fix(memory): send only incremental turns in Hindsight append mode

[estridell]

What does this PR do?

Fixes quadratic token usage in HindsightMemoryProvider.sync_turn() when the Hindsight API supports update_mode="append" (>=0.5.0).

The bug: sync_turn() sends the full accumulated _session_turns on every retain. With append-mode documents, this means turn 1 is sent N times, turn 2 is sent N-1 times, etc. — O(n²) token usage that burns through API credits on long sessions.

The fix: Track how many turns have already been retained (_last_retained_turn_count). In append mode, send only _session_turns[_last_retained_turn_count:] (the new turns since last retain). Legacy overwrite mode is unchanged — full-session snapshots are still required.

_session_turns itself is NOT cleared, so on_session_switch() can still flush partial buffers when retain_every_n_turns > 1.

Related Issue

Fixes the token-cost half of the append-mode integration gap introduced in 3082fa0 (feat: probe API for update_mode='append' support, dedupe across processes).

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• plugins/memory/hindsight/__init__.py: Add _last_retained_turn_count counter, use it to slice only new turns in append mode, reset in __init__/initialize()/on_session_switch()
• tests/plugins/memory/test_hindsight_provider.py: Rename existing test to clarify it covers legacy overwrite mode, add new test test_sync_turn_append_mode_sends_only_new_buffered_turns that verifies incremental sends

How to Test

1. pytest tests/plugins/memory/test_hindsight_provider.py -q — 97 tests pass
2. pytest tests -q -k 'hindsight' — 102 tests pass, 1 skipped
3. The new test test_sync_turn_append_mode_sends_only_new_buffered_turns specifically verifies that the second retain does NOT contain turns from the first retain

Checklist

• My commit messages follow Conventional Commits (fix(memory): ...)
• I searched for existing PRs to make sure this is not a duplicate
• My PR contains only changes related to this fix
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes

Notes

The on_session_switch() flush in append mode will re-send some already-retained turns (redundant but safe). This is the correct tradeoff — data safety > perfect efficiency. The flush code is upstream and should not be touched in this PR.

[teknium1]

This looks implemented on current main by the later Hindsight delta-retain fix. Automated hermes-sweeper review found the requested behavior in the provider and covered by regression tests.

Evidence:

• plugins/memory/hindsight/__init__.py:577 initializes _last_retained_turn_count for append-mode retain tracking.
• plugins/memory/hindsight/__init__.py:1471–1483 sends only self._session_turns[self._last_retained_turn_count:] when update_mode == "append", while legacy overwrite mode still sends the full session.
• plugins/memory/hindsight/__init__.py:1533–1536 advances the append watermark after enqueueing the retain.
• tests/plugins/memory/test_hindsight_provider.py:805 verifies the second append retain does not include already-retained turns.
• Commit 09d66037f8f7bc5bd879ed8128273fb6780a009f landed this on main as fix(hindsight): send only new-turn delta on append retains instead of whole session (#40605).