Adds `NeuronEventIndex` — bounded drop-oldest per-neuron ring of
every structural event the neuron participates in (design §9.2), so
"show me the history of neuron 47" is O(index) instead of a scan of
the global event ring. Wires a third bus subscriber in the archive
worker plus a `request-events-for-neuron` / `events-for-neuron`
protocol pair.

Also bumps the archive worker's local-bus `maxSubscribers` default
to 8 so the upcoming footprint-history subscriber fits without
another edit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds the archive-side receiver for the hybrid log-spaced +
change-triggered footprint scheme (design §9.3). The new
`FootprintHistoryStore` keeps per-neuron drop-oldest rings of
`(t, sparse A column)` snapshots; typed-array payloads are retained
by reference to avoid per-snapshot copies.

Harvests footprints from three sources through the archive worker's
local event bus: (1) birth events, (2) merge-survivor events, (3)
split children, plus a new `footprint-snapshot` `PipelineEvent`
variant that W2 will emit on the log-spaced schedule in task 5.
Adds a `request-footprint-history` / `footprint-history` protocol
pair for the scrubber UI in Phase 7. Default sizing matches the
§9.3 ~5 MB per-session budget.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Fit worker now emits the five header-bar vitals (cell_count, fps,
memory_bytes, residual_l2, extend_queue_depth) every vitalsStride
frames (default 8). residual_l2 is computed from `fitter.step()`'s
return; fps is wall-clock-derived over the last interval so it
reflects what the user sees on the sparklines. Metric names live in
a new `lib/vitals.ts` module shared with the upcoming vitals bar so
emitter and UI can't drift.

Also adds `calaMemoryBytes()` to the cala-core wasm-adapter — a
single source of truth for the WASM heap size now that a consumer
(W2) actually needs it.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds the log-spaced floor of the §9.3 hybrid scheme: a new
`FootprintSnapshotScheduler` emits `footprint-snapshot` events at
ages 1, 2, 4, 8, … frames after each neuron's birth, using the
latest footprint attached to a mutation (register / merge / split)
as its cached snap.

The change-triggered branch (‖A_curr − A_last_snap‖_F drift) is
left as a deliberate TODO — it needs a new wasm-bindgen accessor on
`Fitter` to read per-component A columns. Until that lands the
log-spaced schedule gets us scrubber data during quiet periods and
exercises the archive's footprint-history store end-to-end.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Extends the main-thread archive client with three new promise-based
request methods that forward to the W4 queries added in tasks 1-3:

  - requestTimeseries(name)     → tiered sparkline data for a metric
  - requestEventsForNeuron(id)  → per-neuron structural history
  - requestFootprintHistory(id) → per-neuron (t, sparse A) scrubber data

Refactors the existing dump path onto a generic `issueRequest`
helper so every reply kind shares one requestId-correlation store
without type drift.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds the §12 header vitals bar: five Canvas-based sparkline widgets
bound to a new `vitals-store` that polls the archive client every
500 ms for cell_count / fps / memory_bytes / residual_l2 /
extend_queue_depth. L1 + L2 tiers merge into a single flat window
(default 120 samples ≈ 60 s of history) so the component doesn't
know about the retention scheme.

Lifecycle is tied to the run: the bar spins up an archive client
on transition to running, polls while the run is active, and tears
down on stop/error. Sparklines auto-scale per series so each vital
gets its full vertical range regardless of magnitude.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds the §12 event feed: a structured, grid-laid-out log of every
pipeline event from the dashboard store — kind, id, time, and a
compact human-readable description per row. Birth / deprecate /
reject rows get distinct accent colors so the feed reads as a
narrative rather than a wall of text.

Format helpers are extracted into `event-format.ts` and
unit-tested directly so the visual component stays untouched by
string-formatting churn. Task 9 composes this + VitalsBar +
SingleFrameViewer into the final dashboard layout.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Composes VitalsBar + SingleFrameViewer + EventFeed into a single
grid layout (design §12): vitals along the top, preview canvas on
the left, event feed on the right, each cell independently
scrollable. App.tsx hands off to `DashboardLayout` on any
active-run state; the old SingleFrameViewer side-panel + inline
event list come out since their roles are now covered by the
dedicated components.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Extracts the Phase 3 cold-start E2E's inline `run_extend_cycle`
helper into a public crate function at
`extending::driver::run_cycle`, then wraps it behind a new
`Extender` wasm-bindgen class. The browser W3 worker can now own a
`ResidualRingBuf`, push residuals per fit frame, and call
`runCycle(fitter, queue)` on whatever cadence it chooses —
proposals land on the shared `MutationQueueHandle` ready for
`drainApply`.

