buzz-relay lib tests are archived but not run in CI — 771 of 804 never execute, including the mesh and tunnel/lease suites #3461

Summary

buzz-relay's library tests are compiled in CI but almost none are executed. Of 804 tests in cargo test -p buzz-relay --lib, CI runs 12 — the #[ignore]d subset of api::invites::tests. All 771 non-ignored tests, including the entire mesh, tunnel/lease, subscription, media-auth, and pubsub-fanout suites, execute nowhere in any workflow.

This looks unintentional rather than a deliberate policy: the backend-integration job already provisions Postgres, Redis, and MinIO, and the nextest archive is explicitly built with -p buzz-relay --lib. The filter that selects tests from that archive was written for the invite suite and appears never to have been broadened.

Evidence

Every test invocation across .github/workflows/:

The archive is created at ci.yml:351 with -p buzz-db -p buzz-relay -p buzz-test-client --lib, so the binaries exist; only those two -E filters ever select from it, and both pass --run-ignored ignored-only.

Reproduce the counts:

cargo test -p buzz-relay --lib -- --list | grep -c ': test$'                    # 804
cargo test -p buzz-relay --lib -- --list --ignored | grep -c ': test$'          #  33
cargo test -p buzz-relay --lib -- --list --ignored | grep -c 'api::invites'     #  12  ← what CI runs

Why it matters

This is the reason #2458 survived in main. api::mesh_demo::tests::demo_join_forwarded_arm_round_trips_echo fails roughly half the time on a machine with Redis — I measured 5/10 failures on origin/main @ 22be8bb, and @tlongwell-block and two others reproduced it independently this week. A test with that failure rate cannot survive a CI that runs it. It survived because CI never has.

The same blind spot covers tunnel::directory (lease acquisition, fencing, TTL takeover) and tunnel::reliable (ownership routing, renewer loss signalling) — the coordination primitives underneath the mesh.

There is a second, related trap that @tlongwell-block flagged on #2458: 11 of these tests call redis_directory_if_available() and return early, silently, and green when Redis is unreachable (tunnel/directory.rs ×4, tunnel/reliable.rs ×4, api/mesh_demo.rs ×2, handlers/event.rs ×1). So even running them locally without infra proves nothing, and the output is indistinguishable from real coverage.

Cost to fix

Approximately zero. The full non-ignored suite against local Postgres + Redis:

test result: ok. 771 passed; 0 failed; 33 ignored; 0 measured; 0 filtered out; finished in 2.22s

2.22 seconds, all green, in a job that already has the services running.

Suggested fix

Add a step to backend-integration (which already has Postgres/Redis/MinIO up and the archive downloaded) running the buzz-relay lib suite from the existing archive.

Ordering matters: enabling this today makes CI flake about half the time on the #2458 test. #2745 fixes that flake — I isolated the mechanism to cancel-safety in run_demo_echo with a 2×2 matrix at N=10 per cell (detail) — so the clean order is #2745 first, then switch this on.

Making the silent Redis skips loud is worth doing alongside, so a green run without infra can't read as coverage. #2745 already does this for the one live-transport variant it touches.

Happy to send both as a PR if the direction is right — I did not want to change CI policy unilaterally.