In-session and on the landing/new-session draft, exactly one full-bleed Brain background layer ([data-component="brain-atmosphere"]) exists per Chat window, painted behind the conversation timeline, the composer, and the draft; the Rail, titlebar, and Panels contain no Brain layer.
• No inline brain strip renders anywhere in the timeline: for any session, no [data-component="amico-brain-strip"] and no [data-timeline-row="brain"] element exists, and the timeline row model produces no "brain"-keyed row.
• While a turn is working, the timeline still renders its working-indicator shimmer ([data-slot="session-turn-thinking"] / the cycling-gerund + live-elapsed line), unchanged.
• Deriving Brain events from the session sync store for a given session's messages + parts yields: touch events in message order (mapped via the tool→brain-ref lookup), a chart marker after each completed assistant message that committed ≥2 non-consider touches, replay=true for completed turns and false for the in-flight busy turn, and no duplicate ids. A fresh session with no assistant activity yields an empty stream.
• Feeding that derived stream through the branch Brain engine grows its atlas / claimed-node count (data-true: touches light nodes); a replay-only first flush restores prior turns quietly (no live pulse travel).
• Switching the active session swaps the background to that session's derived stream via a keyed remount; only one engine / render loop is alive at a time (the prior engine is destroyed, no leaked animation frames).
• A chat-theme flip (data-color-scheme) repaints the Brain via a lossless theme swap without tearing down the atlas; both dark and light render.

Testing Decisions

• CREATE a headless unit test for the extracted event-derivation function (the inline strip's events logic, lifted to a pure function): feed sync-store-shaped fixtures (messages + parts) and assert message-order touches, consider-vs-commit counting, chart-after-≥2-commits-on-completed, the replay flags, id de-duplication, and empty-on-fresh-session. This is the "live event feed" contract (AC 4) and the reused strip logic.
• EXTEND the existing headless Brain-engine unit suite (the recording-2d-context engine test — no DOM required) with a background-mount driving case: replay the derived stream through the engine and assert stats() shows atlas/claimed growth, that a replay-first-flush restores quietly (claimed lands with stats().queued === 0 — no traveling pulse), and that a theme swap preserves the atlas (AC 5, AC 7). Flag: extends the brain-engine test surface.
• CREATE a headless timeline row-model test (no message-timeline*.test.* exists today): against the pure constructMessageRows, a working turn (isActive, busy, no error) emits a Thinking row — a fail-first guard that breaks if the strip-removal refactor regresses the shimmer (AC 3). The strip's removal (AC 2) is enforced by dropping the Brain member from the TimelineRow union, which makes the append/render-case/keep-mounted/key-branch/framed-exclusion deletions typecheck-required (the app typechecks before the PR); the runtime absence of [data-component="amico-brain-strip"] and [data-timeline-row="brain"] rides browser-first and the optional DOM assertion below.
• Only if the app harness already supports component/DOM mounting, add render assertions for the AC 1/2/3/6 selectors (brain-atmosphere present behind the pane; no amico-brain-strip / brain row; session-turn-thinking present while working; background re-keys per session). Do not stand up a DOM harness that does not exist — the derivation + row-model + engine unit tests above carry the slice.

Key Decisions

• Mount seam: one full-bleed Brain per Chat window, absolutely positioned inside the shared Chat pane that hosts the conversation-timeline/landing switch and the composer dock (and the standalone new-session draft surface), painted above the pane's own opaque surface fill and below all Chat content, pointer-events: none, keyed on the active session id — empty on the landing.
• Engine retarget: reuse the old-branch full-bleed wrapper's shape and environment plumbing but point every engine call at the engine already on the branch. Method map: the old suspend() → the branch pause(); the old refreshTokens() → the branch setTheme(scheme) with scheme read from data-color-scheme. The old-engine-only setActive / setReducedMotion are dropped here (active-tempo + reduced-motion tempo are slice 4); the old requestRender font-ready hook degrades to a resize()/no-op repaint (non-load-bearing this slice).
• Live feed: reuse the inline strip's exact derivation — touches in message order via the tool→brain-ref map, a chart after each completed assistant message with ≥2 commits, replay=true for completed turns, live for the busy turn — sourced from the session sync store and scoped to the active session. Extract it as a pure function so it is unit-testable and shared by the mount.
• Data contract: derived event is { kind:"touch", id, label, type?, consider?, replay } | { kind:"chart", id, title } (the strip's existing shape), diffed/de-duplicated by id, first flush restoring charts silently.
• Strip removal: delete the Brain timeline row — its tagged-class, its append into the row model, its render case, its keep-mounted special-case, its row-key branch, and its exclusion from the framed-row type. Keep the Thinking row / working-indicator shimmer untouched. The strip's click-to-expand/collapse is obviated (the engine has no node hit-testing, so nothing interactive is lost).
• Tab swap: a keyed remount on the active session id restores that session's atlas (completed turns replay silently on first flush), guaranteeing exactly one live engine/render loop.

Constraints & Invariants

• The full-bleed layer must be pointer-events: none and must never intercept scroll, wheel, touch, or click events bound for the timeline or composer — the timeline's scroll, bottom-lock, and history-load machinery must behave exactly as before.
• Exactly one Brain engine / render loop per Chat window at any instant; a session switch destroys the prior engine with no leaked requestAnimationFrame chain and no parallel loops. Carry over the wrapper's visibility + intersection hard-suspend verbatim (hidden window or off-screen pane must stop frames) — it is reused behavior, not new.
• Do not change the engine's motion, tempo, at-rest, or palette behavior and do not add any glass tint in this slice.
• Rail, titlebar, Panels, and the timeline's shimmer line are untouched.
• The Brain graph palette is exempt from the single-yellow chrome law (#fff676 stays reserved for live thought) — do not restyle the graph.
• data-color-scheme drives the engine scheme (dark default); a flip is a lossless repaint; both themes are first-class from day one.
• Browser-first iteration (vite dev on the app package), not VSIX rebuilds. npm run lint clean and the app typechecks before the PR.

Source

• Part of Make the Brain the Chat's living background; float components on legibility-floored Glass #56
• Durable record: ADR packages/app/docs/adr/0002-brain-as-chat-background.md