test(rr_parity): MODEL2002070001's invalid-reference premise keys on the reference, not on bngsim (#485)

[wshlavacek]

Closes #485.

What the sweep was saying

INVALID_REFERENCE["MODEL2002070001:ode"] is stale, and the 0.15.0 sweep says so:

STALE invalid-reference entry: natural=DIFF, rr_finite=False

Half the premise ("bngsim failed and the reference is unusable") died in 0.14.0,
when the #353 unit-volume substitution for a non-finite compartment size let bngsim
integrate the model. So the natural verdict resurfaced as a DIFF with
max_rel_err = inf, 606/707 cells failing — RoadRunner's NaN scored against bngsim
as a divergence, which is the one thing it is not. The other half is untouched:
RoadRunner still returns non-finite output.

Evidence, re-verified here

The model's NaN is not in its dynamics. Both compartments declare size="NaN", but
every species is hasOnlySubstanceUnits="true", the dynamics are 7 rate rules with 0
reactions, and no <math> block in the file names mucosa or lumen — so nothing
reads a size and the IVP is well posed in amounts.

At the sweep tolerance (rtol 1e-9 / atol 1e-12, t = 0..100, 101 points), against
bngsim 0.15.0 / RoadRunner 2.9.2 / AMICI 1.0.1:

engine	result
bngsim	runs; 707/707 finite cells
RoadRunner	runs without raising; 101/707 finite — x1..x6 NaN from t = 0
AMICI	0/707 cells failing against bngsim, max_rel_err = 0.0

RoadRunner's one finite column is x7, the single rate-rule variable that is a
parameter rather than a species in a NaN-sized compartment — and it agrees with
bngsim cell for cell. That is a clean control: the NaN reaches RoadRunner's state
through the compartment, not through the dynamics.

The change

_apply_invalid_reference now keys the premise on the reference alone — RR ran and
emitted non-finite output — instead of also requiring that bngsim failed.

Two deliberate departures from the issue's rr_finite is False or rr_finite is None:

• None is not admitted. A segfaulted child records the same absence as an RR
  raise, and test_invalid_reference_does_not_fabricate_bad_test_from_a_dead_worker
  exists for that; REFERENCE_FAILED is already the right bucket for "RR raised while
  bngsim ran". Nothing is lost — the both-raised case is still covered by the natural
  BAD_TEST clause.
• A natural PASS is not admitted. rr_finite is whole-array, so a reference
  non-finite only in a column the comparison never used was in fact usable for the
  comparison; flipping that to BAD_TEST would throw away a real pass.

What the dropped "bngsim failed too" clause was really doing was guarding against
burying a bngsim defect, so it is replaced, not dropped. _reference_nonfinite_covers
re-runs the same _core.differ verdict over just the columns the reference kept finite,
and the override holds on a DIFF only if that passes. A real bngsim divergence beside
the NaN leaves a failing cell in a finite reference column and the row stays a DIFF,
flagged stale. Restricting the columns can only tighten the re-run (scale is the peak
over the retained block), so the claim it licenses is never a loosening.

BAD_TEST rather than PASS, as the issue asks: PASS is the #482/#483 treatment for
a divergence attributed to the reference engine, and here it would score RoadRunner as
having validated a model it answered NaN on. Both buckets are non-scoring either way.

One note worth flagging

amici_parity makes the opposite call for a superficially similar shape — its
INVALID_REFERENCE lands on REFERENCE_FAILED, on the stated grounds that BAD_TEST
means neither engine could run the model. That README sentence cited this override's
"bngsim had also failed" clause as the difference, so it is updated here: what separates
them is whose defect it is. There, AMICI returns finite numbers its own defect makes
wrong. Here, the model's own SBML put the NaN in the reference's state — the model is
the bad test. Both buckets are non-scoring, so nothing hangs on it numerically, but if
you would rather this row read REFERENCE_FAILED for consistency with that suite, it is
a small follow-up (it would want an invalid_result refusal class the way AMICI has one).

Verification

• Full ODE sweep on this branch: 1237 PASS / 27 REFERENCE_FAILED / 54 BAD_TEST /
  5 TIMEOUT over 1323 jobs — no DIFF, no EXCEPTION, and no override flagged stale.
  This was the last DIFF after test(rr_parity): BIOMD0000000627's divergence is RoadRunner stepping over an ungated ramp (#482) #483.
• parity_checks/tests: 494 passed, 17 skipped (13 new).
• Suite-side only; no bngsim behaviour changes. build_ode_jobs.py cannot run in a
  checkout without $BIOMODELS_SEDML_DIR (and would churn _meta.generated), so
  ode_jobs.json is re-baked for the one record through the same Override.to_dict()
  • json separators write_manifest uses, verified to leave the other 1322 records and
    _meta byte-identical.