The LTX-2.5 phase-table SUM gate's 95% floor sits inside its own measurement, and it is RED on main

[localai-bot]

ltx2 video: a render through the ABI emits a phase table that SUMS to wall (tests/vllm/multimodal/test_ltx2_video.cpp:2854) asserts CHECK(leaves >= 0.95 * wall). It is red on origin/main at 89261c955 in five of six runs on one x86 box, and the branch that found it is not the cause.

Measured

One box, one build directory, CMAKE_BUILD_TYPE empty as CI has it. For the main rows the four W0-live files were reverted so the binary IS main.

binary	run	wall (s)	leaves (s)	residue (s)	ratio	verdict
origin/main 89261c9	full suite	0.235589	0.222210	0.013379	94.32%	RED
origin/main	-tc alone	0.220766	0.207936	0.012830	94.19%	RED
origin/main	full suite	0.579684	0.551848	0.027836	95.20%	GREEN
origin/main	full suite	0.280686	0.263109	0.017577	93.74%	RED
origin/main	full suite	0.229418	0.216104	0.013314	94.20%	RED
origin/main	full suite	0.264471	0.250421	0.014050	94.69%	RED
W0-live merge	full suite	0.333903	0.313273	0.020630	93.82%	RED
W0-live merge, VLLM_RENDER_PROGRESS=0	full suite	0.246568	0.232059	0.014509	94.12%	RED
W0-live merge	-tc alone	0.282217	0.269225	0.012992	95.40%	GREEN
W0-live merge	full suite	0.279554	0.261876	0.017678	93.68%	RED
W0-live merge	full suite	0.352022	0.332084	0.019938	94.34%	RED
W0-live merge	full suite	0.356531	0.337360	0.019171	94.62%	RED
W0-live merge	full suite	0.393683	0.371605	0.022078	94.39%	RED

main's failing runs report 93 cases / 3321 assertions / 1 failed, exit 1. The W0-live merge reports 96 cases / 3555 assertions / 1 failed — the same single failure, with 3 cases and 234 assertions added and none of them failing.

VLLM_RENDER_PROGRESS=0 still fails, so the W0-live emitter is not the term. Reverting to main still fails, so nothing in that lane is.

It is not box contention, and one run proves it

The obvious reading is a flaky timing case losing to load. The single green main run refutes it: that run had wall=0.579684s, more than double every other run, because the box was loaded at the time. A slower render passes.

The reason is in the third column. The un-named residue is 4.80% to 6.32% of wall in all thirteen runs — 0.0128 s to 0.0278 s against a wall of 0.220 s to 0.580 s — and it grows more slowly than the wall it is divided by. So a longer render has a better ratio, and the floor is easier to clear the worse the box behaves.

What is actually wrong

The gate is a RATIO and its 95% floor sits inside the measurement's own range at the 64x64 / 9-frame fixture scale, where the whole instrumented render is 0.22 to 0.58 s. The case therefore decides by coin flip, and mostly red. The tolerance was argued for the 21.004 B render the instrument exists for, where the same residue would be a far smaller fraction of a wall measured in hours.

What would settle it

Naming the un-named time, or bounding unaccounted_seconds absolutely in addition to (not instead of) the ratio, so the assertion says the same thing at both scales. Either is a change to a gate's semantics and needs its own row, spec and red-first evidence per AGENTS.md "Changing the rules or a checker", so it is not folded into the W0-live recovery that found it.

Owner

Row LTX25-DEVICE-RESIDENCY, listed under ## Owed in ltx25-device-residency.md. Found while merging origin/main into row/LTX25-RESIDENCY-W0-LIVE after #1419 was auto-closed by its base branch being deleted.

[localai-bot]

Measured where the residue actually is, and it is not noise. Instrumenting the
leaf intervals of the same case on one render (wall = 0.343348 s,
unaccounted = 0.0123761 s, 19 leaf records):

• the first leaf record starts at 0.0113197 s, so 11.3 ms -- 91% of the whole
  residue -- is ONE contiguous un-named interval at the head of the load span,
  before load.dit opens;
• the interior gaps between named leaves total about 0.92 ms, the largest
  being 0.51 ms before load.video_vae and 0.33 ms before generate.setup;
• the tail past the last leaf is about 0.12 ms.

That interval is the ~164 lines between phase::Scope load_span("load")
(ltx2_video.cpp:766) and phase::Scope dit_phase("load.dit") (:930): device
resolution, the platform probe, file discovery and option parsing. It is a real,
nameable startup phase, not instrument scatter.

This closes off the tolerance route. An absolute unaccounted_seconds slack
big enough to stop this assertion flapping would have to exceed 12 ms, and would
therefore hide that phase -- a floor below the real value is a mute switch. So the
or bound the residue in seconds half of this issue's own proposal is refuted by
the measurement, and the FIRST half is the one left standing: name the time. A
load.open scope over that prologue drops the residue to about 1 ms of 343 ms,
0.3%, which makes the 0.95 floor honest rather than marginal.

Polarity, so nobody reads this as the CI red. On the two main runs that
carried the denoise coverage red this assertion was GREEN, at 95.52%
(96506970274) and 97.69% (96466616360) -- margins of 0.52 and 2.69 points. It
is the local boxes that see it fail, and it remains a thin margin in CI.

Measured while repairing #1494 in #1503,
which deliberately does NOT touch this assertion for the reason above. The full
probe and the sibling's repair are recorded in
ltx25-device-residency.md.