wave 703: one suspended check hid five defect classes — 10 of 28 Coq proofs became 25 #4881

Re-filed from #2138. An attempt claimed that issue on 2026-09-24 and left an empty branch, and a dispatch row that has spent its retry ceiling keeps its issue for good - so no bee could take it again, however much the brief improved. The work below is carried over unchanged.

Prop. 201 found a continue-on-error: true step reading "non-blocking until observed green" that had never been green. Continuing behind it until the class boundary:

.vo files: 10 → 25 of 28. One repair unblocked seven dependents, because a build stops at its first error and everything downstream is reported by nothing.

Five classes, all mechanical, none related. They accumulated because the flag made the exit status invisible — nothing distinguishes one defect from five when the observable is "workflow passed".

Stopped at the class boundary. It now fails on Physics/StochSkipSafe.v with Tactic failure: Cannot find witness — a proof obligation, not bookkeeping. The flag is not flipped.

My own scanner reproduced the campaign's oldest mistake. The class-4 scan skipped any line already containing %nat. The defect was on a line with two conjuncts annotated and one not. An exclusion whose granularity is coarser than the defect's cannot see a partial fix — and partial fixes are the most likely state of any file someone has already touched.

Theorem (suspension bookkeeping). A suspension (step, flag, promise) is sound only if the promise is checkable. Running the suspended check every invocation is the direct test and is often prohibitive. The cheap sound alternative: make the set of suspensions ratcheted — a new one fails the build until recorded, a removed one shows as a baseline entry with no step. That doesn't prove any suspended check still fails; it guarantees no suspension is ever undocumented, which is the property that was missing.

Gate 29 does that, and deliberately does not classify a step as check vs infrastructure — YAML can't distinguish them and guessing would be classification-by-name (Prop. 197).

Boundary

• trios-coq/Physics/StochSkipSafe.v

The instruments

Every one of these reads and prints; none of them writes. Run them from the repository root, with <spec> replaced by the file in ## Boundary:

.t27 is not Rust and not Zig, and a bee fluent in either writes one by
accident. Measured 2026-09-20: a bee filled eight bodies in
specs/file/watcher.t27 with return Ok(());, Err(FileError::WatcherNotFound)
and for i in 0..watchers.length. The parse ratchet refused the whole file --
Unexpected token in expression: RParen, line 123 -- and a spec that does not
parse generates nothing, so every test it already carried stopped running.

At the top level the parser accepts exactly these eight forms, with an optional
pub, and no others:

const   var   fn   enum   struct   test   invariant   bench

There is no trait, no impl, no type X = ..., no generics, no macro. What
a bee reaches for from another language, and what happens:

• Rust's result sugar (Ok(()), Err(E::V), ?) is not a construct here.
• Rust macros (println!, format!, vec!) are not constructs here.
• A range loop (for i in 0..n) is Rust and Zig, not this language.
• Zig builtins (catch unreachable, @intCast, @constCast) are not constructs here.
• Cross-module reuse (use other::fn;) parses and then generates a comment and an unqualified call, so the Zig fails with use of undeclared identifier (use a::b; generates no import, so no spec can reuse a function from another spec #4298). It does not work yet.

t27c parse <spec> answers in one line whether what you wrote is the language.
Run it before you report, every time.

• t27c spec-status <spec> - the compiler's one-word verdict: IMPLEMENTED, PARTIAL, UNWRITTEN, NOPARSE, NOFN. Exit code is 0 whatever it says, so read the word.
• t27c symbols <spec> - every name the file declares, with its kind. Answers "does this already exist here?" before you add it.
• t27c outline <spec> - per function: its locals, what it calls, what it returns. The contract you are implementing against.
• t27c coverage <spec> - which functions have a test and which do not. The issue asks for tests; this is how you check you wrote them.
• t27c lint <spec> - style and shape warnings. It OVER-REPORTS has no test or invariant: measured 2026-09-20 over 60 specs it printed 677 of those where coverage found 190 untested functions, disagreeing on 59 of the 60 - it warns about bit_to_trit_pair in specs/base/ternary_encoding.t27, which test bit_to_trit_pair_zero calls on line 249. Read it as a hint; coverage is the answer.
• t27c typecheck <spec> - types, before generation. Prints Typecheck OK (0 errors, 0 warnings) or the errors.
• t27c test-report <spec> - builds this spec and runs its own tests, which is what the oracle does. BLOCKED means the generated Zig does not compile, with the error beside it.
• python3 tools/dupe_scan.py --name <function> - where that function already lives, if it does. 576 of 4021 bodies here are byte-identical copies.

The rest of the toolbelt is in docs/BEE_TOOLBELT.md.