fix(dispatcher): restore dispatch.decision logging in CI

[thinmintdev]

/diagnose root cause

tests/dispatcher/test_router.py::test_decision_logging_runs_on_every_resolution failed only inside the full pytest suite (the target test passes in isolation, and so does the whole tests/dispatcher/ subtree). Test pollution. Captured-log output proved the production code DID emit dispatch.decision — but the test's _capture structlog processor never received it:

INFO     hal0-dispatch:router.py:778 {'event': 'dispatch.decision', ...}
AssertionError: assert 'dispatch.decision' in []

Bisection narrowed the polluting set to tests/memory/* + tests/omni_router/test_api_wiring.py. Cognee (imported transitively by the memory tests) calls structlog.configure(cache_logger_on_first_use=True) (see cognee/shared/logging_utils.py:435). With that flag set globally, the dispatcher's module-level log = structlog.get_logger("hal0-dispatch") proxy caches its .bind to Cognee's processor list on first use. A subsequent structlog.configure(processors=[_capture]) in the test does NOT retroactively rebind cached BoundLoggerLazyProxy instances, so log.info("dispatch.decision", …) keeps firing through Cognee's processors and _capture never sees the event.

Fix is in the test: pop the cached bind off the dispatcher's log proxy before exercising it (so the next .info() materializes a fresh BoundLogger bound to the _capture processor), then restore the cached bind in the finally block so subsequent tests see the same proxy state. Production code is unchanged — the dispatcher's logging is correct.

Offending commit

69418d7 feat(memory): ADR-0014 [memory.graph] schema + cognify gate + /api/memory routes (#257) (#287) — 2026-05-23. Introduced tests/memory/test_cognee_wrapper.py + tests/memory/test_graph_gate.py, which were the first tests in the suite to exercise Cognee initialization and therefore the first to set cache_logger_on_first_use=True globally.

The string dispatch.decision itself is unchanged since dc8891a feat: Phase 1 port (history verified via git log -S 'dispatch.decision'). No production regression.

Verification commands

From /tmp/hal0-dispatcher-fix on a fresh Python 3.11 venv (uv venv --python 3.11, uv pip install -e '.[dev]'):

• python -m pytest tests/dispatcher/test_router.py -xvs — 13 passed
• python -m pytest tests/dispatcher/ — 81 passed
• python -m pytest tests/dispatcher/ tests/api/ -x — 408 passed, 3 skipped (broader smoke per brief)
• python -m pytest tests/memory tests/omni_router/test_api_wiring.py tests/dispatcher/test_router.py::test_decision_logging_runs_on_every_resolution — 25 passed (exact polluting combo that previously failed)
• python -m pytest — 1649 passed, 3 skipped (full suite; was 1648 passed + 1 failed before the fix)
• ruff check src/hal0/dispatcher/ tests/dispatcher/ — All checks passed
• ruff format --check src/hal0/dispatcher/ tests/dispatcher/ — 15 files already formatted

Diff

One file, 16 insertions, 1 deletion — tests/dispatcher/test_router.py only. No production code changed.

Coordination

Four sibling PRs were blocked on this failure (#306, #307, #308, #309). They should re-run CI once this lands; their changes don't touch the dispatcher or memory areas so a green base should clear their python (3.11) runs.