Parallel workers inflate the mutation score: overlay directories collide on non-unique mutant IDs

[mcgoughprogress]

Summary

Running gomu run with more than one worker reports a higher mutation score than a serial run of the same code, and the inflated figure is not stable between runs. The extra "kills" are phantoms: mutants that were never actually executed under test.

Serial runs are correct, so the defect is invisible unless you compare worker counts.

Reproduction

Measured against a single package (internal/ha of a private module, 108 mutants), --incremental=false, .gomu_history.json deleted between runs:

workers	killed	score
1 (ground truth)	79/108	91.9%
4	80/108	93.0%
8	80/108	93.0%

That package generates 108 mutants but only 93 distinct mutant IDs.

Root cause

Two things combine:

1. internal/mutation/engine.go does not produce unique mutant IDs — see GenerateMutants produces non-unique mutant IDs #2 for the mechanism. Duplicates are pervasive: 87 of this repository's own 123 source files generate at least one repeated ID.

2. internal/execution/overlay.go:53 derives each mutant's temp directory from that ID:

   mutantDir := filepath.Join(om.baseDir, fmt.Sprintf("mutant_%s", mutant.ID))

Two mutants with the same ID therefore share a directory. They overwrite each other's mutated source and overlay.json, and whichever finishes first calls CleanupMutation and deletes the directory out from under the other. The victim's go test -overlay ... then fails to load the package and exits non-zero, which internal/execution/engine.go:175 classifies as StatusKilled.

Net effect: phantom kills that inflate the score, non-deterministically, whenever --workers > 1.

Suggested fix

Fixing (1) alone would stop the collisions, but (2) is worth fixing on its own terms: mutant.ID carries no uniqueness contract that execution can rely on, it is owned by a different package, and it embeds the source file path — so the "directory" was really a nested path, mutant_internal/ha/ha.go_12.

Allocate the mutant directory from a counter private to the OverlayMutator instead. With that change applied, 4- and 8-worker runs return 79/108 (91.9%), matching the serial ground truth exactly.