[Sad Catapella B3][Memory] Blockwise pyramid region selector does not enforce a viable minimum budget #76

Baseline: d7a538e56d41ac1093ba79adcad397c21b3dfd32. Classification: confirmed failure to enforce supplied memory bound at region selection; actual peak RSS not measured.

Evidence: OmeZarrWriter._region_side() lines 393–403 reduces side until it reaches 2, but does not check whether the estimated cost at 2 still exceeds memory_limit. _write_blocks() then reads, float64-converts and writes these regions. The estimate adds five full chunks; even if the region is 2×2, a Zarr partial write can involve full chunk processing. Existing test constructs a writer with memory_limit=1 and expects _region_side()==2, without exercising the read/write under that bound. Current ConversionSettings permits chunk sides from 16 to 8192.

Impact: The advertised memory limit can be violated during TIFF→OME-Zarr conversion, even when check_planes() approves source plane reading. Some configurations may exhaust RAM without a clear PlaneMemoryError.

Fix/acceptance: compute the minimum feasible region/chunk memory footprint before writing (including input chunk decode, output chunk RMW, float64 intermediates and compression buffers), and fail fast with a named, actionable error when the requested budget is impossible. Either implement a genuinely bounded alternative or demand larger memory/smaller chunks; never silently exceed the limit. Test positive/negative boundaries, large chunks with tiny budgets, full and blockwise source reading, and peak RSS. No Numba fix: this is allocation/I/O planning.