[Sad Catapella N7][Performance] image_summaries eagerly retains all FOV histograms instead of supporting bounded-memory per-image aggregation #58

Baseline and classification

Normalization Batch 7 at f784c95e3049483f87466150ffc5764375fa631d. Confirmed eager-memory architecture; benefit of streaming is unmeasured. Coordinate with #49 (CSV reader/writer streaming), #34 (PooledHistogram merge allocations), preprocessing #77 (large histogram tables), and preprocessing #97 (notebook caching). No benchmarks or code changes.

Evidence

• image_summaries() lines 162–175 first calls records = list(records), holds all histograms, and then builds images: dict[key, list[PlaneRecord]]; iterating over a streaming generator does not bound memory.
• _summary() lines 178–197 materializes per-channel histogram references and invokes PooledHistogram, which allocates concatenated bin arrays for each image; the full table remains live while each FOV is reduced.
• Tests use small unsorted datasets and enforce deterministic sorted output; there is no streaming/large-FOV memory test.

Impact / scale

For 10,000 single-Z FOVs × 4 channels × 65,536 bins of int64 counts, retained bin arrays alone occupy ~19.5 GiB, before Python records and PooledHistogram temporaries. This differs from #49 (CSV reader/writer): even if the reader becomes streaming, this analysis function immediately materializes the complete iterator.

Optimization / acceptance

Provide an explicit API for per-image sorted/grouped record streaming, with a documented precondition that all Z/T rows of an image are contiguous and stable; on arbitrary unsorted data use bounded external sort/staging or retain an explicit eager compatibility mode. Emit one ImageSummary per complete FOV and free per-image histogram buffers promptly; preserve global deterministic ordering, channel validation, missing values and percentile definition. Because global channel-cardinality consistency may be learned late, design a fail-safe validation pass or staged publication; never silently emit partially valid results without a clear error contract. Compare eager vs streaming summary rows exactly and verify no duplicates/omissions; benchmark 1k–100k FOVs, 1/4/8 channels, 256/65,536 bins, 1/10/100 Z/T slices, peak RSS, throughput and sorted-vs-unsorted input. The principal gain is memory architecture, not Numba. Respect source/identity checks (#43, #45).