Separate dense SV discovery from complete-input candidate support

[iskandr]

Dense SV inputs can exhaust the record budget before an event is discovered, and reported support is then restricted to the discovery subset. Add opt-in dense_support=True / isovar sv-rna --dense-support to separate those stages: spool the complete supplied alignment input on disk, select deterministic event-first sequence/placement representatives under the discovery budget, then recount direct junction and exact full-path/ORF witnesses across all eligible originals.

Original SAM records, query intervals, missing QUAL, visible mate-label conflicts and input-scoped cell/UMI and signal ancestry remain auditable. Distinct sequence alternatives stay separate. Discovery votes, omitted records, query/path/extension limits, segment-grouping limits and bounded ORF-start/inclusion assessments remain explicit. Normal API/CLI defaults and schemas stay compatible. The temporary spool is cleaned on success and failure; retained support witnesses and output can still scale with actual event support.

Validation: ./lint.sh and ./test.sh pass (5,699 tests; 97% coverage). Regressions cover a late rare junction, support invariance across record order/discovery budgets, distinct sequences, repeated/alternative records, conflicting mates, missing QUAL, both partner orientations, distant signal ancestry, interrupted scans, CLI/ORF export, and the original ATP5MG–KMT2A fixture's pinned peptide. A synthetic million-background-record run discovers the junction with two discovery records and counts all 13 event fragments in 37.4 seconds at 160.7 MiB peak process RSS; it is a background-scale check, not a Sid pilot throughput claim.

Version 1.44.0, following the minor-release policy for additive features (draft #433 currently uses 1.43.0). Addresses #437's reconstruction/support separation in the library and CLI; adoption and a Sid pilot rerun in draft #433 remain follow-up work. The reproduced default SAM-payload identity collision is tracked separately in #438; dense mode rejects conflicting originals instead of choosing an arbitrary payload. Design and measured evidence.

[coveralls]

coverage: 96.547% (+0.02%) from 96.531% — feature/437-dense-sv-support into master