Closes #3485 `fn test_booth_encode_zero` and `test booth_encode_zero` are DIFFERENT NAMESPACES IN t27 and the SAME IDENTIFIER IN C, where a test block becomes `void test_{name}(void)`. The duplicate gate compares raw names, so it was silent while the corpus printed `redefinition of 'test_booth_encode_zero'`. MEASURED PER BACKEND, NOT ASSUMED, and the message names only the backends it is true of: gen-c `void test_X(void)` / `void bench_X(void)` COLLIDES on both gen (zig) `test "X"` is a STRING, not an identifier no collision `fn bench_X()` COLLIDES gen-rust tests are not lowered at all no collision gen-verilog a test is emitted as a comment no collision Repeating the existing "every backend rejects a redeclaration" would have been false for three of the four. POPULATION: functions named `test_*` 66 ... with a test block of the matching name 4 in 2 specs functions named `bench_*` 0 test blocks named `test_*` 532 ... whose stripped name is also a test 0 The 4 split, and neither half is repaired here. The two in igla/race/backend.t27 are NEVER CALLED -- each appears once in the whole tree, as its own declaration -- while the two in math/property_test_template.t27 ARE called from test blocks. One pair is dead code that also collides; the other is a live helper that needs one of two names changed. Both are a spec author's decision. THE `bench` HALF HAS POPULATION ZERO AND IS KEPT ANYWAY, written, tested and shown reachable rather than left out -- the same reasoning already recorded in this function for the type/function finding: a zero nothing could have produced is not evidence. THE RULE KEYS ON `fn`, AND 532 TEST BLOCKS NAMED `test_*` ARE WHY THAT MATTERS. `test "test_thing"` beside `test "thing"` gives `void test_test_thing(void)` and `void test_thing(void)` -- two identifiers, no collision. A mutant letting any declaration kind trigger the rule survived every test until a fixture for that case existed. It gets ITS OWN RATCHET rather than a number shared with the existing one: they are different claims, and a shared count lets one kind be paid for with the other. The Spec Guards `paths:` filter is extended to the new baseline, which otherwise would not have triggered the workflow that reads it. The corpus is unchanged -- 11 642 before and after, all 582 generated headers identical. This is a `check` change. Tests: 4 new Rust cases, 3 new self-check cases (8 of 8 pass), full suite 3537 passed 0 failed. Five mutants, all killed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Commits

Commits on Sep 8, 2026

check: a collision that only lowering creates