Exports `Extender` through `@calab/cala-core`. The Phase 3 test
switches to the shared driver to prove numerical parity: same code
path serves both the native cold-start recovery proof and the
browser worker.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Fit worker now owns an `Extender` + residual ring: on every frame
it pushes the residual, and on `extendCycleStride` (default 32
frames ≈ 1 s at 30 fps) it calls `Extender.runCycle(fitter, queue)`.
Real `register` proposals land on the Rust mutation queue and get
applied on the next `drainApply`, advancing the fitter's epoch and
populating `cell_count` / `numComponents` — the vitals bar actually
moves now.

The per-cycle proposal count is published as `extend.proposed` so
the event feed shows extend activity even before structural events
surface. W3's heartbeat stub stays in place for the orchestrator
lifecycle handshake.

Architectural trade-off: running the cycle inside W2 avoids a
cross-worker snapshot transport (design §7.2). True W2/W3
separation needs a WASM binding that serializes the Snapshot across
workers + a `Fitter.drainApplyEvents()` to surface `register`
payloads as `birth` events — both deferred to Phase 7.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds `e2e/phase6-extend.e2e.test.ts`, mirroring the Phase 5 exit
harness (real AVI bytes, real SabRingChannel, real worker modules)
but configured with a non-zero `extendCycleStride` so fit's new
extend path actually fires. Asserts:

  * Extender.runCycle is driven every stride
  * fitter.drainApply is called per cycle → epoch advances
  * `extend.proposed` metric events reach W4 with the expected count

Also fixes a bug the E2E surfaced: task 11's `runExtendCycleIfDue`
pushed proposals onto the Rust mutation queue but never triggered
`drainApply` — the existing drain loop only fires when the JS-side
queue has items. Now the extend path drain-applies inline so epoch
advances as soon as the cycle proposes anything.

Phase 5 E2E mock picks up a no-op StubExtender since the fit
worker's Extender import became unconditional; no behavior change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds a main-thread authoring path for pipeline mutations. New
types and flow:

  * `UserMutation` in the runtime's worker-protocol (narrow to
    `deprecate` for Phase 6 — register / merge need a footprint
    picker, deferred).
  * `RuntimeController.pushUserMutation(m)` forwards to the fit
    worker as a `user-mutation` inbound.
  * Fit worker handles it by pushing through the Rust-side queue
    (`pushDeprecate`), drain-applying so epoch advances, and
    publishing a `deprecate` structural event so the UI feed shows
    the user's action.

UI affordance: a "Deprecate latest" button in the event feed
toolbar that targets the most recently born neuron id. The real
click-a-footprint UX lands with the Phase 7 overlay.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Drops `public/coi-serviceworker.js` — an inlined, BSD-licensed
service worker that re-issues top-level navigations with COOP +
COEP headers so the Pages deploy boots `crossOriginIsolated` and
`SharedArrayBuffer` works without server-side header control.
Registered synchronously in `index.html` before the module script
so the page is isolated before any SAB-using worker is constructed.

Vite dev + preview already set the same headers directly, so this
is a no-op there.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…sk 16)

Ships `e2e/phase6-exit.e2e.test.ts`: drives the full decode → fit
→ archive pipeline on a real miniscope AVI and asserts every
Phase-6-shipped surface is reachable end-to-end:

  - vitals timeseries populate (tiered store from task 1)
  - per-neuron event index resolves a birth (task 2)
  - footprint history returns entries (task 3)
  - extend.proposed metrics + structural events land in the archive
    dump (tasks 4 + 5 + 11)
  - user-authored deprecate mutations reach the fit worker (task 13)
  - no worker errors during or after the run

Also fixes `pixel_size_um` missing from W2's metadataJson (task 11
introduced an Extender construction that parses RecordingMetadata
but run-control.ts only forwarded height/width). Surfaced during
live browser testing; without the fix the fit worker panics on
init in production.

Updates `.planning/CALA_DESIGN.md` §12 with the Phase 6 exit date,
the shipped artifact list, and explicitly deferred items (real
birth events need `Fitter.drainApplyEvents`, cross-worker
snapshot transport, ε change-trigger, Playwright, click-a-
footprint UX).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

W1's `frame-processed` heartbeat hardcodes `epoch: 0n` (it has no
view of the fit pipeline's epoch), but `run-control` was routing
that to `recordFrameProcessed` — so the viewer label read
`epoch 0` forever while the fit pipeline silently advanced. Move
the dashboard listener onto the fit worker, which is the only
worker that knows the real epoch. W1's listener stays for
`frame-preview` only.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The `initCalaCore` resolver now reads `mod.memory` to expose
`calaMemoryBytes()` (Phase 6 task 23 / vitals bar), and the
adapter re-exports the new `Extender` binding (task 23). The
mocked init resolver needed to return a stub `WebAssembly.Memory`
instead of `undefined`, and the stub module needed to export an
`Extender` class. Adds one test for the new `calaMemoryBytes()`
helper.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>