test(rr_parity): retire three override keys that could never match, and assert it (#482)

[wshlavacek]

Follow-up to #482 / #483, independent of both — branched off main and mergeable
on its own.

The three keys

stale_keys reports three ODE override keys naming models the manifest does not
build. SSA is clean.

key	dict	corpus status
MODEL0913003363:ode	TOL_OVERRIDES	drop,no_model
MODEL1006230083:ode	NO_ORACLE_ADJUDICATED	drop,no_model
MODEL6963432821:ode	NO_ORACLE_ADJUDICATED	drop,no_model

All three are dropped by the stage-1 corpus filter
(benchmarks/suites/biomodels/manifest.csv) for want of a file: the recorded
sha256 is e3b0c442...b855, the empty string's, so the fetch produced zero
bytes. They are 3 of 305 corpus entries in that state. build_ode_jobs.py
therefore builds no job for them and the keys cannot match. No commit ever added
them to ode_jobs.json, so they have been stale since the initial public
release.

I checked that before removing anything, because the other thing a stale key
can be is a typo'd or renamed model_id — where deleting the entry would quietly
un-mask a real divergence rather than tidy a dead one. These are not that: the
ids appear in the corpus manifest, explicitly dropped.

Retired, not deleted

Each becomes a comment carrying its disposition forward — the tolerance rationale
for MODEL0913003363, the GH #117 scipy adjudications for the other two — so
none of it is re-derived if the corpus ever fetches the model successfully. Same
treatment MODEL2205030001:ode got when GH #119 retired it.

Why three accumulated unnoticed

Nothing fails on them. The only detection is a print("WARN: ...") at the end of
build_ode_jobs.py / build_ssa_jobs.py — advisory, and unreachable in a
checkout without $BIOMODELS_SEDML_DIR.

So the cleanup comes with a guard: test_committed_overrides_have_no_stale_keys
asserts the committed overrides against the committed manifests for both regimes,
needing neither the mirror nor the build. I verified it fails on an injected
stale key rather than passing vacuously.

Scope

Suite-side only; no bngsim behaviour changes, and no job's disposition moves —
none of these keys was reaching a job to begin with.

parity_checks/tests: 481 passed, 17 skipped. ruff clean.