…nt drops (NousResearch#43936)

The session-DB flush in `_flush_messages_to_session_db` was position-based
(`messages[max(start_idx, _last_flushed_db_idx):]`). It assumes
`messages == conversation_history + this turn's new messages`, which breaks two
ways, both reproduced live:

1. Overlapping turns on the cached agent corrupt the shared
   `_last_flushed_db_idx` (it indexes the turn-local `messages` array) → the
   earlier completed turn flushes an empty slice and its delivered assistant
   row is never written.
2. `repair_message_sequence` compacts `messages` in place below
   `len(conversation_history)`, so `start_idx > len(messages)` permanently →
   self-reinforcing drop loop.

The merged NousResearch#44837 fix (NousResearch#45260) clamps `_last_flushed_db_idx` to `len(messages)`
but leaves `start_idx = len(conversation_history)` unbounded, so shape 2 is
still an empty slice. No index arithmetic survives the two arrays diverging.

Replace the positional slice with identity-based dedup: stamp each persisted
dict with `_db_persisted`, recognize this turn's `conversation_history` entries
by object identity (`id()`) and stamp instead of re-write. A message is written
exactly once regardless of how the arrays shift; the NousResearch#860 and NousResearch#31507 guards
hold by construction. `_db_persisted` is underscore-prefixed and already
stripped from API payloads. The compression split clears the stamps so the
surviving messages re-write into the new session row. `_last_flushed_db_idx` is
kept updated for legacy readers but no longer decides what is written.

Complementary to NousResearch#43962 (backfills delivered text after an interrupt in
turn_finalizer.py); this prevents the drop at the flush layer and also covers
the non-interrupt repair-shrink path.

Regression: tests/run_agent/test_identity_flush.py (5, incl. an explicit
test that the NousResearch#44837 clamp still drops where identity flush persists) +
test_message_sequence_repair, test_860_dedup, test_compression_persistence all
GREEN.