`Fitter.drainApplyEvents(queue)` now exists alongside `drainApply`.
Returns `{ report: [applied, stale, invalid], events: AppliedEvent[] }`
via serde-wasm-bindgen — the event shape mirrors the JS
`PipelineEvent` birth/merge/deprecate variants with `kind` as the
discriminator.

Changes:
- AppliedEvent / DeprecateReason / ComponentClass gain conditional
  serde derives with `rename_all = "camelCase"` so the wire shape
  matches the existing TS union types.
- `packages/cala-core` adapter exports a typed `WasmAppliedEvent`
  union + `drainApplyEventsTyped` wrapper so callers don't repeat
  the `as` cast that wasm-bindgen's `any` return forces.

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

Replaces the Phase 6 `extend.proposed` metric-only emission with a
real `drainApplyEvents` call. Every successfully-applied mutation
now surfaces as a `birth`/`merge`/`deprecate` `PipelineEvent` on
the event bus — archive worker picks them up through the existing
subscriber path, so the event feed finally shows `born @(y,x)`
rows end-to-end.

The `extend.proposed` metric stays (still useful as a flat-line =
quiet-FOV signal), but the JS-side placeholder `mutationToEvent`
that Phase 6 used to fake ids + `patch: [0,0]` is no longer the
source of structural events for the extend path.

Test stubs (`fit.worker.test.ts` + phase5/6 E2Es) pick up the new
`drainApplyEvents` surface + `drainApplyEventsTyped` adapter
export so they continue to cover the fit loop without real WASM.

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

On preview-stride frames, W1 now calls the new
`Preprocessor.processFrameF32WithStages` WASM binding and posts
three `frame-preview` messages tagged with `stage`:
'raw' (decoded grayscale), 'hotPixel' (post hot-pixel median),
'motion' (post-motion, what fit sees). The fourth canvas
(`reconstruction`, Ãc) lands in T6 from W2.

Changes:
- New Rust `PreprocessPipeline::process_frame_with_stages` that
  emits the hot-pixel + motion-corrected intermediates alongside
  the final frame. The hot path (`process_frame`) is unchanged —
  the stage-capture cost only pays on preview-stride frames.
- WASM `Preprocessor.processFrameF32WithStages` returning a flat
  `[final || hot || motion]` buffer (3·pixels), sliced on the JS
  side into `subarray` views — no copy on the JS boundary.
- `WorkerOutbound.frame-preview` gains a `stage` field so the
  dashboard wiring can route each stream to its own canvas.
- `run-control` exposes `latestFrames: Accessor<Partial<Record<FrameStage,
  LatestFramePreview>>>`; the legacy `latestFrame` accessor keeps
  pointing at the `motion` stage for the existing SingleFrameViewer.

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

New WASM binding `Fitter.reconstructLastFrame()` returns `A · c_t`
as a `Float32Array`. Fit worker emits a `frame-preview` with
`stage: 'reconstruction'` at `framePreviewStride` cadence (default
2 to match W1's preview cadence via `run-control`).

`run-control` now also listens to fit's `frame-preview` posts and
routes them into the same `latestFrames` signal that W1 posts land
in, so the 4-canvas panel (T7) can read all four stages uniformly.

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

New `FrameQuad` component composes four `FrameStagePanel` canvases
in a 2×2 grid (design §8 "Frame panel"):

  raw             · hot-pixel
  motion-corrected · reconstruction (Ãc)

Each panel reads a stage from the `latestFrames` signal populated
by W1 (raw / hot-pixel / motion) and W2 (reconstruction). Caption
shows the current frame index + epoch from the dashboard store.

Scrubber is deferred to a later polish task — needs main-thread
frame history per stage, which is a separate data-plumbing change.

`SingleFrameViewer` stays in the tree as a back-compat surface
(reads `latestFrame` = motion stage) but is no longer rendered by
`DashboardLayout`.

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

New bus event `trace-sample` carries `(t, ids, values)` — fit
emits one per vitals-stride frame after the per-metric emissions.
Archive subscribes, routes samples into a new `NeuronTraceStore`
(drop-oldest ring per neuron id), and exposes them through a new
`request-all-traces` query.

Supporting bits:
- New WASM binding `Fitter.componentIds()` so samples carry the
  correct ids even when extend inserts / removes components across
  cycles.
- Trace samples are *excluded* from the event-log ring and the
  neuron-event index so they don't flood the feed or the
  structural-history queries — the traces store is their sole
  sink.
- `archive-client` gains a typed `requestAllTraces(idFilter?)`
  call, routing a new `all-traces` reply with parallel
  `ids[] / times[][] / values[][]` arrays back to callers.

T9 wires the TracesPanel uPlot chart on top.

