fix: quarantine corrupt goal state

[danyel117]

Summary

• preserve non-empty whitespace/BOM/NUL-only state files as durable, owner-only goals.json.corrupt-<timestamp>-<uuid> artifacts before recovered state is written
• re-read the canonical file after quarantine and refuse to overwrite it if another process repaired or replaced it
• keep async and sync reads side-effect free and retain fail-closed behavior for malformed non-empty JSON
• report quarantine success, concurrent replacement, and quarantine failure through the OpenCode V1 application log, scoped to the configured state path and detached on plugin disposal
• degrade safely when quarantine storage is unavailable: report the failure and continue recovery instead of making every prompt fail indefinitely
• keep logging failures from blocking recovery and expose the quarantine path through console diagnostics for V2
• document the recovery behavior and cover V1, V2, exact 28,454-byte NUL evidence, whitespace/BOM evidence, reporter lifecycle, and forced quarantine-write failure

Verification

• bun run lint
• bun run typecheck
• bun test (221 passing)
• bun run build
• bun run pack:dry-run
• internal standards/spec review: approved
• Claude Opus review round 1: requested safe degradation when quarantine creation fails; fixed in 833a391 with regression coverage

Closes #35

AI assistance: implemented and reviewed with OpenCode agents and independently reviewed with Claude Opus.

[danyel117]

Claude Opus final review: APPROVE.

The first review found that a quarantine-write failure could reintroduce the original every-prompt failure loop. Commit 833a391 resolves it structurally: quarantine failures are reported and recovery continues, while only a successfully observed concurrent content replacement remains fail-closed. Verification read errors and ENOENT no longer create an indefinite failure.

Opus independently reproduced the source-change refusal, confirmed exact preservation for NUL/whitespace/BOM content, verified listener scoping/disposal/failure isolation, rebuilt dist byte-identically, and ran the complete gate: 221 tests, lint, typecheck, build, and pack. All GitHub CI checks are also green. No merge-blocking findings remain.