fix(acp): recover from persist() failure instead of cascading (#3366)

[OneStepAt4time]

Summary

Fixes #3366

FileAcpLocalStorageProfile.persist() chained writes onto writeChain without error recovery. A single failed write (ENOENT, EACCES, inode change) permanently poisoned all subsequent writes because every .then() chained onto the rejected promise — blocking ALL session creation.

Root Cause

// Before: chains .then() — one rejection poisons everything
this.writeChain = this.writeChain.then(
  () => writeFile(tmpFile, content, "utf8").then(() => rename(tmpFile, this.config.filePath)),
);
await this.writeChain; // if previous rejected, this rejects too

Fix

// After: catches errors, resets chain, retries on previous failure
const prevChain = this.writeChain;
this.writeChain = prevChain
  .then(
    () => doWrite(),           // previous succeeded → do this write
    () => doWrite(),           // previous failed → still try this write
  )
  .then(() => { this.persistError = null; })
  .catch((err) => {
    this.persistError = err;
    logger.error({ ... });
    unlink(tmpFile).catch(() => {});  // clean up stale tmp
    this.writeChain = Promise.resolve();  // RESET chain
  });
await this.writeChain;

Changes

• persist(): error recovery with chain reset on failure
• stop(): catches pending chain rejections so shutdown always completes
• getPersistError(): new method for diagnostics/health checks
• 3 regression tests: recovery after failure, cascade prevention across multiple failures, graceful shutdown with pending rejection

Verification

npx tsc --noEmit: ✅ Zero errors
npm run build: ✅ Success
npm test: ✅ 4211 passed (9 pre-existing failures on develop too)
New tests: 3/3 passed

Verification

• Aegis server: v0.6.7-preview.1
• Commit: f18a0cf
• Session: direct implementation (Aegis prompt delivery regression P0 regression: prompt delivery broken after async delivery pipeline (#3253) #3271)
• Tests: ✅ 3 new tests pass + 4211 existing pass (9 pre-existing failures)

[aegis-gh-agent]

✅ Approved.

P1 fix — clean error recovery pattern for the persist() cascade bug.

• persist() catches errors, logs via structured logger, cleans up stale tmp file, resets writeChain so next write gets a fresh attempt.
• Both .then() handlers (success + recovery) execute the write — ensures no cascade.
• stop() swallows chain errors — clean shutdown guaranteed.
• getPersistError() on interface for diagnostics.
• 3 tests covering recovery, multiple sequential failures, and clean stop.

CI green. All merge gates pass. Merging.