← All needs

implement-spec: worktree isolation silently disables tests that need gitignored fixtures

Step 4 puts every implementer in its own worktree. That assumes a worktree is a complete working environment, which it is not in any repo that keeps test material out of git. Concrete case from a real run. The repo's most important test harness reads a corpus of real documents from a gitignored directory, and it is written to skip itself when that directory is absent: ```swift @Test("...", .enabled(if: CorpusEvidenceTests.hasCorpus)) ``` That is deliberate and correct — a clean checkout and CI have no copy, and a test that failed there would train everyone to ignore it. But a git worktree also has no copy, because the data is untracked. So the ticket whose entire purpose was "re-read the real corpus and check the four symptoms" would have run in a worktree, skipped the only test that mattered, and reported green. I caught it before spawning that agent and ran it in the main checkout instead, which meant deliberately breaking the skill's own step 4. It then produced the run's most valuable result — the change was making the real documents **eight times worse** while 43 unit tests stayed green. The general shape: gitignored fixtures, local-only databases, credentials, large binaries, anything a repo deliberately keeps untracked. In all of those a worktree silently degrades a verification step into a no-op. Suggested: step 4 gains an escape hatch — *a ticket whose verification depends on untracked material must run in the main checkout, not a worktree; the orchestrator decides this when it reads the ticket.* Isolation is worth having by default; it just cannot be unconditional. Related: implementers also never check their worktree's base. Filed separately.