Address presentation feedback: submitted-SCV corpus, benchmark rework, cleaning rules, positional drift

[davmlaw]

Responds to all seven items in the 2026-08-08 presentation feedback plan (the plan file is deleted by this PR now that it is complete).

Client code (CHANGELOG updated)

• Cleaning rules (Build a corpus of real-world bad HGVS strings to evaluate cleaning/resolution #112). Measurement of the bad-accession residuals showed wrong separators, gene+space forms, dangling versions and leading junk dominate (genuinely dropped prefixes were only 2 rows). New string ops for each, plus a data-provider-aware resolve_missing_accession_prefix() that restores a dropped NM_/XM_/NR_/XR_ prefix only when exactly one candidate exists in the loaded provider. Production corpus: rescued 1,678 to 1,721, residual 1,118 to 1,075, zero regressions. 23 new tests.
• get_tx_exons cache (Do some benchmarks #5). Bounded per-instance LRU on the per-variant hot path, ~4% throughput gain, correct for lazily fetched REST providers.

Paper

• Submitted-SCV ClinVar corpus (new, Tier 1). build_clinvar_submitted_pairs.py builds 2.93M unique lab-submitted (AlleleID, HGVS) pairs from the VCV release. 81.8% cite a superseded transcript version; cdot holds 99.3% of those. On a seeded 3,000 sample cdot resolves 98.9% vs UTA 80.1% (the old current-version corpus is kept as an explicit ceiling row). Public residual taxonomy shows failures are version age, not formatting.
• Benchmark rework. The REST-beats-local inversion was a sequence-layer artifact (SeqRepo fd cache off; biocommons per-instance LRU memoising small sets). New methodology: identical 500-pair set, N=5 repeats, median (IQR). Local JSON 748 vs warmed REST 753 HGVS/s (equivalent within 1%); local UTA corrected to ~4x slower, not 30x.
• Mutalyzer facts corrected. Now matches Lefter 2021 body: 50.4% correct / 41.2% error / 7.1% auto-corrected per unique description; the unsupported "missing transcript data" clause removed; R4 motivation reframed around it.
• Related work on HGVS repair added to the discussion (VariantValidator, Mutalyzer, LOVD syntax checker, ClinGen Allele Registry, biocommons parser leniency), positioning clean_hgvs on offline use, auditable HGVSFix output and the no-regression guarantee.
• Positional drift (Figure S1). Conditioned on partial drift, preservation declines strictly monotonically toward the 3' end (RefSeq 98.9% to 37.5% by decile; Ensembl 93.2% to 17.1%); unconditioned curves are nearly flat, so most version risk is all-or-nothing. The ClinVar residuals show no 3' concentration (reported honestly). R5 gains the 5'-safer-than-3' fallback rule of thumb.
• Figure 1 panel B now shows a greyed UTADataProvider/PostgreSQL on the same biocommons interface.

Full test suite: 407 passed, 5 skipped. Paper quick build renders with all facts substituted.

🤖 Generated with Claude Code

https://claude.ai/code/session_015662nbf1iezxdL9EMcckU7