[Montana]

Fixes #1312.

Stacked on #1322 (#1310), which adds the method, n_rep and seed fields of Job and checks_for that this arm uses. Until #1322 merges, this PR also shows its commits. This PR's own are 01e46ff and 0e273d4. It adds one row to CHECKS (nf_reject_vs_exact), as the other differential PRs each add theirs. I'll resolve it.

What it adds

bngsim's two network-free engines, NFsim (nf_reject) and RuleMonkey (nf_exact), against each other, in distribution.

The models. An NF job is one of bng_nf's models, read from its pinned reference archive (reference_archive.py fetch bng_nf). Each entry holds the BNG-XML BNG2.pl wrote for the model and the run it was recorded for (t_start, t_end, n_steps, gml, nf_bscb, n_rep, seed_base). So the arm runs the NF track's own models and windows, and the job needs no BNG2.pl. Both engines are bngsim's, so the archive is inputs only, whatever the nightly's mode. A multi-segment entry (1 of 235) is left out, since one segment is not its run.

The check. nf_reject_vs_exact runs n_rep replicates on each engine, with the same seeds, block_same_complex_binding and molecule_limit. It judges the observables' ensembles with differ.ensemble_verdict at ensemble_k_for(n_rep), the SSA suites' verdict:

• a DIFF is escalated the way bng_parity's NF track escalates one, by its own _run_escalation: 30, then 100 replicates each, stopping where agreement stalls. Finite-sample scatter clears, and a real divergence stays. Escalation runs within NF_ESCALATION_BUDGET (300 s), and a rung that would not fit is not walked; the row says so;
• the note names the observable furthest apart, on its own peak;
• NFsim failing is no_base (the NF track owns it);
• RuleMonkey refusing by name is UNSUPPORTED: compartments, or "cannot honor this BNG XML";
• the engines reporting different observables is DIFF.

Its own nightly job. "Network-free engine agreement" runs under its own baseline, differential_nf. Over the 234 models the two engines' ensembles cost 9,793 CPU-s here, 92% of it in 48 models, which the path-agreement job, already at 52 of its 180 minutes, could not also hold. The path-agreement job is unchanged.

+  differential-nf:
+    name: Network-free engine agreement
+    ...
+      - name: Fetch bng_nf's models
+        run: python3 parity_checks/nightly/reference_archive.py fetch bng_nf --dest "$RUNNER_TEMP/bng_nf"
+      - name: Run differential_run.py (NF arm)
+        run: |
+          uv run --no-sync python parity_checks/differential/differential_run.py \
+            --fd-sens-jobs '' --rr-jobs '' --ssa-jobs '' --no-locale \
+            --nf-archive "$RUNNER_TEMP/bng_nf" --workers 4 --timeout 900 \
+            --out "$RUNNER_TEMP/differential_nf.json" || true

Over the corpus

bng_nf's 234 single-segment models, local run (--workers 4 --timeout 900):

PASS	DIFF	UNSUPPORTED	TIMEOUT
227	5	1	1

The run was made before the refusal fix, when ft_total_rate read EXCEPTION. I reran it on this head: UNSUPPORTED.

The 5 DIFFs:

model	at 10 replicates	escalation	reading
original/.../ode/error.bngl	65% of cells	30 each: 67%, stalled	a real divergence: max z 1137, Initiator_I2_SSA
glacial/.../corpus/fceri_ji.bngl	95.8%	over budget (205 s for 10)	unresolved
original/.../nf/fcr.bngl	98.7%	over budget	unresolved
original/.../ode/population_scaling.bngl	98.8%	over budget	unresolved
slow/.../corpus/tcr.bngl	98.5%	over budget	unresolved

The four unresolved ones sit just under the suites' 99% bar on 10 replicates, so they are likely sampling scatter. Their base ensembles take 200–300 s, too slow to escalate in the budget. With fixed seeds they are reproducible, so the baseline holds them steadily rather than flaking. error.bngl is worth an issue: the two engines disagree on Initiator_I2_SSA well past noise. I haven't yet attributed it to either engine.

TIMEOUT: AMPK_MTORC1_ULK1_v2.bngl, past 900 s.

Elsewhere simple_model_1 read DIFF at 10 replicates (96%) and cleared at 30 (100%): the escalation doing its job.

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

On tests/data/nfsim/simple_system.xml:

• an NF job runs the NF check and no other;
• NFsim and RuleMonkey agree in distribution;
• negative control, as the issue asks: one rate (kcat) ×3 on RuleMonkey only reads DIFF. Escalation walks to 30 replicates and stalls (55%, 55%), with seeds 1–30, new draws, not the base's again;
• an escalation past its budget is not walked, and the row says so;
• engines reporting different observables is DIFF;
• NFsim failing is no_base;
• RuleMonkey refusing (compartments; TotalRate) is UNSUPPORTED;
• the runner reads bng_nf's archive as NF jobs, leaving a multi-segment entry out.

The tests skip where the tree lacks the fixture or the build lacks RuleMonkey.

Mutations, each caught:

• always pass;
• NFsim both ways;
• no escalation;
• the budget ignored;
• seeds reused;
• the refusal not read;
• the observable names unchecked;
• NFsim's failure not read as no_base.

Suites: parity_checks/tests and the shapes give 1,108 passed, 33 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.

Still to do

The differential_nf baseline (parity_checks/nightly/baselines/differential_nf.json) comes from this PR's own nightly run, from its nightly-differential_nf artifact. That run also measures the job's cost on a CI runner, which I have only estimated.