test(syntax): pin SCFM equation content per-axiom, and per colour (#317)

[tsondru]

SCFM equation content was unpinned — closes #317.

Diff

git diff --stat main..HEAD (base 27be055): 3 files changed, 195 insertions(+), 2 deletions(-).

• catgraph-syntax/tests/frobenius.rs: +130/−1 (a scfm_goldens() table + each_scfm_equation_matches_its_own_golden).
• catgraph-syntax/tests/colored_frobenius.rs: +59/−1 (both_functors_agree_on_the_nine_axioms_at_each_colour, renamed from …_decide_…).
• catgraph-syntax/CHANGELOG.md: +6.

What changed

Confirmed by direct falsification before briefing: replacing scfm_equations' equation #2 (left unitality) with a reflexive/duplicate pair, keeping the returned Vec at length 9, left cargo nextest run -p catgraph-syntax at 131/131 green — every existing consuming assertion (laws.len()==9, mk(lhs)==mk(rhs), eq_mod==Some(true), the 18/10 hypergraph_presentation counts) is satisfied by any true parallel pair, not the specific named axiom.

Fix, per axiom (all nine): a literal d=2 MatKron image golden (shared across the five slots the algebra makes coincide — {2,5,9} all identity(2), {7,8} both μ;δ) plus the printed concrete syntax of both sides (which does individuate every slot — confirmed in review). Per the implementer's own falsification table (18 duplicate/reflexive substitutions across the nine axioms, in 5151976's dispatch report): a matrix-only golden, tried first, caught 4 of the 18; the syntax goldens caught all 18, which is why both are shipped.

Per-colour: to_mat_kron and CospanFunctor pinned to agree on all nine axioms at each of Hue::A/Hue::B (previously validated only at the trivial () colour), plus a [B, A] mixed-colour braid pin — the boundary to_mat_kron's braid had no prior test at (the pre-existing per_colour_dimensions covered only [A, B]).

Review

Six rounds via rust-v2:falsifying-reviewer (commits 5151976 → f32c4b1; round 4 landed mid-branch, after which the branch was rebased onto main 27be055, 3 unrelated commits, none touching catgraph-syntax — all SHAs below are the post-rebase ones on the pushed branch, git log --oneline main..HEAD):

• R1 (against 5151976, the implementation commit): 2 important (I1 — a test name claimed to_mat_kron "decides", contradicting the crate's own rustdoc calling it a sound-but-incomplete semantic check; I2 — the new colour-loop test never reddened alone across 27 perturbations) + 4 minor (module-doc overclaim/scope, a doc misattached to the wrong item, a missing per-axiom comment, an unmessaged assertion). I1/M1-4 fixed directly; I2 required real construction (a mixed-colour witness), dispatched to the domain agent — which added a δ-image assertion and a braid-order assertion, then honestly flagged the δ one as redundant under its own falsification (identical failure set with/without it). Cut the δ half, kept the braid half. → 30b12ec.
• R2 (delta on 30b12ec): 1 blocking (a comment claiming the braid assertion is caught "here and nowhere else in the suite" was refuted — per_colour_dimensions also catches the same perturbation) + 2 important/minor. Cut the false clause; tightened a count assertion to a literal 9 matching its message. → f0f6f53.
• R3 (delta on f0f6f53): 1 important (no CHANGELOG entry, against repo precedent) + 1 minor (the tightened count assertion lost an upper-bound: a bogus 10th golden entry would now pass silently, since zip drops it). Added the CHANGELOG bullet; restored the upper-bound check as a second assertion. → fff2be3.
• R4 (delta on fff2be3): 2 important, both in the new CHANGELOG bullet — it re-introduced the "own image" overclaim R1's I1 sibling had already scoped down in code, and credited the braid pin to CospanFunctor when only to_mat_kron is involved there. Rewrote the bullet; also cut an uncited rationale parenthetical. Branch rebased onto current main in the same pass (3 unrelated commits, confirmed non-overlapping). → b50da34.
• R5 (delta on b50da34, post-rebase): 1 important — the braid clause's "mixed colour boundary" phrasing credited syntax: scfm_equations content unpinned — deleting the left-unitality axiom leaves 136/136 green, zero warnings #317 with coverage that already existed pre-syntax: scfm_equations content unpinned — deleting the left-unitality axiom leaves 136/136 green, zero warnings #317 at [A,B]; what's actually new is the reversed [B,A] boundary. Rewrote to name the boundary explicitly. → f32c4b1.
• R6 (delta on f32c4b1): no blocking/important findings — the rewritten clause verified accurate (git grep 'braiding(3, 2)' at the pre-branch main SHA returns only the pre-existing [A,B] assert_ne!, none at [B,A]). One minor wording nuance noted by the reviewer as "not false as written" and left as-is — six rounds of substantive fixes is a reasonable stop point for a nuance the reviewer itself declined to press.

Across all rounds: the original defect (#317's headline) is caught by exactly one test, each_scfm_equation_matches_its_own_golden, in every reproduction, including a full-workspace --no-fail-fast run.

Gates

All run with --manifest-path <wt>/Cargo.toml --target-dir /tmp/cargo/catgraph-syntax against f32c4b1:

• cargo fmt --check — clean.
• cargo clippy -p catgraph-syntax --all-targets -- -D warnings — clean.
• RUSTDOCFLAGS="-D warnings" cargo doc -p catgraph-syntax --no-deps — clean.
• cargo nextest run -p catgraph-syntax — 133/133 passed.
• scripts/check_version_refs.py — all references agree on 0.16.0.

No production code changed — tests/frobenius.rs, tests/colored_frobenius.rs (new tests + one rename), and CHANGELOG.md.

Follow-up filed, out of #317's stated scope (not colour-specific): #384 — to_mat_kron's Tensor-arm factor order is unpinned anywhere in the workspace (swapping fm.kron(&gm) → gm.kron(&fm) leaves 2140/2140 tests green).

Closes #317.