Signal viewer: stable y-scaling, width-relative detail, multi-panel layouts — plus a written visual language

[RaulSimpetru]

Three things that grew out of each other: making the y-axis stop jittering, making the
"how much to draw" control mean something, and making one stream viewable through several
panels. The last one exposed how much of the visual language lived only in the code, so it
is now written down and partly enforced.

Scaling

• Auto no longer hands the axis to ImPlot's per-frame auto_fit. The range snaps out
  instantly to contain a new peak (never clipped) but contracts over ~5 s, so a variable
  signal in a small window stops zoom-jittering. Per-channel mode gets the same easing — its
  lane ranges no longer breathe — and its range scan is vectorised (~15× cheaper at 256 ch).
• Per-Ch is now a scaling basis, not a mode that greys out Auto/Manual. Per-channel
  Manual freezes each lane's range so a weakening channel stays visible against its captured
  reference; Rescale re-fits and locks; Gain is live in per-channel Manual, inert in Auto.
• Range fitting rejects artifacts by duration, not amplitude ("Artifact", default
  20 ms), because a 10 ms artifact and a 10 ms real event are indistinguishable by amplitude
  alone. Mode-aware, per-channel then unioned, ~11 ms at 256 ch. 0 ms restores plain min/max.

Detail

The fixed "Point cap" slider (100–10000 pts) becomes "Detail": draw density relative
to the plot width, shown as a percentage of full. The old cap fought the width-derived target
(min(n_pixels, width × 3)), so the top half of its range was inert on a typical plot and the
default under-resolved wide ones. n_pixels survives as an optional hard-cap override
(default: no cap).

Multi-panel

• Several viewers can share one stream. State and ImGui ids key off a new widget_id
  (defaulting to stream_name) rather than the stream, so one panel per electrode grid works,
  each keeping its own channels, scale, filter and pause. Previously every SignalViewer("emg")
  resolved to one state and rendered identically. This also restores design-principle 8 —
  state keyed by widget identity, not by the data it happens to show.
• Panel chrome collapses (title, controls, channel bar, footer) behind the header toggle,
  leaving a tiled panel almost all plot.
• The channel-grid selector tiles grids near-square (6 → 3×2) instead of one tall stack.
• Heatmap.ui(vrange=…) — an explicit colour range. Several heatmaps must share one to be
  comparable; with per-instance autoscaling a quiet electrode array and a loud one render
  identically.

Visual language, and two bugs it caught

docs/concepts/visual-language.md writes down what _theme.py / widgets/common.py already
did — type stack, the tint+underline selection cue, panel headers, plot styling, units, icon
meanings, pop-out vs collapse, widget identity. It sits beside design-principles.md, which
stays the code contract.

Writing it down made parts of it checkable, and tests/test_visual_language.py (AST-based, no
allowlist) found real bugs:

• RawSignalViewer rendered as stock ImPlot — chart border, opaque background, heavy grid
  — for want of one ensure_implot_style() call.
• 19 hardcoded colours across 8 widget files could not follow the theme. Several were wrong
  on the light theme specifically: a near-white session label, a white-on-white channel-grid
  hover outline, and a prediction readout that flashed toward white — into the card it sits
  on. They now read the theme (muted() / primary() / hairline()) or a named token.
  Colours deliberately fixed in both themes (console surface, status pill) are named in
  common.py rather than inlined.
• ProcessLauncher used its own red/green instead of the shared DANGER / SUCCESS.
• The panel chrome toggle didn't swap its glyph, unlike every other toggle here
  (PLAY↔PAUSE, autoscroll, pop-out), so it read as a one-shot action.

Rules needing judgement — is this the right cue, does this comparison need a shared range —
are deliberately left to review.

CLAUDE.md points agents at both contracts before they touch a widget.

Notes

• 696 tests pass, ruff clean. Both enforcement rules were verified to fail when a violation is
  reintroduced.
• tests/test_stream_lsl.py::test_stream_reconnect_swaps_buffers_atomically is flaky under
  full-suite load (LSL multicast contention) and passes in isolation — pre-existing.
• Known gap: initial_channels seeds a viewer's channels but does not restrict them, so
  in a tiled per-grid layout "All" still enables every channel and [Edit…] lists every grid.
  A real channel_scope parameter is the fix; left for a follow-up.