test(ep0f): registry hermetic classifier mock + logger timeout diagnostics (kata ep0f)

[danshapiro]

Summary

Root-cause fixes for two of the five test-hygiene load-flake families tracked as kata ep0f's timing-pin class (observed under host load 54–94 across four full-suite executions during the rust-sidecar-sdk-0-3 gate campaign). Test-files only; zero production changes.

logger.separation — exit-before-flush race

The three tsx -e probes ended setTimeout(() => process.exit(0), 25|50), which races rotating-file-stream's lazy open + async first write. Measured under induced load ~90: logger.ts import takes 368ms; its 'Resolved debug log path' record lands at 390ms — a ~15× overshoot of the 25ms exit window; historical passes survived on ~3ms of luck. On a miss the record can never land, and waitForFileContent treats empty as not-yet, burning its full 30s gate on an impossible condition.

• Probes now self-poll their own resolveDebugLogPath() target and process.exit(0) only once their expected content is on disk (cap 10s; PROBE-FATAL via synchronous fs.writeFileSync(2, …) + exit(1) otherwise).
• waitForFileContent timeout errors now include the logDir listing with per-file sizes and the probes' captured stdout/stderr.
• All five tests assert probe exitCode === 0 and absence of PROBE-FATAL after content waits; the multi-probe waits run concurrently.

terminal-registry :3317 — real worker spawn + 12.5 GB DB inside a 1s window

vi.waitFor's default budget (interval 50ms / timeout 1000ms, Vitest 3.2.4) wrapped a REAL fire-and-forget classification: nested node:worker_threads spawn opening the host's 12.5 GB opencode.db. Measured 222ms idle / 638ms loaded (single-shot) against the 1000ms cap. CI containers lack the DB + fast path, so the flake was dev-host specific.

• File-wide vi.mock('../../../server/coding-cli/providers/opencode-subagent-query.js') mirroring the two sibling test files; mockResolvedValue(false) in a describe-local beforeEach (the outer describe's vi.resetAllMocks() wipes factory-set implementations — the trap is documented in the test).
• Stale 'no opencode.db here' comment corrected; the misleading "both directions" test renamed to its true child→root scope (real both-directions coverage stays in terminal-registry.rebind-metadata-resync.test.ts; real classifier coverage in opencode-subagent-query.test.ts).
• Sibling header comment in bind-reclassify-guard.test.ts reworded to match.

Verification

• RED: neither flake reproduced in 15 loaded iterations (load 74–91); fallback ladder used: kata-recorded failures + captured margin measurements + live zero-files reproduction of the mechanism.
• GREEN under induced load (48 detached CPU burners, load 74–94): registry focused ×50, registry full-file ×5, logger full-file ×20, sibling files ×3 each, combined back-to-back pass.
• Failure branches exercised as one-off manual receipts (PROBE-FATAL marker delivery + rc=1; diagnostics branch forced inside the real harness then reverted).
• Coordinated full-suite gate: cloud backend, green at the test-file HEAD 42085e4b8; the three later commits are plan-doc-only.
• Failure-mode discovery filed as kata zqz3 (uncreatable FRESHELL_LOG_DIR → infinite rfs mkdir retry starving the event loop) — found via strace while generating receipts, explicitly out of scope here.
• Kata f884 filed: nothing lint/typechecks test/** today.

Review journey (the-usual-beta)

• Plan loop: FAILED/non-converged at its 3-round cap; all 13 findings (Majors incl. a phantom-GREEN criterion and a silently-degrading vitest CLI form mistake) dispositioned with committed plan repairs.
• Delta loop: PASSED at round 3 (r1/r2 Majors repaired; each repair episode passed the beta focused-recheck loop).
• Recap with full receipts: .worktrees/.the-usual-logs/ep0f-deflake/recap.md.

Kata ep0f stays open for the three remaining families (remote-proxy, opencode-serve-manager, codex-session-flow).