[Montana]

Fixes #1310.

Stacked on #1321. The first nightly here failed Path-agreement on TIMEOUTs only, at the cap on MODEL1006230049, 0053 and 0077, which any added check tipped over. #1321 bounds each check by the model's cap, keeps a killed worker's finished rows, and raises the job's cap to 1500 s. It is merged in here; until it merges, this PR also shows its commits.

Builds on #1291 (merged). It adds two rows to CHECKS, as #1315, #1316, #1318, #1319 and #1320 each add theirs, so whichever lands second keeps both. I'll resolve it.

What it adds

An SSA arm of the differential report: compiled SSA propensities against interpreted ones.

check	path	reference
ssa_cc_vs_interp	the default Simulator, which compiles the propensities of an eligible model	codegen=False
ssa_backend_as_asked	ssa_diagnostics["propensity_backend"]	what was asked: codegen=False interpreted, the default compiled (#1213's class)

codegen=True is refused for SSA, so the compiled side is the default Simulator. Each job runs n_rep replicates at seeds seed up to seed + n_rep − 1:

• every seed identical: PASS;
• an all-mass-action model, any seed differing: DIFF. Both sides compute the same doubles, so the event sequence must match draw for draw;
• otherwise: judged in distribution, by differ.ensemble_verdict at ensemble_k_for(n_rep), as the SSA parity suites judge. That branch is for when eligibility widens to functional laws; today none is compiled.

Cost, which shaped the design. Compiled propensities serve exact SSA only on mass-action models without events, up to a size (GH #190). The Simulator decides that before any run, and builds no propensity library for a model it finds ineligible. So:

• no library: both checks SKIP, and no replicate runs. On a first design that ran ensembles both ways, 31 of the first 98 models timed out at 600 s; even one replicate each way, 18 of 101 ran past the 900 s cap;
• a library the engine still declines: one replicate each way decides both checks;
• otherwise: the full ensembles.

Job gains method, n_rep and seed; checks_for(job) gives an SSA job the SSA checks only, and an ODE job none of them. The runner reads rr_parity/ssa_jobs.json (--ssa-jobs, '' leaves the arm out), each job at its own window and seed, capped at --max-reps (20).

     "backend_as_asked",
+    # The SSA arm (#1310), on SSA jobs only.
+    "ssa_cc_vs_interp",
+    "ssa_backend_as_asked",
 )

Over the corpus

rr_ssa's 268 jobs on disk, local run at the nightly's settings (--workers 4 --timeout 900): 34 s in all (median 0.12 s per model, worst 2 s).

	PASS	SKIP	DIFF	EXCEPTION	TIMEOUT
ssa_cc_vs_interp	23	245	0	0	0
ssa_backend_as_asked	23	245	0	0	0

The 23 are the models the default compiles: every seed of 20 is identical draw for draw, and the default ran cc. No finding.

Tests (python/tests/test_differential_shapes.py, appended)

• an SSA job runs the SSA checks and an ODE job none;
• compiled and interpreted mass action agree draw for draw (a dimerization, 10 seeds);
• negative control: one rate constant 1% off on the compiled side only reads DIFF, "mass-action". A change of 1e-9 moved no draw: it shifts each event time by as little, which crosses no sample;
• functional laws are judged in distribution: the same ensemble at other seeds passes, and one with a propensity tripled does not;
• codegen=False reporting compiled propensities reads DIFF;
• a model with no propensity library runs no replicate;
• a library the engine declines takes one replicate each way;
• the runner reads rr_ssa's jobs as SSA jobs.

The compiled-side tests skip on a machine with no C compiler, which the skip audit declares.

Mutations, each caught (rerun on the final commit):

• a differing seed read as a pass;
• the mass-action rule off;
• the ensemble verdict always passing;
• codegen=False not passed to the interpreted side;
• codegen=False reporting compiled propensities left unread;
• the no-library skip off;
• the one-replicate probe off.

Suites: parity_checks/tests and the shapes give 1,104 passed, 28 skipped, 1 xfailed. One local failure, test_jacobian_characterization_gate, needs main's fill_dense_fd_jacobian_half_step (#1002), which my local build predates; CI builds fresh. ruff 0.14.0 is clean. Merged with lanl main.

Still to do

The baseline (parity_checks/nightly/baselines/differential.json) gains these 536 rows from this PR's own nightly run. I'll record them from its nightly-differential artifact.