Restore cargo check --all-targets: drop the test importing a deleted module

[briansrls]

cargo check -p v1-compiler --all-targets has not compiled since #8486, which deleted the .dag witness v1_tests_claim_caret_parse_smoke_test and its generated module but left caret_parse_smoke_native_compile_emit_witnesses importing it:

error[E0432]: unresolved import `crate::v1_tests_claim_caret_parse_smoke_test`

Only --bin was still compiling, so -D warnings covered one target while dead code accumulated unobserved in the other — #8517 removes 18 such functions found this way.

The part worth reviewing

The orphaned test had three homes, not two:

1. src/v1/compiler_tests_rust.dag — the authority
2. src/v1/stage0/src/v1_compiler_compiler_tests_rust.rs — its generated projection
3. src/v1/stage0/src/compiler_tests.rs — output of running (2)

compiler_tests.rs is produced by running the generator, not by emitting from the .dag. So editing (1) and (3) alone leaves (2) stale, with the authority silently disagreeing with its projection — and --all-targets still goes green in that state. My first attempt did exactly that; I caught it only because regenerated output still contained a test I thought I had deleted.

Verified by regeneration rather than by a green check: compiler_tests.rs is byte-identical to generator output (diff exit 0).

Coverage, stated exactly

dag/test/claim/caret_syntax_witness_test.dag covers the 3 endpoint claims (compile-clean, symbol-literal emit, discriminant-sugar emit). The 7 tokenize/parse intermediate claims have no .dag equivalent and have not executed since #8486 deleted the module; the Rust suite left CI in July. This deletes unbuildable code, not live coverage — but those 7 are a real gap, not parity.

Two findings NOT fixed here

Found while verifying; both need their own disposition:

1. --required-regen refuses unconditionally on any tree. generated_basenames_from_emit inserts the full emit key (src/foo.rs) while committed_generated_basenames inserts a basename (foo.rs), so the sets can never intersect. It refuses before writing a candidate or comparing content.
2. Real drift behind that bug. With the prefix noise patched locally, 127 files clear and a genuine residue remains: 9 committed-but-never-emitted (required_regen_host.rs, v2_compiler_compile.rs, expected_red_roster_join.rs, +6) and 1 emitted-but-not-committed (v1_compiler_trait_bound_witness.rs).

DESIGN.md names --required-regen-fixed-point as the sole answer to the self-host fixed point, and that flag only reads a receipt this producer writes — so the authority is currently answering nothing, and (1) has been concealing (2). No fix for either is included here; (2) needs a per-file disposition rather than a patch.

🤖 Generated with Claude Code

[gunbai-bot]

Thanks — approving read matches the change. One correction to the record, because it inverts the fact this PR is mostly about.

The review describes the three files as "the .dag source, the generated Rust, and the hand-authored twin." There is no hand-authored twin: all three are generated, in a chain.

src/v1/compiler_tests_rust.dag                       authority
  -> v1_compiler_compiler_tests_rust.rs              generated projection (committed, compiled in)
       -> compiler_tests.rs                          output of RUNNING the projection

compiler_tests.rs is not hand-maintained and is not emitted from the .dag either — it is produced by running the compiled-in generator. That coupling is the whole trap: edit the authority and the output, leave the middle file stale, and the authority silently disagrees with its projection while cargo check --all-targets still passes. That is exactly what my first attempt did, and a green check did not catch it — regenerating and diffing did (byte-identical, diff exit 0).

Flagging it because "hand-authored twin" would license editing compiler_tests.rs directly next time, which is the one thing that reintroduces this class silently.

— sent from fierce-ram-721