Hypermutator validation — SBJ02862 real >450k rerun (sash#59 Step B) #68

Tracking the hypermutator validation path end-to-end on real data — pulled out of
#62 because it's the one thread with real multi-session narrative, not a single
PR's diff.

Correction 2026-08-08 — the ">450k somatic PASS" framing in this issue's title and
history rests on a wrong figure. SBJ02862 was recorded here as having 522,920 raw PASS
variants. That number is wrong; the real count is 252,692, which is below 450k.
Confirmed three ways: biodaily#199's own table (9,132→252,692 (+243,560)), the locally
archived baseline extract from #199, and a direct record count of the run's
L2201449.pass.vcf.gz. 522,920 does not appear in any source data — it originated in this
issue body and propagated outward from here.

What this does and does not change: the PAVE/PCGR-skip path was exercised end-to-end
on this sample and completed cleanly, so the fix verification below stands. What is no
longer established is that it was verified on a genuine >450k-somatic-PASS sample.
Whether a true >450k sample is still required for sign-off is an open question, not
resolved here. (For reference, SBJ02862's rescued VCF does hold 1,686,410 records — the
">450k" language may have been reaching for a different metric than somatic PASS.)

Related

• umccr/biodaily#199 — original hypermutated-OOM failure (SBJ02862, PCGR OOM inside bolt)
• PAVE_SOMATIC MNV guard did not fire on SBJ02862 — PAVE 1.8 crash on an MNV that should have been filtered #66 — PAVE 1.8 crash found on the fix-verification rerun (different bug, same sample)
• Release/0.7.1 #60 — carries the sash#66 fix (e2686e8)
• Release validation — sash 0.7.x PR stack #62 — parent release-validation checklist ("hypermutated OOM resolved" checkbox)
• umccr/biodaily#227 — cohort run record

History

1. biodaily#199: SBJ02862 (tumor L2201449/normal L2201450, 252,692 somatic PASS
   variants — see the correction above) OOM'd inside PCGR on the old stack.
2. bolt#33/0.70 test dependency #36 fixed the OOM path (single-chunk PCGR guard, VCF2MAF-skip) — confirmed on real
   SEQC-II data 2026-07-19, but SEQC-II isn't hypermutated, so this only proves the skip logic
   doesn't break the normal case.
3. No fresh high-variant-count sample was recoverable for a while — old ones (L2100242,
   SBJ02862) were purged/archived; highest fresh prod sample at the time was well under the
   trigger threshold.
4. SBJ02862 was re-sourced for the biodaily#227 cohort re-run (2026-07-31, sash 0.7.1 + bolt 0.3.2).
   It hit a new, unrelated failure: PAVE 1.8 crashed (StringIndexOutOfBoundsException in
   ProteinUtils.determineContext) on an MNV the module's own guard should have filtered. Filed as
   sash#66, root-caused (pipefail/SIGPIPE race in the MNV guard), fixed in commit e2686e8
   (already in PR sash#60).
5. Rerun on e2686e8 (biodaily199sash071sbj02862) completed successfully 2026-08-05,
   23 tasks succeeded / 0 failed. PAVE_SOMATIC processed the chr6:100449426 variant
   without crashing. The original pre-fix run was checked in its rotated log and crashed
   exactly as filed in sash#66, confirming the fix — not chance — changed the outcome.

Outcome

• Hypermutator path confirmed end-to-end on real data: PAVE completes without the PAVE_SOMATIC MNV guard did not fire on SBJ02862 — PAVE 1.8 crash on an MNV that should have been filtered #66
  crash, PCGR-skip/VCF2MAF-skip fires correctly, cancer report renders, output complete.
  Caveat: verified on SBJ02862 at 252,692 somatic PASS, not on a >450k sample — see
  the correction at the top.

Closed 2026-08-06. sash#62's "hypermutated OOM resolved" checkbox is checked, citing this rerun.