6.4.0 — wrong-length indel rule, orchestration fail-fast, splice-aware RNA evidence

The changes below alter reported counts at wrong-length indel loci and at
spliced RNA positions — the minor version bump this release delivers.

⚠️ Changed — span-aligned REF testimony at spliced deletion loci (user-visible)

• Structural REF evidence survives fragment consensus at quality zero.
  Span-aligned REF's carried quality is the first exon base after the
  junction — exactly where BAQ can zero it — and FragmentEvidence gated
  REF existence on quality > 0, silently dropping such fragments from
  rdf (the mirror of the structural-ALT divergence fixed in the
  splice-aware change below). Span-REF observations now carry a structural
  flag that keeps them alive in consensus; a structural ALT in the same
  fragment still wins. On the FORTE locus this recovers one real fragment
  (rdf 726→727 = dpf).

• Junction reads that observe every deleted-span base now count REF at a
  pure-deletion locus whose anchor base is spliced out (typical for deletion
  annotations left-aligned onto the last intronic base at an acceptor). The
  anchor base is not the discriminating fact for a deletion — the span is —
  and these reads demonstrate the span is present. Previously they counted
  DP as neither (conservative cluster-A behavior): at a validated FORTE
  acceptor locus, 823 of the 824 anchor-spliced junction reads convert
  (rd 24→847 of dp 848, fragment rdf 23→727 of dpf 727, DP
  unchanged); the one
  residual read's exon2 alignment covers only one of the two deleted-span
  bases (95M92N1M) and honestly stays neither. Guards: the FULL span
  must be aligned (partial coverage stays neither), and reads carrying a
  competing shifted/wrong-length indel candidate keep their existing
  arbitration paths. Read- and fragment-level, mirrored in both engine
  paths via the shared checker.

⚠️ Changed — consensus intron snipping of ref_context removed (user-visible)

• ref_context is always genomic now. The RNA-mode step that drained
  consensus introns from the variant's reference context before Phase 3 had
  no coordinate map: the context shrank while ref_context_start stayed
  genomic, so shifted-indel sequence verification, the large-deletion band's
  context guard, and haplotype offsets all read garbage right of a snipped
  intron — so exon-contained reads were scored against haplotypes whose
  offsets no longer matched their genomic coordinates, and pre-mRNA /
  intron-retention reads against a haplotype missing bases their sequence
  genuinely contains. Removing it
  fixes windowed shifted-deletion matching near acceptors (contract test
  test_windowed_deletion_after_junction_in_repeat_rna, committed red) and
  erases the binned-vs-legacy RNA divergence around it; junction reads
  needing alignment-based scoring stay conservatively neither under the
  splice-aware evidence rule. Re-validated on local (non-repo) clinical RNA
  data — the b37 dedup and FORTE hg38 smoke loci report unchanged counts. A
  coordinate-mapped spliced-haplotype rework remains tracked in issue #94,
  gated on real-data measurement; it must keep pre-mRNA / intron-retention
  reads genomically scored.
• mq0_count now tallies before the RNA strandedness filter in the binned
  path, matching the legacy path: an antisense MAPQ-0 read is still a
  physical read at the locus, and the two paths previously diverged on this
  diagnostic in stranded RNA mode.

⚠️ Changed — splice-aware evidence in RNA counting (user-visible)

