[dpadr]

Frame-time percentiles over a long window hide a slow stretch. In a 6-minute run, two 1000-frame stretches at 70–90 ms move p50 and p95 only a little, while the player sees about a minute of stutter. This adds what a long run needs, for any game:

• Per-run counts: every run's summary now has long_frames (frames of 1 s or more), holds (a GPU queue held by an event wait, the host's an event wait has held it line), and the median of each stretch of --stretch frames (default 1000), with worst_stretch. results.md shows long_frames, worst_stretch and holds.
• --verdict: judges each run CLEAN or BAD on those counts. A stretch is slow when its median is over --slow-factor (1.5) times the run's own frame p50, or over --slow-ms when given. The relative default works across games and frame targets.
• Report and exit code: results.md gets a verdict line, and the exit code is 1 when any run is BAD. Without --verdict, the exit code and the existing columns are unchanged.

bench.summary() keeps its signature (a new optional stretch argument), so scripts/regress.py's call is unchanged.

What it catches. I applied the branch's summary and judge to the frames.tsv and log of long runs of Ocarina of Time on an M2 Air (a 12,000-frame Kokiri walk, judged from frame 1000):

run	frame p50	verdict	why
plain main ccd834e + haze fix, run 1	48.8	BAD	2 of 11 stretches over 73 ms (worst 91 ms)
plain main ccd834e + haze fix, run 2	45.4	BAD	1 of 11 stretches over 68 ms (worst 77 ms)
a build with a signal-cut revert	49.5	BAD	97 frames of 1 s or more (worst 23.5 s), 8 event-wait holds
a build with fences stripped	33.2	BAD	2 of 11 stretches over 50 ms
the reference build, 4 runs	27.6–30.3	CLEAN

These agree with the absolute rule we used for those runs (no 1000-frame median over 40 ms); --slow-ms 40 gives that rule.

Tests (dm/tools/tests/test_bench.py):

• test_stretches_and_judge: stretches, the half-stretch tail, the factor and absolute limits, holds, a 1.2 s frame, no frames.
• test_verdict_cli: a fake host writes a slow stretch and a hold line → BAD in results.md and exit 1; a clean run → exit 0.
• test_bench.py: 9 tests OK. test_regress.py: 65 OK (Python 3.14; on 3.9 one test fails on main too, shutil.rmtree(onexc=)). test_docs.py OK.

Not included: the GPU watchdog line (#1550, not on main). When it lands, add its line to the holds count.

Regress: bench.py is a measuring tool. It changes no chain, host or rewriter, and regress.py's call into bench.summary is unchanged and tested, so no game run is needed.

🤖 Generated with Claude Code