test(applied,dl): add verify_rig_axioms call sites for Z and Dual (#292)

[tsondru]

Closes #292.

verify_rig_axioms (catgraph-applied/src/rig.rs:772) had sweeps for BoolRig, F64Rig, Tropical, UnitInterval, and Checked<i64>, but never ran on Z (BigInt, backing SNF) or Dual<T> (forward-mode AD) — both Rig carriers via the blanket impl<T> Rig for T where T: Clone+PartialEq+Zero+One+Add<Output=T>+Mul<Output=T>.

Change

• catgraph-applied/src/rig.rs: added verify_axioms_z_sample, sweeping Z::from(i64) samples (negative, zero, positive) through the existing nested-loop pattern.
• catgraph-dl/src/para/dual.rs (feature ad): added verify_axioms_dual_f64_sample, sweeping 11 Dual<f64> samples (new/variable/constant/zero/one, 6 with nonzero du) through the same pattern.

Review (4 rounds, rust-v2:falsifying-reviewer, perturb-in-copy)

• Round 1 (base fb8426b → ae8e74b, the two call sites): no blocking/important. Two minors: (a) an epistemic limitation, not a defect — for Z, no single-line perturbation the reviewer tried isolated "additive associativity" or "absorbing zero" as the first-firing axiom (earlier-ordered checks always fired first); recorded here as a disclosed limitation of the check-order + .unwrap()-on-first-failure mechanism, not something a test edit can fix. (b) the Dual<f64> sample set was thin on nonzero-du coverage (3/8) — a Mul-cross-term bug's detection rested on one data point. (b) applied in f2b5fc6 (8→11 samples, 3→6 nonzero-du).
• Round 2 (ae8e74b→f2b5fc6 delta): confirmed the widened set independently redundant (both the old-3 and new-3 nonzero-du subsets catch the perturbation on their own, with distinct witnesses). One important finding: the catch mechanism is narrower than "cross-term margin" framing implies — verify_rig_axioms aborts on the first failing axiom in a fixed order, so any Mul-formula bug that also breaks the multiplicative-identity axiom is caught there, never via associativity/distributivity between two nonzero-du operands. Documented in 903c275.
• Round 3 (903c275→640dc67 delta): the round-2 doc comment's "always caught via identity" was itself falsified — a Mul corruption that stays isomorphic to a valid Rig (e.g. folding an ε² contamination term into both channels, making Dual<f64> isomorphic to R×R) satisfies every rig axiom and passes undetected; verify_rig_axioms is a rig-axiom oracle, not a chain-rule oracle. Narrowed the claim and named the blind spot in 640dc67.
• Round 4 (640dc67 delta): no findings — the narrowed wording held up against the same class of falsification (3 independently-constructed isomorphic-to-a-valid-Rig corruptions, all correctly missed; 2 independently-constructed genuine rig-law breaks, both correctly caught). Converged.

Every perturbation round ran in a throwaway cp -r copy with its own --target-dir; the worktree stayed clean throughout (git status --porcelain empty before/after each round).

Gates (worktree)

• cargo nextest run -p catgraph-applied (--target-dir /tmp/cargo/catgraph-G5-T3): 669 passed, 0 failed, 12 skipped
• cargo test -p catgraph-dl --features ad (nextest skips doctests; Dual's doc example is a live doctest): 121 passed + 12 doctests passed, 8 ignored, 0 failed
• cargo clippy --all-targets -- -D warnings: clean on catgraph-applied (default, --no-default-features) and catgraph-dl (default, --features ad)
• RUSTDOCFLAGS="-D warnings" cargo doc -p catgraph-applied -p catgraph-dl --features ad --no-deps: clean
• cargo fmt --all --check: clean
• scripts/check_measured_claims.py / scripts/check_audit_counts.py: not applicable — diff carries no MEASURED markers and touches no audit doc

🤖 Generated with Claude Code