test: no wall-clock budget assertion carries the performance marker — all 73 run on the PR gate #13445

Finding

pytest.ini declares a performance marker, ci.yml deselects it on the PR gate
(-m "not integration and not slow and not distributed and not performance"), and
marker-tests.yml exists to run the excluded set on a schedule.

An AST sweep of every *_test.py / test_*.py for assertions comparing a measured
elapsed delta (a subtraction of two clock reads) against a numeric literal found:

TRUE elapsed-vs-literal assertions: 73  |  on PR gate: 73  |  marker-excluded: 0

Not one of the 73 carries an excluding marker. Every wall-clock budget in the repo
runs on the PR gate, where CI queue depth inflates wall-clock — the same shard was
measured at 82.49s quiet vs 132.55s loaded. Even the two files named for it,
system_benchmarks_performance_test.py and performance_benchmarks_test.py, are
unmarked.

The full file:line list is in #13444 (comment).

Why this is one issue, not 73

The decision is systemic, not per-test. Each of the 73 falls into one of three buckets:

1. A genuine performance budget → mark performance, so it runs in
   marker-tests.yml where timing is meaningful, not on a contended PR runner.
2. A concurrency proxy — the test wants "these ran in parallel", and time is a
   stand-in → rewrite as a rendezvous, the way file_locking_test already was
   (test: file_locking_test asserts a 0.04s wall-clock budget — fails 2 of 13 runs on unmodified base #13366 / 3726bc6): park every worker inside its critical section until all are
   inside. No clock, survives any load.
3. A timeout guard — the test wants "this was bounded / the timeout fired" → assert
   ordering and causality (the timeout path ran), not a duration.

Bucket 1 is a one-line change per test once the bucket is decided.

Scope

• Excludes system_benchmarks_performance_test.py (8 hits) — pre-existing failures
  tracked under ci: Python unit-test suite is not a required check — red tests merged to Dev_new_gui twice #13162.
• test(backend): replace wall-clock threshold assertions with the properties they guard (#13399) #13444 already fixed the 5 that were a single one-line judgement (JWT exp
  vs a literal tolerance), so those are not in the 73 remaining.

Related

• flaky: two backend tests assert wall-clock timing with no headroom and fail intermittently under xdist #13399, test: file_locking_test asserts a 0.04s wall-clock budget — fails 2 of 13 runs on unmodified base #13366 — the instances that prompted the sweep
• ci: Python unit-test suite is not a required check — red tests merged to Dev_new_gui twice #13162 — the standing-failure baseline these flappers inflate
• test(backend): replace wall-clock threshold assertions with the properties they guard (#13399) #13444 — the PR carrying the sweep