fix: thread outcomes through router fallback, widen response cache key

[shrdgn]

What & why

Two small, unrelated correctness fixes found during a scheduled repo review, bundled together since each is a one-line-cause / self-contained change:

1. router.route_async silently disabled the routing learning loop on classifier fallback

router._call_classifier's two fallback paths — a classifier error (text is None) and an unrecognized classifier reply — both called route(body, config) without the outcomes argument, even though route_async itself receives outcomes and the normal (non-model-mode) heuristic path passes it through correctly. Since route_request() (the server's actual entry point) always goes through route_async, any time router.mode: model's classifier errored or replied with neither SOLO nor FUSE, that request's routing decision silently bypassed the EMA-based learning nudge (openfusion.outcomes.OutcomeStore) instead of falling back to the heuristic with the learned nudge still active.

Fix: pass outcomes through on both fallback call sites (openfusion/router.py).

2. responsecache.cache_key omitted self-fusion and debate config fields

cache_key already accounts for strategy, aggregator, panel/judge models, tools, and sampling params, but not config.self_fusion.n / temperature_spread / seed_offset, nor config.debate.rounds. Two server configs differing only in self-fusion sample count/spread or debate round count produce identical cache keys and can serve each other's cached answer, even though they make a different number of upstream calls and synthesize differently — the same class of staleness bug as the already-open #153/#155 (which covers ToolsConfig fields specifically; this covers a different, still-missing set of fields).

Fix: add the four omitted fields to the payload dict in cache_key (openfusion/responsecache.py).

Both fixes include regression tests that fail on main and pass with the fix.

How it was tested

• ruff check . passes
• pytest -q passes (493 passed, no live network)
• New behavior has a test (tests/test_router.py::test_route_async_classifier_error_still_consults_outcomes, ::test_route_async_unrecognized_response_still_consults_outcomes, tests/test_responsecache.py::test_cache_key_differs_on_self_fusion_and_debate_config)
• Docs updated (CHANGELOG.md Fixed entries)
• No secrets, prompts, or response bodies added to logs or metrics
• Quality/cost claims backed by a bench/run.py number — not applicable, no quality/cost behavior change

Notes for reviewers

Neither fix touches request/response shape or config schema — both are pure bug fixes within existing modules (router.py, responsecache.py). No follow-ups needed; this closes both gaps completely as scoped.

---

Generated by Claude Code