fix(agents): add file_not_found correlation status for missing transcripts

[gokhanozdemir]

Summary

Interactive Claude sessions can fail to persist transcripts at their expected file paths. PTY-driven sessions weren't validated for file existence, so the correlation logic incorrectly marked non-existent files as "matched," producing empty session metrics. This adds a file_not_found correlation status so missing transcripts are handled gracefully instead.

Fixes #523

Changes

• Add a file_not_found status to the correlation type system
• Validate file existence before marking correlations as matched in the hook command
• Update the session synchronizer to skip gracefully when files are missing, instead of failing outright

Impact

No user-facing behavior change for existing successful correlations; missing-transcript cases now report a clear status instead of silently producing empty metrics.

Checklist

• Self-reviewed
• Manual testing performed
• Documentation updated (if needed)
• No breaking changes (or clearly documented)

[gokhanozdemir]

Adversarial review

Critical regression: event.transcript_path ? existsSync(...) : false forces correlationStatus to 'file_not_found' whenever transcript_path is empty. But codemie-code, codex, copilot-cli, opencode, and pi all fire SessionStart with transcript_path: '' by design (path not known yet):

• src/agents/plugins/codemie-code.plugin.ts:233
• src/agents/plugins/codex/codex.plugin.ts:191
• src/agents/plugins/copilot-cli/copilot-cli.plugin.ts:161
• src/agents/plugins/opencode/opencode.plugin.ts:210
• src/agents/plugins/pi/pi.plugin.ts:376

Nothing anywhere ever rewrites correlation.status after creation (verified repo-wide), so these sessions get stuck at file_not_found permanently, and SessionSyncer.sync() silently (success: true) skips them forever. This breaks metrics/conversation sync for most agents, not just interactive Claude PTY sessions — a much bigger regression than the bug being fixed.

Broken pre-existing test: tests/integration/agent-model.test.ts:329 skip-guards on correlationStatus === 'failed', written specifically to tolerate this exact issue (#523). The new code never produces 'failed' here — it produces 'file_not_found' — so the guard is dead and the test will fail instead of skipping gracefully.

Other issues:

• Race on re-entrant SessionStart (e.g. compact) can downgrade an already-matched live session to file_not_found with no recovery path.
• One-shot existsSync check at record-creation time doesn't cover a transcript appearing moments later, or one deleted/rotated after being marked matched.
• Duplicated existsSync/ternary/warn block across two branches in hook.ts; redundant dynamic import('node:fs') for a builtin already statically imported elsewhere.

Fix applied in a follow-up push: only mark file_not_found when a non-empty transcript_path was reported and is missing; keep prior behavior when no path was reported (the deliberate "not known yet" case); dedupe the two branches; use a static existsSync import; update the test's skip guard to also accept file_not_found.

[gokhanozdemir]

Cherry-picked the Windows CI teardown fix from #537 (110add2f) onto this branch to unblock the flaky migration-runner-ordering.test.ts failure we hit here.

That commit matches an existing established pattern already used in several test files in this codebase (claude.metrics-processor-names.test.ts:104, claude.metrics-processor-clear.test.ts:90, codex-models.test.ts:141, metrics-upload-contract.test.ts:112): wrap rmSync(..., { recursive: true, force: true }) with maxRetries: 5, retryDelay: 100, plus a try/catch that swallows the cleanup error. Reason: on Windows, rmSync's force option only suppresses ENOENT, not the ENOTEMPTY/EBUSY that can occur when a file handle (e.g. from a logger stream) is released a moment after the recursive delete already removed the children.

Not a hack — it's the codebase's already-established idiom for this class of flaky Windows teardown, applied here to the migration runner test that was failing on our Windows CI job.