libkernel: wait tracing reports inactive sites with NaN averages #7155

Resolved — 4 November2026: #7156 merged as 5924ecf2dfa195608920b8beb90edd5a5d6a5b40 and closed this issue. The prior reproducer/evidence remains below. Current maintenance report; no new title-performance claim follows from the merge.

Current proposed correction -31October2026

#7156 remains open,Depends on:none. Refreshed/published head70f7250c3b95 incorporates main5857e2e98fda8af3a2ee61676bd40c8e97a8df1e with the identical one-line correction and preserved history. Main's newer TSC/timed-wait changes are retained.

Fresh13positive CPU processes and one intended control failure:literal main five shared synchronization checks,candidate the same five plus three trace passes;main-production plus identical trace fixture still emits inactivex0 rows,while candidate removes them and preserves active zero-durationx1. Maximum6.240s,no skips/timeouts;full native/relinker/real selected-fixture builds and conventions pass. Complete source/control evidence,logs,hashes and one inherited analyzer advisory.

Diagnostic correction only;no fresh title,performance,GPU or other-platform claim. Earlier evidence keeps its actual pins below. Keep open pending merge. AI-assisted:yes(Codex).

What

While investigating Sonic Unleashed Recompiled (PPSA99902) remaining at initial NOW LOADING (#3036), the existing APS5_TRACE_WAITS=1 diagnostic repeatedly reports inactive sites:

[waits] | cond <site> x0 0.0s (avg nan ms, max 0 ms, 0 timeouts, tid 0)

<site> replaces the caller offset. After each report, KernelTraceWait_nid_postfix resets counters but retains the sites. The next report copies zero-call sites into its ranked list and divides their zero duration by zero calls. They can also occupy the eight displayed slots.

Expected: omit sites with no calls in the reporting window; retain active sites even when their measured duration is zero. Current main 47a266cf975f00f7fba94013467d03bfa5e355e9 has this code.

Reproduction and validation

Test-only9b9eb47a31 leaves main production unchanged. Build guest_wait_trace_tests and run ctest --test-dir build -V -R '^guest_wait_trace$'. It creates eight sites, crosses two real three-second windows, then emits an active zero-duration wait. CTest rejects an inactive x0 row and requires the active zero-duration row. Expected failure3/3 on both Windows/GCC15.2r7 and Ubuntu24.04/GCC13.3.

Matched NVIDIA title profiles: original 120.790s has 39 reports / 38 NaN rows; corrected 120.872s has 39 reports / zero inactive or NaN rows. Only libkernel.prx changes; executable, guest module,126 other PRXs and dependencies match. Both reach their120s harness limit; original inputs/configuration remain unchanged. This is diagnostic validation, not a whole-main startup, performance or title-unblock claim. Wait tracing reports completed waits and does not exclude a blocked thread.

Proposed fix

Focused fork947e02f32b filters the temporary reporting vector before sorting/limiting it. One production line; regression and related synchronization controls pass. No change to wait/wake semantics. AI-assisted: yes, Codex; author responsible.