[Sad Catapella B3][Performance] TIFF-to-Zarr conversion rereads the converted pixels to compute histograms #79

Baseline: d7a538e56d41ac1093ba79adcad397c21b3dfd32. Classification: confirmed redundant I/O by design; benefit of a fused pass is a performance hypothesis.

Evidence: _CopyStaging.convert() lines 885–897 reads TIFF planes and writes Zarr. run_experiment() lines 600–610 then invokes _read_file() on the new Zarr copy for histogram calculation. Existing test asserts each TIFF is read once for conversion and the Zarr copy is read again for histograms: intended behavior. It should not be altered without preserving the invariant that the table describes the persisted converted pixels.

Optimization plan: investigate a fused, chunk-aware stream that feeds level-0 writer and histogram accumulator from the same canonical pixel blocks, with a post-write integrity check to ensure copied data and histograms match. Consider direct concurrent accumulation and conversion only if retry/rollback, channel-ordering, clipping, endianness and range-check multi-pass behavior can remain exactly equivalent. For 32/64-bit and float robust range, extra quantile passes are still necessary unless spooling/compact intermediate state is used. Benchmark source TIFF reads, Zarr reads/decode calls, disk traffic, wall time, peak RSS and end-to-end CSV equivalence. Do not add Numba to the I/O orchestrator.