[lnieuwenhuis]

Summary

Test-only fix for the intermittent SEC-04 capture flake (a_contentless_provider_payload_is_not_written_to_logs failing with an empty captured log). Product logging behavior is unchanged; no assertions, markers, retries, timeouts, DB gating or suite parallelism were touched.

Root cause

tracing-core caches each callsite's Interest process-wide on first execution. While at most one dispatcher has ever been registered, a first hit is resolved from the registering thread's default subscriber (Dispatchers::rebuilder() -> Rebuilder::JustOne). Under a parallel cargo-test run, a gateway test thread without a scoped subscriber could therefore cache Interest::never for the shared upstream_photo_failure callsite, silently dropping the WARN event before it reached the capturing fmt subscriber on the test thread (verified interleaving: tracing 0.1.44 / tracing-core 0.1.36 / tracing-subscriber 0.3.23, independent second look in work/security-logtest/second-look.md).

Change

apps/backend/src/legacy_api/tests/mod.rs only:

• Register one permanent no-op (NoSubscriber) dispatch in capture_logs before the scoped capturing subscriber, so the global registry is never on the JustOne fast path while captures are active (Interest::and(never, always) == sometimes, decided per thread by enabled()).
• Add an isolated deterministic regression: a scopeless std::thread first-executes a warn callsite reserved for the test, then the capturing thread re-emits the same callsite and asserts it was captured. It fails on the previous tree (reproduced in isolation) and passes with the guard.
• No retries, no suppressions, no whole-suite serialization, no product-code change.

Evidence

Check	Result
Regression isolated, pre-fix (guard disabled)	FAILED (event dropped)
Regression isolated, with fix	passed
Full backend suite, disposable DB, 50 consecutive parallel runs (237 tests each, 0 ignored)	0 SEC-04 capture failures (baseline before fix: 5/15 runs failed)
Negative control: raw provider body + raw payload temporarily logged	both SEC-04 content tests failed with the markers, unmutated invalid_json test still passed; mutations reverted, tests green
cargo fmt --all --check, cargo check -p macro-tracker-backend --locked	clean
Secrets scan (gitleaks git origin/staging..HEAD)	no leaks

Two remaining failures over the 50 runs were the pre-existing barcode stub transport flake (all_genuine_misses_still_return_not_found, a_successful_fallback_wins_over_an_outage: 503 vs 200, no log capture involved) — explicitly out of scope and not bundled here.

App version stays 3.6.0: test-only change, no feature behavior.