[wshlavacek]

rr_ode left RoadRunner its default maximum_num_steps of 20000, a hard ceiling per output interval. bngsim's max_steps is also 20000, but it is a batch size: bngsim calls CVode again while the time advances, and raises only when a batch makes no progress (retry_while_advancing, #54). So at the same number, RoadRunner gave up on runs bngsim completed. #1654 found BIOMD0000000375 REFERENCE_FAILED for that reason alone. This PR gives RoadRunner 100000 steps per output interval (_rr_common.RR_MAX_STEPS).

What it changes

BIOMD0000000375 goes from REFERENCE_FAILED to PASS. At its TOL_OVERRIDES tolerance (rtol 1e-11, atol 1e-16), one 800-unit output interval needs about 21000 steps: 20500 fails, 21000 passes. With the budget, RoadRunner completes and passes against bngsim at the same tolerance (err/movement 3.4e-3) and at 1e-12 (2.8e-3). bngsim converges there too (against its own 1e-12 run, 5.8e-4).

No other row changes verdict. Measured with RoadRunner 2.11.0 on macOS arm64, on every row that ran out of steps at 20000 in the 0.15.1 paper reports or in the nightly baseline, at four budgets:

row	20000	40000 / 100000 / 1000000	RoadRunner stops at
BIOMD0000000375	REFERENCE_FAILED	PASS	completes
BIOMD0000000338	REFERENCE_FAILED	REFERENCE_FAILED	t = 0.0971 at every budget
MODEL1112050000	REFERENCE_FAILED	REFERENCE_FAILED	t = 10.64 at every budget
MODEL1112050001	REFERENCE_FAILED (adjudicated uncoverable)	the same; the error becomes CV_CONV_FAILURE	t = 3.36e-7, the overflow its adjudication describes
MODEL2001310001	BAD_TEST	the same; the error becomes CV_CONV_FAILURE	t = 43.59, where bngsim stalls too
MODEL2003050002	BAD_TEST	BAD_TEST	t = 0.0586
MODEL0995500644	BAD_TEST	BAD_TEST	an error-test failure at t = 0.377 on macOS

The other eight rows give the same verdict at every budget: six PASS, including 339's known artifact, and two TIMEOUT. The rows that stay are stalls, not slow runs, so a larger budget only makes them fail later in wall time: 338 takes about 1 s at 100000 and 13 s at 1000000. 100000 is also the budget the AMICI sensitivity harness gives both engines (#339).

Table 17

Only libRoadRunner's column moves: Failed −1, Validated +1, for 375. BNGsim's column is unchanged, because COPASI already passes 375 in the 0.15.1 reports.

The nightly

A run that completes comes out bit-identical under a larger budget: the review checked all 1323 jobs, 20000 against 100000. So the budget stays out of the archive's identity, where it would stale every entry at once. An entry where RoadRunner ran out of steps does depend on it, and CVODE names the budget in that message (mxstep (20000)). archive_load now treats such an entry, recorded under a budget other than RR_MAX_STEPS, as archive_stale.

Until the next recording:

• the Sunday live run passes 375 (REFERENCE_FAILED → PASS, not an alert);
• the replays show the rows the current archive recorded as out of steps (338, 375 and MODEL1112050001; on Linux also MODEL0995500644, MODEL2001310001 and MODEL2003050002) as REFERENCE_FAILED archive_stale. None was good, so none alerts;
• the Monday replay after a live night reports 375 as "archive gap after a live night, as the committed baseline records" (nightly: a replay after the live Sunday run alerts on rows the archive can't replay (rr_ssa MODEL2003200001) #1599), not as a regression. I checked this with the nightly's own extract and diff, on a live run passing 375 followed by a replay of a 20000-step refusal.

A recording settles it all; the committed baseline then re-records 375 as PASS.

RoadRunner with the new budget has not run on Linux: pull requests replay the archive. The review ran x86_64 RoadRunner 2.11.0 under Rosetta, and it fails on the same rows, with 375 needing the same steps (it fails at 20961 and completes at 20976). MODEL0995500644 there runs out of steps at t = 0.4029, at the same t at every budget, so it stays BAD_TEST.

Also

• 375's reason and its evidence test now say the row passes, and keep why it was REFERENCE_FAILED. The test fails with the budget set back to 20000.
• MODEL1112050001's reason names both errors RoadRunner stops with at t ~ 3.36e-7.
• rr_ode's timing config records max_steps, so an archived entry says which budget it ran under.
• ode_jobs.json re-baked with build_ode_jobs.py --mirror data/sedml_source/final. Only those two rows change, and _meta is untouched.
• The rr_parity README gains a "Step budget" section.
• ci/test_floors.env: parity floor 1381 → 1382, for one new test. Counted as CI runs: 1392 collected, 83 skipped.

Tests

• test_rr_reference_archive.py, one new test: an out-of-steps entry under another budget is archive_stale, and one under RR_MAX_STEPS replays. It fails without the rule.
• test_rr_known_artifact_evidence.py: 36 passed with the corpus.
• The parity suite with the corpus: 1346 passed, 46 skipped.
• A live rr_run.py on the 15 rows above gives the table's 100000 column.
• The pinned pre-commit hooks pass.

Review

One round. No HIGH findings, and every figure in the first commit reproduced. Check 3 asked whether 375's PASS checks anything: the engines agree at rtol 1e-13 (6.1e-5), and RoadRunner at 1e-11 is 3.2e-3 from its own 1e-13 run, so both sides are converged. Fixed:

1. MEDIUM: a false nightly alert every Monday. Sunday's live run would pass 375, and Monday's replay of the recorded CV_TOO_MUCH_WORK diffed as a regression. Fixed by the archive_stale rule above.
2. LOW: the evidence test pinned RoadRunner failing at 20000 steps. That pins RoadRunner, not the harness, and the test fails with the old budget anyway. Dropped.
3. LOW: the comment's list of stalls left out MODEL0995500644, and said "a budget changes only an entry that raised", which holds only for a larger one. Both corrected, and the budget is now in the timing config.
4. LOW: not run on Linux. Covered under Rosetta, as above. The Sunday live night is the first Linux run.