fix(hindsight): use descriptive context and stable document_id in sync_turn

[mattverlaque]

Summary

• sync_turn() now passes a descriptive context field (includes session_id) instead of the generic "conversation" string
• Uses a stable document_id keyed to session-{session_id} so repeated retains upsert instead of creating duplicate documents

Why

Per Hindsight best practices:

• Context field: "High-impact on extraction quality. Always set it." — the generic "conversation" provides no signal to the extraction LLM
• document_id: "Always use the same ID for growing conversations. Do NOT use random UUIDs per retain call; this creates duplicates." — without document_id, every turn creates a new document, leading to 100+ near-duplicate memories per session

Test plan

• Verified aretain accepts document_id kwarg (hindsight-client SDK)
• Run with active Hindsight instance, confirm single document per session in bank
• Verify no invalid JSON regressions

🤖 Generated with Claude Code

[teknium1]

Thanks for the focused Hindsight retention improvement. This is an automated hermes-sweeper review; current main already provides the requested behavior through a more complete implementation.

• 3082fa0829e0df4ce682358481fb59275b31a46e added capability-gated deduplication and is included in v2026.5.7.
• plugins/memory/hindsight/__init__.py:1631-1686 uses a stable session ID plus update_mode="append" on supported Hindsight APIs, while retaining configurable descriptive context.
• plugins/memory/hindsight/__init__.py:1181-1200 preserves a per-process document ID for legacy APIs, avoiding the resume-overwrite behavior that an unconditional stable ID would cause.
• tests/plugins/memory/test_hindsight_provider.py:1346-1378 covers both the legacy fallback and stable append-capable path.