Summary

• pass honcho_context(query=...) through to focused Honcho peer context retrieval
• bound honcho_search(max_tokens=...) output with deterministic token-budget trimming
• add regression coverage for the focused-query and budget controls

Fixes #29402.

Rationale

The Honcho tool schema already advertises a focused query parameter for honcho_context and a max_tokens budget for honcho_search. Before this change, the context tool ignored query, and search_context() assembled representation/card output without applying the requested budget. That made the memory tools noisier than callers expect from focused retrieval.

This PR keeps the behavior intentionally narrow and backward-compatible: the tools still return the existing { "result": ... } shape, and it does not change peer routing semantics or introduce a new ranking API.

Test plan

• RED verified before implementation:
  • scripts/run_tests.sh tests/honcho_plugin/test_session.py::TestPeerLookupHelpers::test_search_context_applies_max_tokens_budget tests/honcho_plugin/test_session.py::TestConcludeToolDispatch::test_honcho_context_passes_query_to_manager tests/honcho_plugin/test_session.py::TestGetSessionContextFallback::test_query_uses_focused_peer_context_even_when_session_context_is_cached -q -> 3 failing tests
• GREEN / targeted verification:
  • scripts/run_tests.sh tests/honcho_plugin/test_session.py::TestPeerLookupHelpers::test_search_context_applies_max_tokens_budget tests/honcho_plugin/test_session.py::TestConcludeToolDispatch::test_honcho_context_passes_query_to_manager tests/honcho_plugin/test_session.py::TestGetSessionContextFallback::test_query_uses_focused_peer_context_even_when_session_context_is_cached -q -> 3 passed
  • scripts/run_tests.sh tests/honcho_plugin/test_session.py -q -> 118 passed
  • scripts/run_tests.sh tests/honcho_plugin -q -> 270 passed
  • git diff --check
  • non-ASCII added-line guard -> non_ascii_added_lines=0

Follow-ups intentionally left out

• adding structured results arrays / top_k / score metadata
• switching to a conclusions/query search endpoint
• MMR/dedupe or retrieval-quality benchmarks