feat: add optional message embedding for semantic search (config-driven)

[gzsiang]

Motivation

Store embedding vectors alongside each assistant message to enable future semantic search over conversation history. The feature is fully opt-in — no embedding endpoint configured = zero overhead.

Changes

Only one file modified: hermes_state.py (177 insertions, 8 deletions).

• Schema v12: adds message_embedding BLOB column to the messages table (auto-reconciled by existing column reconciliation)
• Config: reads embedding.base_url / embedding.endpoint from ~/.hermes/config.yaml — defaults to None (disabled)
• Storage: append_message() computes and stores embedding vectors (packed float32) for assistant messages when configured
• Search: search_messages() re-ranks FTS5 results using cosine similarity when embeddings are available; falls back to pure vector search when FTS returns nothing
• Graceful short-circuit: if embedding.base_url is not set, _compute_embedding returns None immediately — no DNS calls, no overhead

How to use

# ~/.hermes/config.yaml
embedding:
  base_url: http://your-gpu-server:8081
  model: Qwen3-Embedding-0.6B
  dimension: 1024

Notes

• Pure FTS5 and LIKE searches remain unchanged when embedding is not configured.
• The hybrid re-rank weights are 0.3 (BM25 rank) + 0.7 (vector similarity).

[trevorgordon981]

LGTM. This is a clean, opt-in addition for semantic search over conversation history. The implementation is thoughtful:

• Schema v12 adds message_embedding BLOB column (backward-compatible via existing reconciliation)
• Config-driven: embedding.base_url/endpoint in ~/.hermes/config.yaml — disabled by default (zero overhead)
• Hybrid search: FTS5 BM25 + cosine similarity re-ranking (0.3/0.7 weights), with pure vector fallback
• Graceful: _compute_embedding short-circuits if base_url is unset — no DNS calls

Verified:

• Schema migration works (v12 columns present)
• Append with embedding disabled: PASS
• 195/198 state tests pass — 3 failures are expected (schema version assertions need updating from 11→12)

The PR is ready once the 3 test assertions are bumped to expect v12. Nice work.

[gzsiang]

Thanks for the review! Updated the 3 schema version test assertions to expect v12.

[gzsiang]

Friendly ping — this PR has been approved and the requested changes addressed. Happy to rebase or make any additional tweaks if needed. Would love to get this merged when you have a chance! 🙏