bootstrap/src/math_compare.rs is compiled by nothing; its plateau test contradicts its own goldens

[gHashTag]

Found while landing the cargo test -p t27c gate for #2288. Not fixed — filing it because the fix is a judgement call about which number is right, and because wiring the file in today would land that gate red.

bootstrap/src/math_compare.rs is compiled by nothing

There is no mod math_compare; anywhere in the crate:

$ grep -rn 'math_compare' bootstrap/src/
bootstrap/src/math_bayes.rs:168:  /// Path to verification data (default: .trinity/experience/math_compare.json)
bootstrap/src/math_bayes.rs:343:  let path = experience_path.unwrap_or_else(|| ".trinity/experience/math_compare.json".to_string());
bootstrap/src/math_bayes.rs:366:  if event.as_str() == Some("math_compare") {
bootstrap/src/math_compare.rs:185:  let path = dir.join("math_compare.jsonl");
bootstrap/src/math_compare.rs:212:  "event": "math_compare",
bootstrap/src/math_compare.rs:331:  .join(".trinity/experience/math_compare.jsonl")

Every hit is a string literal or the file itself. Rust reaches source only through mod declarations, so cargo build, cargo test, and cargo check --workspace --all-targets all skip it. The 10 #[test] functions in it have never been compiled.

It is one of 8 orphan .rs files in bootstrap/src — codegen_python.rs, math_bayes.rs, math_compare.rs, math_pslq.rs, notebook.rs, runtime_minimal.rs, runtime_minimal_test.rs, tooling.rs — and the only one carrying tests.

Built standalone it is 9 pass / 1 fail, and the failing test contradicts the file's own goldens

test_hybrid_v2_plateau (line ~419) asserts:

assert!((n152 - n20).abs() < 1e-9, "N=152 should match N=20");

But GOLDEN_V2 at line 162 of the same file records:

(20,  0.9617435184, 15.8995),
(152, 0.9617435163, 15.8995),

Those two constants differ by 2.100e-9. Recomputing hybrid_v2_cosine independently in IEEE-754 double reproduces the recorded goldens exactly:

N=20   0.961743518402275
N=152  0.961743516301789
|n152 - n20| = 2.100486e-09    tolerance 1e-9    -> FAIL (2.10x over)

So the file asserts a plateau tighter than its own recorded data supports.

Classification: the test is wrong, not the implementation. hybrid_v2_cosine matches all seven GOLDEN_V2 entries, and the sibling golden tests (test_hybrid_v2_golden_n20, ..._n152) pass at their 1e-6 tolerance. The convergence is real; only the 1e-9 plateau bound is unsupported. Pure f64 arithmetic — no fixtures, no network, no binaries.

Deliberately left alone. Per the rule that a wrong test gets reported rather than adjusted, I did not relax the tolerance and did not add mod math_compare;. Adding the mod today would make cargo test -p t27c red, and a gate that lands red gets disabled rather than obeyed (Prop. 26).

What a fix has to decide

1. Is the intended claim "N=152 equals N=20 to 1e-9"? Then the tolerance is right and the plateau claim is false at that precision — the honest change is to widen to 1e-8 (or assert against the goldens directly) and say so.
2. Or is 1e-9 a typo for 1e-6, matching every other tolerance in the module?

Either way the module should then be wired in with mod math_compare; so the answer is enforced instead of assumed.

Related, same shape

bootstrap/src/proxy.rs:326 gates its 8 tests behind #[cfg(all(test, feature = "server"))]. No workflow passes --features server or --all-features, so those 8 are stripped before type-checking everywhere. Together with math_compare's 10 they exactly account for the gap between the 913 #[test] attributes in bootstrap/src and the 895 that compile into cargo test -p t27c.

[gHashTag]

☣️ [VERDICT] Retracting this issue. Its premise is false on master.

Measured on feat/wave-547/host-heapsort — forked 2026-04-04, 2,399 commits behind master — not on master. That is the same wrong-tree error #2293 reverted 34 minutes after landing that morning, and this issue was filed at 03:54Z from the same checkout.

On origin/master:

• mod math_compare; is at bootstrap/src/main.rs:30, live since 767382a1b (PR chore(#969 follow-up): connect math_compare, fix codegen_python masquerade (Closes #1090) #1092, 2026-06-14), and dispatched at main.rs:1874, :10831, :11231. The module is not orphaned.
• The mod line was lost in a refactor (2d23e2a86, a 2380-insertion rewrite) — and was already found and restored by chore(#969 follow-up): connect math_compare, fix codegen_python masquerade (Closes #1090) #1092, whose body names this exact bug including the same 2.1e-9 figure.
• chore(#969 follow-up): connect math_compare, fix codegen_python masquerade (Closes #1090) #1092 also fixed the tolerance. math_compare.rs:433 on master reads < 1e-8, with the drift documented in a comment above it. The 1e-9 assertion exists only on the stale branch.
• Re-measured standalone against master's arithmetic: drift 2.100487e-9; passes at 1e-8, fails at 1e-9; all GOLDEN_V2 entries match to ≤3.1e-11.

The orphan set on master is 7, not 8, and none carry tests: _enrich/mod.rs, math_pslq.rs, tooling.rs, neural/attention.rs, neural/hslm.rs, runtime_minimal.rs, runtime_minimal_test.rs. math_bayes, notebook and codegen_python were all resolved in June by #1092 and #1100. The 913/895 test arithmetic is stale — master has 1622 #[test] attributes.

One finding survives, with corrections, and deserves its own issue: proxy.rs:234 (not :326) gates 2 tests (not 8) behind #[cfg(all(test, feature = "server"))], and no workflow passes --features server or --all-features.

#2288 is unaffected and still valid — formal-yosys.yml carries the cargo test -p t27c step but does not exist on master, so those tests compile and never execute in CI.

Two adjacent facts, not actioned: math_pslq.rs could not be wired in as-is (rug is absent from Cargo.toml), and bootstrap/src/memory/mod.rs:40 include!s the repo-root gen/, leaving the committed bootstrap/gen/rust/memory/*.rs copies unreferenced.

Closing as not planned — the defect it describes does not exist on the default branch.