fix(hindsight): bound the append-mode session turn buffer

[Vissirexa]

Part of #62950. Rebased onto current main and narrowed to the hindsight half only — the TTS model-cache half of this PR landed via #73512 (commit 15673dfce7, authorship preserved), so it has dropped out of the diff. The contributor email mapping commit also dropped: contributors/emails/tikkanadityajyothi@gmail.com now exists upstream, and contributors/README.md explicitly forbids new AUTHOR_MAP entries in scripts/release.py.

What's left is one commit in one subsystem.

What

_session_turns (plugins/memory/hindsight/__init__.py) accumulated every turn's text for the whole session, reset only at session boundaries — so a never-ending session grew without bound, independent of context compaction.

The fix is scoped to append mode only:

• In append mode each retain ships just the delta since the last watermark (_session_turns[_last_retained_turn_count:]), and once retained a turn is never read again — sync_turn always slices from the watermark, and flush-on-switch flushes what's left. So after an append retain the buffer now drops the retained prefix (clear() + reset the watermark), bounding it to the un-retained tail.
• Overwrite mode is deliberately untouched. Legacy/overwrite APIs resend the whole session each retain (list(self._session_turns)) because each retain replaces the document, so that path must keep every turn.

This is distinct from #40605 (which fixed the quadratic re-send on append); the in-RAM list itself is what this bounds.

How to test

scripts/run_tests.sh tests/plugins/memory/test_hindsight_provider.py

TestSessionTurnsBufferBounding covers: append trims the retained prefix while shipping every turn exactly once (no loss, no duplication), the buffer stays bounded across 100+ turns, and overwrite mode keeps the full buffer.

Full tests/plugins/memory/ passes on the rebased branch — 18 files, 555 tests, 0 failures under the canonical scripts/run_tests.sh.

Platforms tested: macOS (Apple Silicon), Python 3.11.

[teknium1]

Thanks for the focused memory-bound fixes. Current checkout HEAD still retains all Piper/KittenTTS cache entries (tools/tts_tool.py:2020-2025, tools/tts_tool.py:2117-2123) and keeps append-retained Hindsight turns in _session_turns after advancing only a watermark (plugins/memory/hindsight/__init__.py:1637-1696). Commit 8be65e07d0fa addresses those exact paths with bounded LRU caches and append-only buffer trimming while preserving the overwrite path.

No verified correctness issue found in this review. Existing integration tests already exercise both provider call paths, and the added tests cover LRU order plus append-mode no-loss/no-duplication semantics.

Automated hermes-sweeper review.

[teknium1]

The TTS model-cache half of this PR landed on main via #73512 (merge f3cc2bc) as commit 15673df with your authorship preserved — _piper_voice_cache/_kittentts_model_cache now run through a small LRU. The hindsight turn-buffer half is a different subsystem and remains for the memory-subsystem triage, so this PR stays open for that half.

[Vissirexa]

Thanks — rebased onto current main and narrowed this PR to the hindsight half, so the remaining diff is scoped to the memory subsystem for triage.

• The TTS model-cache commit drops out cleanly (already on main via fix(tts): honor config.yaml endpoints across all providers + bounded upstream buffering #73512).
• The AUTHOR_MAP chore commit also drops: contributors/emails/tikkanadityajyothi@gmail.com is upstream now, and contributors/README.md says not to add new entries to scripts/release.py.

What's left is a single commit touching plugins/memory/hindsight/__init__.py (+9/-1) and its tests. scripts/run_tests.sh tests/plugins/memory/ — 18 files, 555 tests, 0 failures. Title and description updated to match.