Expose per-transition compiled propensity + stoichiometry artifact on CompiledRhs

[jc-macdonald]

Closes #184.

Summary

Exposes a new CompiledRhs.reactions: tuple[CompiledReaction, ...] artifact for kind: transitions specs — one compiled propensity evaluator + axis bookkeeping per named transition, built entirely independently of the existing deterministic per-state equation summation. This is the compile-side building block a hybrid deterministic/CTMC engine needs (tracked separately, downstream, against flepimop2's diffrax_engine.py); this PR does not add any execution semantics, sampling, or opinion about what a consumer does with the artifact.

Why a separate pass, not an extension of _build_transition_equations_ir

Traced the existing per-state equation construction in detail before writing anything: _build_transition_equations_ir deliberately shares Expr objects by identity across template cells (the tpl_uniform fast path) and deduplicates summed terms by id()-tuple (documented against issue #145's performance work). A shared object can legitimately serve multiple transitions/cells at once, so there's no way to recover per-transition identity by decomposing the already-summed equations_ir — and tagging provenance onto shared IR nodes would either be wrong (for legitimately-shared nodes) or require unwinding that optimization. This PR adds a new, independent pass instead (op_system/_reactions.py) that reuses the same normalization primitives (expand_inline_templates, expand_reduce_pointwise) but never touches _build_transition_equations_ir itself — confirmed zero risk to the existing hot path by running the full pre-existing test suite unchanged after each commit.

What's in scope (v1)

A transition is included in the artifact iff:

• it has a name: (unnamed transitions aren't addressable — no join key)
• it has a from: (no from: null source-only transitions — no well-defined "how many independent source cells are firing")
• its to-side wildcard axes are a subset of its from-side wildcard axes (no axis "created" from -> to)
• its rate expression doesn't reference an axis outside that from-side wildcard set

Transitions outside this scope are silently omitted from reactions/reactions_ir, not an error — mirrors how history_requirements is built opportunistically elsewhere in this package. This covers both the same-shape case (S[age,vax] -> E[age,vax]) and the "collapse to a pinned target coordinate" case (C[age,vax] -> S[age,vax=f], i.e. every firing cell's event lands in a single fixed stratum regardless of its own axis value) — the latter is exercised heavily by the diphtheria model this was built for (recovery/treatment-success transitions that all land in a vax=full-equivalent stratum).

What CompiledReaction carries

name: str
from_base: str            # state base name depleted
from_axes: tuple[str,...] # shape of the compiled propensity
to_base: str               # state base name replenished
to_axes: tuple[str,...]    # subset of from_axes
sum_axes: tuple[str,...]   # from_axes not in to_axes -- summed away when scattering
pinned: tuple[tuple[str,int],...]  # (axis, to-side coord index) for each sum_axis
propensity_fn: (t, y, **params) -> array   # shaped like from_axes; rate * from_state

Important correctness note, found by testing rather than by inspection: the propensity is rate * from_state (the actual per-cell hazard), not the bare rate: expression — caught this exact bug by hand-computing an expected value before wiring the compile step in, where an early version returned just the bare rate. The strongest verification in this PR is a cross-check that didn't exist as a plan item originally: summing each reaction's propensity_fn output, scattered via its own sum_axes/pinned bookkeeping, exactly reconstructs the existing deterministic pytree_eval_fn's output for the same spec/state/params — done for a plain SIR-shaped spec, the collapse-to-pinned-target case, and a factorize_axes + shaped-param-rate (delta[vax] * tau) case matching the diphtheria model's real shape.

Where it hooks in

• src/op_system/_reactions.py (new): ReactionArtifactIR + build_reaction_artifacts_ir, called once from normalize_transitions_rhs, stored on a new TransitionsRhs.reactions_ir field (default empty tuple — ExprRhs/existing callers unaffected).
• src/op_system/compile.py: CompiledReaction dataclass, ReactionPropensityFn protocol, _build_reaction_artifacts (mirrors _build_history_artifacts's exact pattern — dedicated helper, compare=False/hash=False/repr=False field, restored in __setstate__), wired into compile_rhs. Reuses _vectorize.py's _compile_ir_expr/_LoweringContext the same way compile.py already reaches _vectorize.py internals elsewhere (lazy importlib.import_module, avoiding the existing import-order constraint between the two modules) and reuses the already-built _VectorPlan's param_templates/extra_param_buffers for shaped-param buffer assembly, so propensity evaluation stays consistent with how make_pytree_eval_fn resolves the same parameters.

Answers to the open questions from #184

• Naming: went with reactions/CompiledReaction (not transitions, to avoid colliding with the existing meta["transitions"] key).
• Unnamed transitions: silently excluded (see scope above).
• Propensity evaluator calling convention: matches PytreeEvalFn's (t, y, **params) exactly (new ReactionPropensityFn protocol), for consistency — it's read-only (never returns a StateDict, just an array), but keeping the same call shape means a consumer can reuse the same param-passing code path for both.

Verification

• Full existing test suite: 442 passed, unchanged, after every commit.
• New tests/op_system/test_op_system_reactions.py (9 cases): exclusion rules, axis bookkeeping + propensity correctness for both same-axes and collapse-to-pinned-target, the deterministic-reconstruction cross-check, factorize_axes + shaped-param-rate, and ExprRhs compiles with empty reactions.
• mypy --strict clean on every touched file, including the new test file.
• ruff check/ruff format clean, modulo pre-existing repo-wide findings from a newer local ruff version's noqa-comments style preference (# noqa: vs # ruff: ignore[...]) that fires identically on untouched pre-existing lines — confirmed not introduced by this PR, not touched.

Test plan

• pytest — 450 passed (442 existing + 8 from a WIP commit, replaced by 9 final)
• mypy --strict — clean
• ruff check / ruff format --check — clean (see note above on pre-existing noise)
• Manual cross-check against the deterministic path for 3 spec shapes, including one matching the diphtheria model's real [age, vax, loc] + factorize_axes structure
• Downstream: flepimop2's diffrax_engine.py CTMC/tau-leaping PR (tracked separately, blocked on this)