fix: preserve valid reviewer coverage during output repair

[zzwong]

Summary

Preserve valid review results when coverage lists repeat a file, and make assignment-scope errors easier to correct. In the observed failure, 73 entries represented 71 assigned files. Two duplicate pairs caused the result to be discarded; the retry then lost coverage and introduced an out-of-assignment claim.

Changes

• Treat repeated inspected or skipped paths as one coverage claim without adding coverage or suppressing findings.
• Report invalid coverage positions using zero-based array indexes without echoing rejected path text into instructions. Preserve the retry-summary length cap.
• Keep out-of-assignment claims, conflicting inspected/skipped claims, missing coverage, and incomplete tool evidence blocking approval.

Verification

• Decoder regression tests failed before the repair and passed afterward.
• Structured-retry, missing-coverage, approval-clamp, and tool-evidence checks passed.
• Recorded-output replay normalized 73 entries to exactly 71 assigned claims; the invalid retry remained rejected.
• Verified the actual reviewer assignment after post-selection coverage expansion against the per-reviewer artifacts.
• Lint and build passed with Go 1.26.
• Full local tests passed except the installed-runtime preflight test, which timed out invoking its help command and reproduced on targeted rerun. No adapter behavior changes here.
• All 10 CI checks passed. The first test job hit a helper-startup deadline; a failed-job rerun passed without code changes. Three targeted local runs also passed; the timeout's root cause is not established.

Live outcome

A scoped local binary completed a live review with valid output from all six reviewers. The repository reviewer reported 71 inspected files, zero skipped files, and zero findings. The replay isolates the decoder repair; the live run demonstrates workflow completion rather than sole causation or guaranteed validation of future output. Application-specific identifiers and evidence are retained privately.

For reproduction, compare the recorded prompt's assignment.files with the corresponding per-reviewer slices and initial structured output. Post-selection expansion precedes reviewer validation-scope construction.

Closes #586.