fix(honcho): dedupe repeated message flushes

[erosika]

Summary

• add bounded per-session batch fingerprints inside the Honcho session manager
• skip repeated Honcho uploads from stale async queue/retry/session-end flush paths after a successful sync
• keep the fix scoped to the Honcho plugin only

Test Plan

• scripts/run_tests.sh tests/honcho_plugin/test_honcho_idempotency.py tests/honcho_plugin/test_session.py::TestConcludeToolDispatch::test_sync_turn_strips_leaked_memory_context_before_honcho_ingest tests/honcho_plugin/test_async_memory.py -q

[alt-glitch]

Implements part of #9404. Related to #17153 (gateway-restart dedup). This adds batch fingerprint dedup for async/retry/session-end flush paths.

[alt-glitch]

Implements part of #9404

[teknium1]

Thanks for narrowing the change to the Honcho plugin. The duplicate-ingest problem remains present on current main: plugins/memory/honcho/session.py:440-446 creates and uploads messages from content alone, with no stable Hermes source-event identity.

Problems

• The new guard at plugins/memory/honcho/session.py:385 treats an equal role/content batch as a replay. That drops legitimate repeated turns; the #9404 discussion explicitly requires repeated content from distinct source events to remain ingestible.
• The check and record operations at plugins/memory/honcho/session.py:385 and :399 are not synchronized. Current main can call _flush_session() via the async writer (session.py:471), flush_all() (session.py:534,544), and the background sync_turn() path (__init__.py:1232-1248), so two callers can both pass the check before either records the fingerprint.

Suggested changes

• Re-scope deduplication around stable Hermes source-event identifiers plus chunk identity, with an atomic claim/commit protocol.
• Test replay of one source event separately from two equal-content turns, and cover concurrent flush entry points.

Automated hermes-sweeper review.

[teknium1]

This key contains only roles and content, so it cannot distinguish a retry from a later legitimate turn with the same text. #9404 requires source-event idempotency while preserving repeated messages; please key this on stable Hermes event/turn and chunk identity instead.

[teknium1]

Recording the fingerprint only after add_messages() and without synchronization leaves a check-then-act race: concurrent async/session-end flushes can both pass the earlier membership test and upload the same batch. Use an atomic per-event claim/commit protocol with rollback on failed upload.