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

Adds `uplot` dependency and a `TracesPanel` component that polls
`requestAllTraces` every second, merges per-id time/value arrays
into a single aligned uPlot frame, and renders one stroke per
neuron with a stable hue-hash color. Clicking a trace selects
that neuron via a new shared `selectedNeuronId` signal —
`FootprintsPanel` (T11) and the per-neuron zoom (T12) will read
it to stay in sync.

Dashboard grid gains a `traces` row under the frame panel so
the chart shares the main column with the 4-canvas viewer.

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

Two complementary pieces feeding the upcoming footprints panel:

1. **Main-thread running max projection**. W1 already posts the
   motion-corrected frame as a `frame-preview` for the 4-canvas
   viewer; `run-control` now folds each motion frame into a
   shared `maxProjection` signal (element-wise max over the u8
   preview). Resets when a new run starts. Kept main-thread —
   W1 already owns the data and no archive round-trip is needed.

2. **Archive `request-all-footprints`**. Returns the most recent
   sparse `A`-column snapshot per *live* neuron. Live = latest
   structural event is not a deprecate, via a new
   `NeuronEventIndex.liveIds()` helper. Client gets a typed
   `AllFootprintsReply` with parallel `ids / pixelIndices / values`
   arrays.

T11 wires both into the `FootprintsPanel` overlay.

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

New component blits the running max projection into a canvas and
strokes each live footprint's 4-connected boundary in its per-id
color. Clicking a boundary pixel (or interior pixel — hit-test
looks up the sparse support) calls `setSelectedNeuronId`; the
selected id gets a thicker white outline so it pops against the
color wall.

Dashboard grid gains a third column: frame quad | footprints |
events, with the traces strip chart spanning the two left columns
underneath.

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

New `NeuronZoomPanel` opens above the event feed whenever
`selectedNeuronId` is non-null. Polls `requestFootprintHistory` +
`requestAllTraces({idFilter: [id]})` every 2s and renders:

- bbox-cropped grayscale footprint shape (bbox + 2px padding)
- sparkline of the neuron's most recent trace samples (reuses the
  existing vitals `SparkLine` for visual consistency)

Click-through wiring: traces panel → `setSelectedNeuronId`, footprints
panel → `setSelectedNeuronId`, zoom's × button clears it. No modal
overlay — the panel renders inline in the events column so the
event feed just shrinks while a neuron is being inspected.

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

Pulls the current footprints + traces from the archive worker and
packs them into a scipy.sparse-CSC-shaped `.npz`:

- `A_data` / `A_indices` / `A_indptr` / `A_shape`: footprint matrix
  as CSC — `scipy.sparse.csc_matrix((data, indices, indptr), shape)`
  loads it directly.
- `footprint_ids`: parallel id vector.
- `C` (K×T), `C_times`, `C_ids`: dense trace matrix padded with NaN
  on the union time axis.
- `height` / `width`: frame geometry.

Events are intentionally omitted from the NPZ for now — JSON-in-zip
is awkward; the structural event log lives in the UI feed only.

Supporting bits:
- `@calab/io` `writeNpy` now dispatches on dtype (Float32 / Uint32 /
  Int32), and a new `writeNpz(arrays)` helper uses `fflate.zipSync`
  as the inverse of the existing `parseNpz`.
- `run-control` keeps the archive worker reference alive after
  natural run completion so export still works in the `stopped`
  state; `stopRun` clears it explicitly.
- New `ExportButton` lives next to the vitals bar. Disabled when
  no archive worker is available; shows an "Exporting…" indicator
  while polling + zipping.

Two-pass (T13) + run-mode toggle (T14) from the original Phase 7
task list were descoped to Phase 8 — full implementation requires
cross-worker `Footprints` state transfer that's non-trivial.

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

End-to-end test `apps/cala/e2e/phase7-exit.e2e.test.ts` drives
the real W1/W2/W4 workers against the test AVI and asserts every
Phase 7 wire-protocol deliverable lands intact:

- real `birth` events on the bus via `drainApplyEvents` (T1-T3)
- 3-stage W1 preview streams: raw / hotPixel / motion (T5)
- W2 reconstruction preview frames (T6)
- `request-all-traces` returns per-id trace samples (T8)
- `request-all-footprints` returns live-id sparse-A (T10)
- `buildCalaExportNpz` round-trips through `parseNpz` with the
  expected CSC + K×T shapes (T15)

Stub Fitter now returns real ids from `componentIds`, populates
`lastTrace` with a fixed amplitude, and emits a non-empty
reconstruction frame so the preview-stride path actually posts.

Design doc §12 updated with the Phase 7 exit status, explicit
deferrals (two-pass, scrubber, events-in-NPZ, Playwright, extend
tuning), and the callouts from the live-testing session (loose
redundancy gate + missing SlowBaseline seeding together drive the
+4 cells per cycle behavior on real recordings).

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