[Sad Catapella B7][Performance] IntensityPanel recomputes image pooling and distribution curves for every setting or channel change #97

Scope

Sad Catapella Batch 7, preprocessing main d7a538e56d41ac1093ba79adcad397c21b3dfd32. Confirmed repeated-work structure; latency/RSS unmeasured. This is distinct from normalization's leave-one-out algorithm cost (filed in its owning repository).

Evidence

• IntensityPanel._settings_changed() and _check() lines 2776–2809 recompute check_intensities(self.records,...) on every threshold change. check_intensities() reruns image_summaries(records), a histogram-pooling operation independent of thresholds, before computing warning flags.
• _draw() lines 2811–2822 invokes repeat_distributions on every threshold or channel change, rebuilding every repeat's PooledHistogram and its large histogram arrays, even though underlying records and cumulative distributions do not change with thresholds.
• IntensityPanel.show() can initialize the necessary cache for a specific table. Current notebook tests exercise rendering/updated settings but do not measure repeated calls or scaling.

Fix / acceptance

Split (a) immutable image summaries and per-channel pooled repeat distributions from (b) threshold-dependent warnings and (c) plot rendering. Cache summaries and the precomputed CDFs for the current table, invalidate only when show(records) or clear() changes data, not when outlier limit/saturation threshold/min_images changes; invalidate per channel only when needed. Avoid repeated full PNG redraw if only status/warning annotations change, where feasible. Do not cache by fragile id(records) across replacement; use explicit lifecycle. Compare exact warning flags/curves before/after threshold edits, channel switching, and replaced tables. Benchmark initial render and 10 settings/channel changes on 10²–10⁴ images, 1–8 channels, 256–65536 bins, measuring wall time and peak RSS; use no Numba in UI. No tests executed; no code changed.