• Reads spliced over a variant no longer count depth or REF. A CIGAR N
  spanning every discriminating position (a deletion's deleted span, an
  insertion's junction flanks, an SNV/MNP's REF bases) is asserted splicing —
  no observation — so the read classifies neither AND is excluded from
  DP/fragment depth, matching samtools pileup's zero coverage inside an N gap.
  Previously such reads counted definitive REF (deletions) or inflated DP with
  zero aligned bases (intronic positions): on a junction-dense real RNA locus,
  DP at an intronic SNV drops from ~826 span-overlapping reads to the 2 with
  aligned bases there. Excluded totals are logged per variant at debug level
  (Phase stats … splice_skip_excluded=).
• The REF/ALT call no longer flips on the aligner's D-vs-N representation
  choice. A D(100) at the expected span is deletion evidence (ALT); the
  same gap as N(100) is splicing (excluded).
• Indels directly after a splice junction are now examined. An indel op
  whose only neighbor is a splice N (M-N-D-M / M-N-I-M — an event at an
  exon boundary reached through the junction) was structurally invisible to the
  strict and windowed CIGAR scans; carriers now classify through the same
  anchor/windowed inspection as M-adjacent ops (evidence attributed to the
  nearest aligned/inserted base when the anchor itself is spliced out).
• Phase 3 no longer scores across a splice. Raw-window extraction refuses
  windows that an N overlaps, and check_complex's reconstruction classifies
  such reads neither instead of stitching exon arms into a junction-chimeric
  sequence in which the missing intron reads as deletion evidence.
• The exclusion is not silent. Deletion-type loci where splice-skip
  exclusions exceed confirmed ALT are flagged SPLICE_SKIP_DOMINANT(n) in
  gbcms_diagnostic — RNA aligners write large deletions as splices (STAR:
  ≥ alignIntronMin, default 21bp), so alt_count=0 there may mean the
  carriers exist as junction reads. Per-variant totals also appear in the
  debug-level Phase stats line (splice_skip_excluded=).
• Shifted representations across a junction stay in classification. A
  read whose N covers the annotated span but which carries an I/D op inside
  the scan window (repeat-tract shift) is deferred to the windowed scans
  instead of being excluded.
• Fragment consensus recognizes structural ALT independent of base
  quality. BAQ can stack splice and indel penalties to quality 0 on the
  base carried with a junction-adjacent I/D op; FragmentEvidence::resolve
  previously treated qual-0 ALT evidence as no evidence, reporting ad > 0
  with adf = 0 at the same locus.
• check_complex refuses string comparison for structurally anomalous
  junction reads (splice N inside the context window): with the indel
  shifted outside the variant span, the span reconstruction is clean REF
  sequence and Phase 2 would absorb an ALT carrier into rd.
• ASJD junction tallies at intronic splice-region variants now reflect only
  allele-informative reads (spliced reads with no observation at the locus
  no longer contribute fabricated REF junctions); expect LOW_REF_JUNC
  where the REF tally was previously fed by such reads.
• Contract battery: tests/test_rna_splice_contract.py (13 cases incl. a
  legacy-parity case with N-CIGAR reads and one strict xfail pinning the
  known D6 ref_context coordinate corruption for the cluster-B fix).
  DNA-mode classification is untouched (reads without N ops never enter
  the triage), and the legacy parity oracle mirrors every engine-loop
  change.

⚠️ Changed — orchestration fail-fast (user-visible)

• --bam-list entries that do not exist now fail the run (exit 1) unless
  --lenient-bam is given — matching the long-documented fail-fast promise. An
  unreadable list file is equally fatal: previously both cases silently ran a
  partial sample set and exited 0.
• Duplicate sample names are now a hard error (from --bam, --bam-list, or
  a mix): the later BAM silently replaced the earlier one, which was then never
  processed. Use sample_id:path (or two-column list entries) to disambiguate
  deliberate same-stem inputs.
• Ragged MAF rows now raise in the batch readers (gbcms merge inputs)
  instead of having their overflow fields silently truncated — the gbcms count
  columns are the trailing columns, exactly what truncation dropped.

🔧 Fixed

• A rejected (FAIL) variant no longer crashes --mfsd, RNA --gtf, or
  --mfsd-parquet runs: the zero-count stub now carries every column the
  writers read (mFSD q-value and nucleosomal fractions, RNA/GTF annotation and
  ASJD fields), and the mFSD Parquet excludes rejected variants (they have no
  fragment data) with the exclusion logged.
• gbcms merge no longer coerces float-formatted count strings to 0: a
  pandas/R round-trip renders integers as 12.0, which the combined-column sums
  silently nulled to 0; genuinely non-numeric values (e.g. NA) are counted and
  warned about per column.
• Re-genotyping warns about replaced columns: an input MAF already carrying
  gbcms output column names (e.g. ref_count from a previous run) has those
  values refreshed — now with a warning naming every colliding column, and the
  writer's docstring no longer claims originals are never overwritten.
• Failed-sample reports name the exception type and log the traceback
  (KeyError: some_key instead of a bare 'some_key').

