refactor(math): drop Polynomial operator impls unused by production

[diegokingston]

Summary

Stacked on #598. Removes 6 Polynomial operator families from polynomial.rs that no production code reaches — proven by a clean cargo build --workspace.

Removed (319 lines, 24 impls):

• Sub Polynomial − Polynomial
• Mul Polynomial × FieldElement (and FieldElement × Polynomial)
• Add Polynomial ± FieldElement (and FieldElement ± Polynomial)
• Sub FieldElement − Polynomial

Of these, only Sub Poly−Poly had tests (2, self-referential); the other 5 families were fully dead with no test coverage at all.

Why these 4 families stay

Production uses operators at exactly 2 call sites, both building ∏(x − rᵢ):

• boundary.rs:149 — zerofier * binomial → Mul Poly×Poly
• transition.rs:127 — acc * (x_poly - current) → Mul Poly×Poly, Sub Poly−FieldElement

Sub<FieldElement> for Polynomial is implemented as -monomial + self, so Neg and Add Poly±Poly remain as transitive dependencies. Kept families: Mul Poly×Poly, Sub Poly−FieldElement, Neg, Add Poly±Poly.

Test plan

• cargo build --workspace — clean
• cargo clippy -p math --all-targets -- -D warnings — clean
• cargo test -p math — 196 passed (−2: the removed self-referential Sub Poly−Poly tests)

[github-actions]

Codex Code Review

Findings

1. Potential bug: PR removes Sub implementations for Polynomial - Polynomial, but the merged code still uses that operator in transition.rs:

   Polynomial::new_monomial(FieldElement::<F>::one(), 1) - current.clone()

   With the impl removed from polynomial.rs, this should fail to compile. Keep the Sub<Polynomial<...>> impls or update the call sites to use an explicit supported operation.

Notes

I attempted cargo check -p stark, but this sandbox cannot write to rustup’s home temp directory, so compilation could not run here. The source-level break is direct: the exact Sub<Polynomial> impl existed at base polynomial.rs:266 and is gone in this PR, while the operator use remains.

[claude]

Review: LGTM with one merge-order note

Pure dead-code removal - no logic changes, no correctness issues, no security concerns.

Verified:

• Both production call sites (boundary.rs zerofier * binomial and transition.rs acc * (x_poly - current)) use only the kept operator families (Mul Poly x Poly, Sub Poly-FieldElement).
• The Sub that's kept correctly delegates to Neg + Add Poly+Poly, which are also kept - dependency chain is intact.
• Removed tests were entirely self-referential to the removed impls.

Merge order: This PR is stacked on #598 (cleanup/math-dead-code), which is still open. Do not merge this until #598 lands on main and this branch is rebased, or the base ref will be wrong.