Configure structlog once, so captured logs are the logs that fired

[xmap]

Three tests failed intermittently under the parallel suite and passed when run alone: test_agent_subscribers_registration.py (×2) and test_language_model_seed.py. They were not flaky in the scheduling sense — they were order-dependent, and the order that broke them was any contract test building a second app first.

Cause

build_kernel() calls configure_logging() on every create_app(), and each call installed a new processor list. A logger already bound and cached against the first list keeps reading it. structlog.testing.capture_logs() works by swapping the live config's processors, so a cached logger never sees the swap: the capture comes back empty while the log line still prints to stdout.

The failing assertion was len(squat_warnings) == 1 getting 0, with the warning plainly visible in the test's own "Captured stdout call" section. A log that fired and a capture that saw nothing.

Caching on its own does not cause this. A probe confirmed capture_logs reaches a cached logger fine when the configuration is installed once; it is the second configure() that strands the binding.

Fix

Configure structlog once per process. The stdlib handler is still rebuilt on every call, so log level continues to follow the most recent caller — the only part a repeat caller actually wanted.

Fixed at the source rather than with a test fixture, for two reasons: a process that builds two apps gets loggers bound to a dead processor list regardless of tests, and a fixture would have to be remembered at every new call site.

The module's caching note claimed "only the first call's level/handler take effect", which was half wrong — the stdlib handler was rebuilt every call — and said nothing about the capture_logs consequence. It now describes what is true and why the guard is there, so it does not get removed as redundant.

Checks

The 18 tests in the combination that previously failed 3 now all pass. Unit tier 12,331 passed. Contract tier 3,350 passed. Architecture 29,732 passed. Ruff and pyright clean.

🤖 Generated with Claude Code

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
apps/api/src/cora/infrastructure
logging.py
Project Total

This report was generated by python-coverage-comment-action