mirth's own checks over the corpus: --compiler-checks, ui-incr (P6 over the UI suite); finding 56

[zmaril]

Step 1 of leaning on mirth more: the patched compiler's own checks (verify-reuse.patch and report-untracked.patch in rustc-verify12) had only run in the fuzzer, the replays and the flag walks. This PR turns them on over the corpus.

What's added

• --compiler-checks for the sweep driver; used by diag-check and gate-mutate.
  • Each compile becomes an incremental session with RUSTC_VERIFY_REUSE=all and RUSTC_REPORT_UNTRACKED. Both checks hook the dependency graph, so they do nothing without -Cincremental.
  • Reports are classified against rustc/untracked-known.tsv: every (what, file) pair the fuzzer, replays and flag walks reported, plus options whose verdict holds anywhere. Reuse reports are classified against rustc/reuse-known.txt.
  • New library module compiler_checks.rs, plus Compile::compiler_checks.
• mirth-lab ui-incr: P6 over the UI corpus.
  • For each test: a clean incremental session, then three rebuilds in it (unchanged; every span moved by a leading blank line; an unused fn appended).
  • Each rebuild is compared with a clean build of the same source on diagnostics, output bytes, and program output.
  • The reuse check and the untracked-read report are on in every session.
• big-run.sh runs diag-check-cc and ui-incr.

Results

Sweep	Scope	Result
ui-incr	18,502 tests × 3 rebuilds	Finding 56; otherwise no incremental/clean difference in bytes, program output or diagnostics
diag-check --compiler-checks	18,374 tests, every cached value recomputed	No difference; no new untracked read
gate-mutate --compiler-checks	3,000 mutants	Nothing reported

Finding 56. With -Zunleash-the-miri-inside-of-you, the flag's warning and its feature-gate error come from session state that const checking fills as a side effect. An incremental rebuild therefore drops both, and accepts a crate that a clean build rejects. Low severity: the flag exists for testing the const evaluator. The lost warning goes back to at least 1.80.0; the error case was shown on nightly-2026-10-06. Facts are in docs/hunt/unleash-incremental.md, with a repro in docs/hunt/tests/.

Triaged, not findings (details in checks.md, fourth batch):

• Allocation identity in const-eval results across typing modes (222 tests). This is finding 8's mechanism. No output effect was observed.
• The reuse check re-emitting lints when it recomputes values (4 tests). This is an artifact of our patch; it's noted in shadow-mode.md.
• New source-text read sites. 46 sites in 15 queries, all in the known class.
• Two harness artifacts, fixed:
  • the flag order with a trailing --cap-lints;
  • the per-session object-name suffix, which had made 78 tests look nondeterministic.

release-diff is not wired in, because it builds with official toolchains that don't have the patches.

Note: this branch numbers its finding 56. If another branch merges first, it may need renumbering.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QXiEXbESemwqMLYKaWLDbT

[zmaril]

Merged into main with #41–#46; renumbered at merge: finding 56 is now 58 (docs/hunt.md).