fix(honcho): serialize concurrent message flushes

[Diaspar4u]

What changed

This PR is intentionally limited to one current-main concurrency defect that is not covered by #83500:

• serialize Honcho message flushes so two concurrent flush paths cannot select and upload the same unsynced messages twice
• add a deterministic regression that reproduces the duplicate upload on current main

Superseded scope removed

#83500 already carries the original PR's saveMessages, writeFrequency, migration ownership, and normal async-writer shutdown work. None of that scope is resubmitted here.

Validation

• regression fails on current main and passes with this commit
• scripts/run_tests.sh tests/honcho_plugin -q — 294 passed
• focused Ruff — passed
• worktree and upstream-range git diff --check — passed
• upstream-range commit SSH-signature verified

[alt-glitch]

This was generated by AI during triage.

Related to #19650 and #35210: this active salvage combines both documented persistence-contract fixes and adds lifecycle coordination. It is not a duplicate of either open predecessor; maintainers should choose the integration path.

[teknium1]

Thanks for consolidating the two persistence-contract fixes. The underlying premise still holds on current main: sync_turn() directly calls _flush_session() at plugins/memory/honcho/__init__.py:1348-1353, while the writeFrequency routing boundary is HonchoSessionManager.save() at plugins/memory/honcho/session.py:499-522; saveMessages is parsed at plugins/memory/honcho/client.py:574-576 but is not checked by sync_turn().

Problems

• The branch predates bd1a850fa2, which made the async writer lazy and requires _ensure_async_writer() before an async enqueue (plugins/memory/honcho/session.py:511-514). The proposed lifecycle rewrite must preserve that path; otherwise queued async saves have no worker.
• bd1a850fa2 also introduced the package-level network-hermetic Honcho test fixture. The older manager construction pattern in this branch should not be restored.

Suggested changes

• Salvage the two focused sync_turn() changes, then port lifecycle coordination and its tests onto the current lazy-writer implementation while retaining the hermetic test fixture.

Automated hermes-sweeper review.

[Diaspar4u]

@greptileai

[Diaspar4u]

Re-review requested. I rebased this onto current main and removed every behavior superseded by #83500. The PR now contains only the independently reproducible concurrent-flush duplication fix: without serialization, two flush paths can upload the same unsynced messages twice.

The regression fails on current main, passes here, and also passes when the minimal patch is applied on top of #83500. Focused Honcho/memory validation passed (314 tests), along with Ruff and both diff checks. The single commit is SSH-signed. Please review the reduced scope for merge.

[Diaspar4u]

Rebased onto current upstream main at 1d3d021282; scope remains limited to serializing concurrent Honcho flushes so two writers cannot upload the same unsynced messages.

Validation on head a36fc940d9ae:

• Honcho plugin suite: 294 passed
• focused Ruff and both diff checks: passed
• upstream-range commit SSH-signature verified
• independent security/correctness review: passed

@teknium1 @alt-glitch please review the current head.

[Diaspar4u]

Rebased the focused concurrent-flush serialization fix onto current main at head c9c049405a; its upstream-range commit has a verified SSH signature. Scope remains limited to preventing duplicate uploads when flush paths race.

Validation: 294 Honcho plugin tests passed; focused Ruff, git diff --check, and upstream-range diff checks passed.

@coderabbitai review

@teknium1 @alt-glitch please re-review the current head.

[Diaspar4u]

Superseded by #86094, rebuilt from current main with only the remaining concurrent-flush fix and focused regression test.