⚠️ Changed — wrong-length pure-indel evidence is a distinct allele (issue #91)

• A read whose CIGAR proves a pure indel of a DIFFERENT length at the variant anchor now
  counts as partial_alt, never as REF or the queried ALT. In repeat tracts, coexisting
  distinct-length indel populations are distinct slippage alleles; the previous behavior
  routed them to the length-blind Phase-3 haplotype window, which promoted them to ALT and
  inflated VAF several-fold at tract loci (a homopolymer locus with 7 true-ALT reads
  reported 259). alt_count now reflects exact support; the distinct-allele evidence
  surfaces in partial_alt/any_alt, and PARTIAL_DOMINANT flags loci where it dominates.
• Placement-aware ≥50bp deletion band replaces the reciprocal-overlap tier: a
  wrong-length D at the anchor is the annotated event only when the read deletes essentially
  the whole expected span (≤3 retained bases, ≤3 changed outside) — accepting breakpoint
  wobble and split D+M+D representations while rejecting displaced net-matches. The old
  ≥50% overlap rule accepted any deletion sharing half the annotated length (its sequence
  check compared reference against reference, identically true by construction).
• Insertion truncation containment: a shorter insert that is a ≥90%-identity slice of a
  non-low-complexity expected insert still counts as ALT (sequencer truncation smear of long
  insertions); both sequences must be non-low-complexity so repeat-tract slippage is never
  mistaken for truncation.
• Same-length insertions with confidently mismatching bases (≥--min-baseq) are a
  third allele → partial_alt; they were previously absorbed into ref_count with no
  signal. Unverifiable cases (all inserted bases low-quality) go to Phase-3 arbitration.
• Windowed wrong-length ops (deletions ≥5bp — 1–4bp windowed Ds remain
  CIGAR-definitive alignment noise; insertions at any size): repeat tract → partial_alt;
  unique context → REF with the stray op surfaced as partial_alt. Same-length S3-fail candidates keep the Phase-3
  left-alignment rescue (TP53-class), now via the split has_shifted_same_length flag.
• Repeat scan anchors at the first changed base (not the shared VCF anchor, which sits
  one base left of a left-aligned tract), and adaptive context padding covers the whole
  tract plus flank — Phase-3 haplotype windows in repeat regions are no longer too narrow
  to distinguish tract lengths.
• Delins/complex variants are unaffected: they route to check_complex, whose Phase-3
  realignment correctly resolves split and mismatch-absorbed representations of one event.

🔧 Fixed

• --trace never emitted a single per-read Rust trace line: pyo3-log's default filter
  capped forwarding at DEBUG, and the Python side configured a logger name (gbcms_rs)
  that pyo3-log never uses (_rs.…). Both fixed; _rs.reset_log_caching() exposed so
  enabling trace after Rust code has logged is not silently ignored.
• Left-alignment failures are loud: wide-window FASTA fetch failure warns with the
  error (reachable near contig ends) instead of silently skipping normalization; the
  2500bp expansion cap binding without convergence warns; non-UTF-8 alleles from a corrupt
  reference keep the variant fully unnormalized instead of emitting a shifted position with
  reverted alleles; a stray println! on ref_context fetch failure removed.
• --enforce-strandedness warns when variants have no resolved gene strand (no GTF,
  uncovered locus, contig mismatch) instead of silently not enforcing.
• GTF annotation warns per variant chromosome with zero loaded exons, on both the
  text-parse and the --gtf-cache-dir cache-hit paths (a warm cohort cache previously
  silenced the warning after the first sample).
• Per-read debug! diagnostics demoted to trace! (WFA router, marginalized PairHMM).

✨ Added

• Wrong-length contract battery (tests/test_wrong_length_contract.py, 20 tests through
  the binned↔legacy parity oracle) and an end-to-end PARTIAL_DOMINANT reporting-chain
  test driving the full CLI.
• Documentation: "Wrong-Length Pure Indels" rule reference in
  docs/reference/allele-classification.md and a worked case study (Case 4 in
  docs/reference/complex-indels.md);
  partial_alt semantics updated in output docs; fictional RUST_LOG/GBCMS_LOG_LEVEL
  controls removed from docs (logging is --verbose/--trace).