feat(task): per-artifact output equivalence predicate (Closes #246)

[noahgift]

Closes #246.

Byte-identity is exactly right for source and object files — recompiling to identical bytes correctly does not relink. It is wrong for producers that cannot reach byte-identity, and wrong in the dangerous direction: such an artifact is not "uncached", it is content-addressed with the wrong key.

The case, reproduced then fixed

An ASR pass writes narration.srt; a person corrects it. Verified end to end through the binary, differentially — "the artifact survived" proves nothing unless the default destroys it:

	2nd apply	narration.srt
no declaration	1 converged	ASR draft — human work destroyed
narration.srt: external	1 unchanged	Human corrected transcript

Both are tests, not just a transcript — the "destroyed" case is kept as the precondition that makes the fix meaningful.

--force still regenerates

The issue names force as the escape. I built a script-level guard that blocked it first, then removed it: it contradicted the spec, and it tripped the repo's own C3-FORCE-DISTINGUISHABLE invariant, which was right to complain. A guard that made --force mean something different here than everywhere else in forjar would be worse than the problem.

Three predicates, cheapest first, as the issue asks

• none — excluded from the staleness predicate; a missing artifact still is
• external — a modification is an improvement, not staleness
• command — compare a declared normaliser's stdout instead of the bytes, covering structural equivalence without teaching forjar any media formats

A failing normaliser is an error, never a silent fallback to bytes — that would reintroduce exactly the spurious staleness the author declared it to avoid.

Absent declarations mean bytes, so every existing project hashes exactly as before (FALSIFY-OE-002 pins it).

Two things the repo's own guards caught

1. every_resource_field_is_declarable_in_yaml failed — a field on Resource that isn't in RESOURCE_FIELDS makes any forjar.yaml using it fail with unknown field. The feature would have been unreachable from config while every unit test passed.
2. The declaration is now part of the hash. Flipping an artifact from external back to bytes must not look identical to never having declared it, or it stays keyed by the old predicate forever.

Also: a pre-existing flake, fixed in passing

cli::structured_log::tests::test_set_and_get_level asserts on a process-global log level while two sibling tests set it concurrently. It failed once in 13k — the worst frequency: rare enough to look like noise, common enough to erode trust in every other result. Serialised with a mutex that ignores poisoning.

Verified

13013 lib tests
8 equivalence tests (6 unit + 2 end-to-end through the binary)
clippy --all-features --all-targets -D warnings clean
io_tracking.rs 99.1 (A+) — above its own baseline, after splitting the new
                            code into its own module rather than inlining it

🤖 Generated with Claude Code