60 s 1080p WATERMARK SPLICE BENCHMARK
run: 60s-1080p-2026-10-23T10:35:45Z

PRIMARY: 100 sequential warm-cache C wm-splice runs of a 60 s / 1,800-frame / 2-IDR
single-moof 1080p segment. Process-per-request, full C post-probe, output write.
Constraint: taskset -c 0,1 + RLIMIT_AS=4GiB on 16-logical-CPU host.

STATS (60 s single-moof, n=100):
  n:                     100
  Wall min:              143.60 ms
  Wall median:           150.51 ms
  Wall mean:             151.57 ms
  Wall p95:              159.07 ms
  Wall p99:              161.36 ms
  Wall max:              163.24 ms
  Wall stdev:            5.12 ms
  Wall CV:               3.4%
  CPU mean:              146.38 ms
  CPU p95:               149.99 ms
  CPU p99:               150.72 ms
  Peak RSS:              129.8 MiB / 136.1 MB
  Throughput:            6.60 segments/s
  100 segments CPU:      14.64 CPU-s
  100 on 8 cores:        1.83 s
  100 on 2 cores:        7.32 s
  120 on 8 cores:        2.20 s
  120 on 2 cores:        8.78 s
  SLO p50:               FAIL
  SLO p99:               FAIL
  C probe recheck:       PASS (1800 samples, rc 0)
  Output:                22.35 MB (-0.23% vs input)
  SHA-256:               1ccdd36f722379d0d6a25d3d…

COMPARISON (fresh):
  shape                          n    median      p95       p99      CPU mean   peak RSS   SLO
  120 f · 1 IDR · 1 moof          100     10.25     10.73     11.12     10.07      10.9  PASS/PASS
  900 f · 1 IDR · 1 moof           50     68.00     72.06     73.01     67.36      65.8  FAIL/PASS
  1,800 f · 2 IDR · 1 moof        100    150.51    159.07    161.36    146.38     129.8  FAIL/FAIL
  1,800 f · 2 IDR · 2 moof         50    144.80    152.14    155.04    141.61     130.0  FAIL/FAIL

VERIFICATION (60 s single-moof):
  Python probe --expect-substituted: PASS (1800 samples)
  C wm-probe recheck: PASS (1800 samples, rc 0)
  stock ffmpeg strict decode: exit 0, 0 stderr bytes
  keyframes: sample 0, sample 900
  frame 0: non-region diffs=0, region changed=3864
  frame 1500: non-region diffs=0, region changed=3844
  sha256: 1ccdd36f722379d0d6a25d3dc4061da3f1f0dace6cc8d8f58280c56bb3ca795f

CAPACITY (arithmetic, current measured path):
  peak RSS 129.8 MiB = 136.1 MB decimal
  10 GiB / 129.8 MiB = 78.9 concurrent (zero overhead)
  10 GB / 136.1 MB = 73.5 concurrent (zero overhead)
  80% headroom = 59-63 concurrent; 100 does not fit
  one segment = 0.146 CPU-s; 100 = 14.6 CPU-s; 8 cores = 1.83 s
  to finish 100 in 140 ms needs ~105 logical cores

VERDICT: 60 s is technically valid and verified, but fails the per-segment
p50 30 ms and p99 150 ms SLO on the measured path. Keep 4 s for delivery;
6 s is the safe upper bound to measure next.

CAVEATS: synthetic testsrc2 content; process-per-request with full post-probe;
warm cache; PCM placeholder fill (not production text plate); ffmpeg software
decode only; no hardware decoder; no 1,000-concurrent load test; RSS includes
shared pages; 10GB/8-core arithmetic is ideal scaling, not a load test.

TAGS: #60s-1080p #1-moof #1800-samples #2-IDR #multi-GOP #probe-PASS #strict-decode-PASS #pixel-verified #SLO-p50-FAIL #SLO-p99-FAIL #RSS-129.8MiB #warm-cache #C-splice
ARTIFACTS: web60/index.html, web60/stats.json, web60/out-60s-1moof.m4s,
web60/out-60s-2moof.m4s, web60/proof-*.png
PLAYERS: web60/index.html#players — play-60s-1moof.mp4, play-30s.mp4,
play-4s.mp4, play-60s-original.mp4, zoom-60s-original.mp4, zoom-60s-spliced.mp4