fix(gateway): bust cache on auxiliary compression changes

[Glucksberg]

Summary

• include auxiliary compression settings in the gateway agent cache signature, including a sanitized fallback_chain fingerprint
• support nested configuration paths while preserving the existing cache-key behavior
• invalidate a cached gateway agent when construction-time auxiliary compression feasibility can change
• add regression coverage for fallback-chain and nested missing-path invalidation

Why

Gateway agent construction resolves auxiliary compression feasibility and can cache an unavailable-client warning or a reduced compression threshold. Changing the provider, model, endpoint, context length, API mode, or fallback chain must construct a fresh agent instead of retaining that stale state.

The earlier Honcho memoization change was deliberately removed because it overlaps upstream PR #46385; this contribution is now limited to auxiliary compression cache invalidation.

Validation

• rebased onto 4281151ae859241351ba14d8c7682dc67ff4c126 on 2026-07-11; current head: 4ce5cd577eac58ce1f5d913b879bc9e2ede57367
• .venv/bin/python -m pytest -q tests/gateway/test_agent_cache.py - 81 passed
• .venv/bin/ruff check gateway/run.py tests/gateway/test_agent_cache.py and git diff --check upstream/main...HEAD passed

[teknium1]

Thanks for narrowing this to the gateway cache signature; the current-main premise is real: gateway/run.py:15743-15854 has no auxiliary compression fields, while initialization stores auxiliary context state (agent/agent_init.py:1618-1630) and feasibility is memoized per agent (agent/conversation_compression.py:490-497).

Problems

• gateway/run.py:15935-15940 omits auxiliary.compression.api_key. That is a supported config field (hermes_cli/config.py:1594-1600) and _resolve_task_provider_model() reads it for compression clients (agent/auxiliary_client.py:5895-5901). A key-only edit therefore still reuses an agent whose feasibility probe has already been marked complete. Fingerprint this field like fallback_chain, rather than retaining its raw value.

Suggested changes

• Add a signature-regression test where only auxiliary.compression.api_key changes.

Automated hermes-sweeper review.

[teknium1]

Please include auxiliary.compression.api_key too, using a stable digest. It is a supported compression setting and is read by _resolve_task_provider_model; without it, a key-only repair still reuses an agent whose feasibility probe was already memoized.