Add params support to Contractor and Separator

[dpsanders]

Summary

Wire ReversePropagation's parameter support through Contractor and Separator so that the same compiled separator can be re-used at runtime with different parameter values, instead of recompiling one separator per parameter value.

ReversePropagation.forward_backward_contractor(ssa, vars, params) already emits a contractor with signature (__inputs, __constraint, __params) -> .... Until now, ICP's Contractor / Separator exposed no way to use it.

Motivation

Periodic Lorentz gas "first-hit" probabilities, computed via interval-validated paving: each blocker disc contributes a not_blocked separator with the disc centre baked in as a numeric literal. With ~127 blockers along a corridor of length 100, that's ~500 atomic separators each going through Symbolics → ReversePropagation → @RuntimeGeneratedFunction. Build cost is ~10 minutes and grows linearly in k.

Rebuilding not_blocked parametrically in (Bx, By) and reusing it across blockers measured ~16 s build at k = 20 (vs ~190 s for the baked-in version), and stays constant in k.

Changes

• Contractor and Separator gain a params field (default []) and a params kwarg on every constructor.
• For the exact = true codepath in build_fb_contractor, emit the parametric outer wrapper symmetrically (this codepath rebuilds the wrapper itself for decoration / exact-literal handling, so it had to be taught about params explicitly).
• (CC::Contractor)(X, constraint, param_vals::Tuple) and (SS::Separator)(X, param_vals::Tuple) are new call methods.
• Atomic separators built without params silently ignore an incoming param_vals, so parametric and non-parametric separators compose cleanly through ⊓ / ⊔ / ¬.
• CombinationSeparator gains a second closure fp for the parametric call; ⊓ / ⊔ / ¬ build both closures so the same composite can be invoked either way.
• make_function and the separator(...) factory thread params through to Symbolics.build_function and to leaf Separator construction respectively.

Compatibility

Existing 1-arg / 2-arg call signatures are unchanged. New behaviour only kicks in when params=[...] is passed at construction or when param_vals::Tuple is passed at call time.

Tests

Three new testsets (Parametric Contractor, Parametric Separator, mixed composition):

@variables x y r
C = Contractor(x^2 + y^2, [x, y]; params = [r])
C(X, -Inf..1, (interval(1.0),))  # x, y ∈ [-1, 1]
C(X, -Inf..4, (interval(2.0),))  # same compiled C, x, y ∈ [-2, 2]

@variables x y cx cy
S = Separator((x - cx)^2 + (y - cy)^2 <= 1, [x, y]; params = [cx, cy])
S(X, (interval(0.0), interval(0.0)))   # disc at origin
S(X, (interval(5.0), interval(0.0)))   # same compiled S, disc at (5, 0)

Sx = Separator(x >= 0, [x, y])         # no params
Scomb = S ⊓ Sx                          # mixed composition
Scomb(X, (interval(0.0), interval(0.0)))  # works; non-param leaf ignores params

Test plan

• Parametric Contractor testset passes (same compiled C reused at two different r).
• Parametric Separator testset passes (same compiled S reused at two different (cx, cy)).
• Mixed parametric ⊓ non-parametric composition passes.
• End-to-end smoke in downstream Lorentz gas project: parametric not_blocked separator built once, applied across multiple blocker disc centres, returns sensible contractions per blocker.
• Existing Contractors / Separators testsets — currently fail on this branch with add_rev not defined in IntervalArithmetic, but this is a pre-existing test-environment issue (the test env doesn't dev the local IntervalArithmetic that has the new reverse functions); not introduced by this PR.

Note on commits

This branch is stacked on lorentz-no-icontractors-style preparatory work (RP 0.8 / SSAFunction adoption, exact = true decoration handling, IntervalContractors removal). Those changes weren't on master at PR time, so they appear in the diff as separate commits — happy to rebase / split / target a different base branch if any of them have landed in the meantime.

🤖 Generated with Claude Code