TracePreview rendered traces as "Cell 0, Cell 1, …" while the rest of
the UI is 1-indexed (`CellCard` shows `Cell {cellIndex + 1}`,
`CellSelector` labels its input "Cell indices (1-indexed)"). Align the
import preview with the app-wide convention.

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

Two issues in the spectrum panel that showed up during parameter tuning
and dataset re-imports:

1. The outer chart-build effect tracked the whole `spectrumData` signal.
   Cutoff-only updates (from tau changes) produced a new `spectrumData`
   object identity, so every tau tick tore down and rebuilt the uPlot
   chart instead of just repainting the two cutoff lines. Split into
   (a) a rebuild effect keyed on the `freqs` typed-array reference
   (stable across cutoff-only updates) and (b) a redraw effect keyed on
   cutoffs + filter toggle. The filter-band plugin now reads cutoffs
   live via accessors, so a plain `uplot.redraw()` picks up new values.

2. `initSpectrumStore()` registered effects and a debounce `setTimeout`
   on module-level state with no `onCleanup`. Re-importing a dataset
   left the previous timer pending (firing into stale state) and
   retained `psdCache` PSDs for cells that no longer existed. Add an
   `onCleanup` for the debounce timer, plus an exported
   `resetSpectrumStore()` that clears caches and last-seen refs; wire it
   into the reset path alongside `clearMultiCellState` / `resetImport`.

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

`CaTuneZoomWindow` recomputed raw-trace mean/std and deconvolved
min/max from typed-array scans inside `createMemo`s. Because props flow
through the multi-cell store (which replaces the whole `CellTraces`
object on every solver write), the memos saw identity churn per tick
and re-ran on every visible card for every solver iteration — an O(n)
scan × N cards × tick cadence that Peak/FWHM drags felt in the
scripting cost.

Move the work next to the data:

- `CellTraces` gains a required `rawStats: RawTraceStats` (mean, std,
  zMin, zMax) and `deconvMinMax: [number, number]`. Helpers
  `computeRawStats` and `computeArrayMinMax` live in the store module.
- `updateOneCellTraces` fuses a single min/max pass with every solver
  write, so deconv min/max is up-to-date in O(n) per write (as opposed
  to O(n × visible cards) per write).
- `cell-solve-manager` computes `rawStats` once at initial cell
  insertion — raw traces never mutate during a session.
- `CardGrid` / `CellCard` / `CaTuneZoomWindow` thread the two fields
  through. The per-card `rawStats` / `deconvMinMax` / `pinnedDeconvMinMax`
  memos are removed. `gtSpikesMinMax` memo stays — GT data is stable
  and infrequent.

Test helper `seedCellTraces` updated to the new required fields.

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

CI format check caught two lines that fit within the project's width
after the earlier edits.

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

The previous refactor split the chart setup into a rebuild effect keyed
on a `createMemo(() => spectrumData()?.freqs)` and a redraw effect keyed
on cutoffs. In practice this left the chart blank on first load (likely
an HMR/tracking interaction around the memo's undefined→Float64Array
transition interleaving with `container` signal updates from the Show's
ref).

Fold both paths back into a single effect that tracks spectrumData and
container, same shape as the pre-refactor version. Skip the
destroy-and-rebuild path when the freqs typed-array reference matches
the last build (cutoff-only updates): just call `uplot.redraw()` and
let the filter-band plugin pick up the new cutoffs via its live
accessors. Keep a smaller redraw effect for the filter-enabled toggle.

Same perf win for cutoff scrubs, without the initial-load regression.

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