[SkippySteve]

Summary

• The per-phase replay timers called QueryPerformanceCounter on the way into and out of every draw, dispatch, and descriptor build. The WebGPU renderer already measured ten of those reads as about 4% of the GPU thread, and switched its own stats clock to the timestamp counter.
• prof_now now reads the invariant TSC after one calibration against the performance counter. When that counter's tick is a whole number of nanoseconds (Linux reports 1 ns), the conversion is an integer multiply, so a reading past 2^53 keeps its low bits.
• The phase log is unchanged. If the calibration cannot get a span, the old counter path remains.

Test plan

• python -I -B tests/test_prof_clock.py passes
• The same test fails if the conversion always uses the double divide ((1<<53)+3 nanoseconds round away) and fails if prof_now no longer returns the timestamp-counter reading
• CI: "Check the replay clock skips the performance counter"

[NotRexed]

Thanks! This one changes timing/locking on the replay's hot path, so I'll measure it (fps and frame times in Kokiri, plus a few boots) before merging. The two clock-gating PRs (#306, #308) are merged.

[NotRexed]

Thanks! Measured together with the other replay PRs (replay 27 -> 24-25 ms a frame, fps 23 -> 26 in Kokiri) and applied by hand (conflicts only in the CI workflow; #310 and #315 needed one declaration moved), pushed in 40c3f85.