[Sad Catapella B8][Performance] Notebook plan rebuild repeatedly stats every TIFF and reparses the manifest on settings edits #102

Basis

Batch 8 at d7a538e56d41ac1093ba79adcad397c21b3dfd32. Confirmed repeated I/O access pattern, latency impact is a benchmark hypothesis.

Evidence

• _build_run_step lines 1765–1778 attaches _refresh_run_step() observers to multiple independent widgets (bins, quantiles, memory, force-rebuild).
• _refresh_run_step() lines 2021–2064 computes the full plan and calls _refresh_conversion() on every refresh. The plan _plan_html() lines 2184–2197 reparses the full prior run_manifest.json if present (can be very large for many FOVs).
• When conversion is enabled, _refresh_conversion and _space_warning_html lines 1926–1954 recalculate estimate_conversion, which iterates all TIFF files and calls stat() for each plus filesystem-space checks.
• The notebook does cache _plan_layouts() lines 2066–2100, so this issue targets only the uncached manifest parsing, conversion estimate and space rechecks—not reopening every plane layout.

Fix / acceptance

Cache manifest header/count for the current output folder until output path changes or the run publishes new outputs; cache source sizes and conversion estimates keyed by validated preflight snapshot plus pyramid settings (never cache across a new preflight); update only threshold-dependent text when unrelated controls change. Account explicitly for acquired files changing during a session (Batch 5 #90): preview is advisory, final run always revalidates sources. Benchmark 100, 1k, 10k and 100k TIFFs, 10 rapid settings changes, local/remote storage, wall time, stat calls and peak RSS; ensure identical plan text and invalidation. Numba not appropriate.