feat(detect): surface windowed match count/proportion in evidence

[Lioscro]

Summary

• Adds windowed_match_count: usize and windowed_match_proportion: f64 to ComponentEvidence. They sum hits across [pos ± remap_window] on the same mate, surfacing what cyto map --remap-window N would score for each component.
• Plumbs per-file PositionAccumulators into validate_and_aggregate so the aggregated path merges accumulators and re-walks at max_remap_window (centered on file 0's best position).
• Surfaces the new values inline in both stderr log functions. No CLI changes, no mapper changes, no stdout-contract change.

Why

Per-component evidence today reports counts at a single canonical position only. When a library has positional drift -- V2 GEX [:18] spacer drift, V1 [probe] jitter -- the headline proportion looks low even though cyto map --remap-window N would score most of those reads. The windowed values let users see the effective rate at the recommended window.

What this looks like

[2026-06-23T23:00:43.804Z INFO  cyto_map::detect] Detected geometry: `[barcode][umi:12][:13][probe] | [gex]`
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect] Recommended --remap-window: 2
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect] Detection sampled 400000 reads total (4 files)
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]   [barcode] R1 pos=0 count=371755 proportion=0.9294 windowed_count=372522 windowed_proportion=0.9313
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]     alt: R2 pos=17 count=19615
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]     alt: R2 pos=3 count=9423
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]     alt: R2 pos=31 count=6000
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]   [gex] R2 pos=0 count=371995 proportion=0.9300 windowed_count=371995 windowed_proportion=0.9300
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]   [probe] R1 pos=41 count=101140 proportion=0.2529 windowed_count=378998 windowed_proportion=0.9475
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]     alt: R1 pos=40 count=96612
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]     alt: R1 pos=42 count=90595
[2026-06-23T23:00:43.804Z INFO  cyto_map::detect]     alt: R1 pos=43 count=82843

Test plan

• cargo test -p cyto-map -- 76 unit tests (69 + 7 new) + 4 integration tests, all pass.
• cargo test --workspace green.
• cargo clippy -p cyto-map --all-targets --no-deps -- -D warnings -- zero new errors beyond the 10 lib + 1 lib-test pre-existing on base.
• Mutation experiment (two rounds): Mutation A (range-predicate collapse to *p == best_pos) catches tests 2, 3, 5, 6, 7. Mutation B (formula short-circuit to 0) catches tests 1, 2, 3 + integration assertion. Union covers every new test; production code restored after each.
• Manual fixture smoke (cyto detect gex and cyto detect crispr): windowed tokens emitted on every per-component line; probe shows windowed_count > match_count on V1 fixture; recommended-remap-window line still emitted; stdout contract unchanged.
• Integration test test_detect_gex_geometry_from_binseq asserts probe positional drift via strict windowed_match_count > match_count.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces windowed match counts and proportions to the geometry detection module to better estimate read counts within a remap window, addressing positional drift in short references like probes. It updates ComponentEvidence and PositionAccumulator to calculate and aggregate these metrics across files, and adds comprehensive unit and integration tests. The feedback suggests using saturating_add when calculating the upper bound of the window to defensively prevent potential overflow panics.