[wshlavacek]

Re-baselines parity_checks/nightly/baselines/differential.json. Every Path-agreement run since 2026-10-27 19:00 UTC fails on one alert, BIOMD0000000338|klu_vs_dense PASS → EXCEPTION. That includes main's own scheduled night, 40712357505, and the differential PRs #1315-#1323. The change comes from main, not from any of those PRs.

The file is replaced byte for byte by the verdicts.json of scheduled run 40712357505 on main at 9a26f5b5. Re-extracting that run's report.json with verdicts.py extract gives the same bytes.

• All 13,530 cases are on both sides, with none added or dropped.
• Three cases change outcome, and no others do:

case	before	after	why
BIOMD0000000338 klu_vs_dense	PASS	EXCEPTION	#1302: the dense jacobian="fd" leg fails on Linux (below)
BIOMD0000000385 ajac_vs_fdjac	TIMEOUT	PASS	#1321: per-check deadlines and the 1500 s cap
BIOMD0000000385 klu_vs_dense	TIMEOUT	PASS	#1321: per-check deadlines and the 1500 s cap

• The tally moves only there: PASS 11,225 → 11,226, EXCEPTION 7 → 8, TIMEOUT 20 → 18. DIFF 8, SKIP 2,217 and UNSUPPORTED 53 are unchanged.
• verdicts.py diff of that night against this file gives no alerts.

BIOMD0000000338: what changed it

It changed at a commit boundary, not with the runner. The case reads PASS on every run before 2026-10-27 18:31 UTC and EXCEPTION on every run from 19:00 onward. Runs on each side used AMD EPYC 7763 and 9V45 runners, and two Xeon runs after 19:00 fail too:

run	when (UTC)	CPU	klu_vs_dense
40521487203 (scheduled, main b543f7b)	10-27 15:23	EPYC 9V45	PASS
40537573193	10-27 18:16	EPYC 7763	PASS
40539428325	10-27 18:31	EPYC 9V74	PASS
40565116208	10-27 20:11	EPYC 9V74	EXCEPTION
40586213480	10-27 22:39	Xeon	EXCEPTION
40712357505 (scheduled, main 9a26f5b)	10-28 16:24	EPYC 7763	EXCEPTION

The first failing run is #1302's. The first failing run was created at 19:00 UTC, when the only bngsim change new to main was #1302 (merged 17:31 UTC). #1321's harness change, a timeout= on every run, is not it: on macOS the run is bitwise identical with and without a timeout, for KLU, dense and the default solver.

The failing leg is the dense one.

• klu_vs_dense runs force_sparse_linear_solver=True, jacobian="fd", then force_dense_linear_solver=True, jacobian="fd".
• The model has 57 species, so the default jacobian="fd" run is KLU, bit for bit the first leg. That run is ajac_vs_fdjac, and it passes on the same night (1.0e-7). So the leg that raises at t = 0.72 (CV_CONV_FAILURE) is the dense one.
• That is the matrix Difference quotients that step across a jump in a rate law (#1002) #1302 replaced: CVODE's own dense difference quotient became cvode_dense_dq_past_jumps, plus the jump test.
• Before Difference quotients that step across a jump in a rate law (#1002) #1302 the two legs agreed within 1.1e-4 on Linux.

It does not reproduce on macOS arm64. At 9a26f5b the dense leg completes, and Fg at t = 2 is 0.734, which agrees with libRoadRunner. The KLU leg ends at 311: #1304's branch split. With BNGSIM_JAC_DEBUG=1, neither leg's quotient has an entry the jump test mends.

The model is #1304's. KLU and dense take different branches on it past a step-size collapse, so a bit-level change in either quotient can move a leg. The new failure is a loud one: an exception, not an answer. Recording it here stops it alerting every night, and #1304 carries the finding.

A baseline-only change: the nightly's sweeps are not re-run for it.