ci(mir): run the Michaelis-Menten sensitivity tests, and correct the Windows note

[wshlavacek]

Two follow-ups to #87 and #88, both confined to mir.yml.

test_codegen_mm_sens.py joins the run list

#87 added the other three analytic-sensitivity files. #55's Michaelis–Menten stage (#88) merged after it, so its 17 tests are currently covered only by the pre-push hook — the gap #87 existed to close, reopened by timing.

It needs no #85 quarantine, and the header now records why rather than leaving it to be rediscovered:

#85 fires on bngsim_codegen_output_sens, and generate_output_sens_from_model emits that only when the model has at least one function. The MM fixtures declare none, so the symbol is absent even with emit_output_sens forced on:

MM:         n_obs=3 n_func=0 -> want_outputs=True want_output_sens=False
MM_DERIVED: n_obs=2 n_func=0 -> want_outputs=True want_output_sens=False
   forced emit_output_sens=True -> symbol present: False

Same reason test_steady_state_dfdp_stops_differencing passes under the JIT today (your note on #85). So the Simulator(..., sensitivity_params=...) cases in TestEndToEnd should compile on this backend.

Corroboration that c2mir handles the arithmetic: mm_tqssa.net already goes through test_mir_equivalence and test_codegen_jacobian on this job, and the MM sensitivity block is the same closed form as the Jacobian block, emitted by the same builder (_mm_jacobian_groups). The ctypes/compile_rhs harness is copied from test_codegen_functional_dfdp.py, which passes 29/29 under the JIT.

I could not run a MIR build to confirm end to end. If a leg does go red on this file, the fix is one blocked_on_mir_jit marker on TestEndToEnd — the same shape #87 used — and I'd want the failure text first, since it would mean #85 is broader than bngsim_codegen_output_sens.

The Windows paragraph was wrong

It claimed the needs_cc cc-baseline tests skip on windows-latest, and that the compiler-free MIR-vs-ExprTk check was therefore the whole of the Windows verification. Neither holds:

• the runner ships MinGW gcc on PATH, so _CC = which("cc") or which("clang") or which("gcc") resolves;
• 0 skipped on every leg — feat(codegen): analytic sensitivity RHS for Michaelis-Menten rate laws (#55) #88's Windows run was 156 passed, current main is 229 passed, 10 xfailed;
• test_mir_equivalence.py::test_mir_matches_cc[...] runs and passes on Windows.

So all four legs assert cc-vs-MIR bit-identity on the same generated source. The comment understated the job's coverage, which is the direction that matters — someone reading it would think Windows was weakly covered and might not add a test there.

Replaced with what is true, plus the consequence worth keeping: because the runner has compilers, this job never exercises the no-compiler-at-all path that motivates MIR. Guarding that needs a leg with the compilers taken off PATH, not this one. Might be worth its own issue; I have not filed one.

Test plan

test_codegen_mm_sens.py, test_codegen_switch_condition_sens.py, test_codegen_functional_sens_rhs.py, test_codegen_functional_dfdp.py — 100 passed locally on the cc backend. The MIR legs of this PR are the real check, since the change is what they run.