Bug description

The Honcho memory tool schemas advertise focused retrieval controls that are not fully honored by the current implementation:

• honcho_context exposes an optional query parameter described as focusing/filtering context, but the handler currently calls get_session_context(..., peer=peer) without passing the query through.
• honcho_search accepts max_tokens, but HonchoSessionManager.search_context() returns the assembled representation/card blob without deterministic budget trimming.

This makes the Hermes tool surface noisier than the underlying focused Honcho/pgvector retrieval path: broad peer representation/card content can dominate even when a caller asks for a focused search or smaller token budget.

Expected behavior

• honcho_context(query=...) should pass the focused query into Honcho peer-context retrieval rather than silently returning broad session context.
• honcho_search(max_tokens=N) should bound its returned raw context to the requested approximate token budget.
• The existing JSON shape ({"result": ...}) should remain backward-compatible for a narrow bugfix.

Proposed first slice

Keep the first PR intentionally narrow:

1. Pass query from the honcho_context tool handler into HonchoSessionManager.get_session_context().
2. When a query is present, fetch focused peer context via peer.context(search_query=..., target=...) instead of broad session context.
3. Apply deterministic post-fetch trimming in search_context() based on max_tokens.
4. Add regression tests for the drift above.

Out of scope for the first PR

• Switching honcho_search to a new conclusions/query API.
• Adding top_k, min_score, MMR, or structured ranked result arrays.
• Changing peer routing semantics.

Those would be good follow-up work after the schema/implementation drift is fixed.