feat(hindsight): add on_pre_compress hook to preserve context before compression

[pyrate-llama]

Summary

When Hermes compresses its context window to save tokens, the original messages are summarised and discarded. This adds an on_pre_compress lifecycle hook that fires just before compression, capturing the last 10 user/assistant messages and persisting them to Hindsight.

Messages are tagged [Pre-compression context] and retained with context=pre_compress. Runs in a background thread — never blocks compression.

Changes

• on_pre_compress() method on HindsightMemoryProvider (follows MemoryProvider ABC optional hook)
• _precompress_thread with proper join on shutdown
• on_pre_compress registered in plugin.yaml hooks list

Test plan

• Tested locally with Hermes v0.7.0 + Hindsight
• Pre-compression messages appear in knowledge graph after compression
• Background thread confirmed non-blocking

Motivation

Closes vectorize-io/hindsight#870 — standalone plugin discontinued, migration to official Hermes location.

[alt-glitch]

Related to #11753 (pre-compression task state preservation hook) and closed #3741 (pre_context_compress hook). P3 per plugin-provider-demotion — Hindsight memory plugin feature.

[alt-glitch]

Related to #11753 (pre-compression task state preservation hook) and closed #3741 (pre_context_compress hook).

[teknium1]

Thanks for the Hindsight preservation work. This is an automated hermes-sweeper review; current main already provides the requested persistence guarantee through a stronger lifecycle.

• plugins/memory/hindsight/__init__.py:1603 retains every completed user/assistant turn by default through the serialized writer queue.
• plugins/memory/hindsight/__init__.py:1819 flushes buffered turns during a session switch, including a compression boundary; commit 0a5ee01e487a5a4e0e3637ecce6c2a41546c9457 specifically serialized that flush behind queued retains.
• agent/context_compressor.py:2885 summarizes the middle window while preserving the recent tail. The proposed messages[-10:] capture at PR line 897 would therefore retain messages that normally survive compaction rather than the discarded region.

Closing as implemented on main.