test_ltx2_video false-reds once on main under load: 1 of 96 cases failed on a first run, then 4 consecutive greens, and the case identity was not captured

[localai-bot]

Observed by the operator while rerunning the W0 gate on main at fdefb4529
(the merge of #1441), immediately after a clean ninja build.

What was measured

first run  EXIT=1  96 cases | 95 passed | 1 failed   4017 assertions | 4016 passed | 1 failed
rerun      EXIT=0  96 cases | 96 passed | 0 failed   4017 assertions | 4017 passed | 0 failed
repeat 1-3 EXIT=0  96/96, 4017/4017 each

Same binary, same tree, no rebuild between runs. test_capi (65/65),
test_ltx2_vae (43/43) and test_ltx2_tiling (10/10) all passed on their first
run in the same batch.

Box state during the failing run: load average 22-37 on 20 cores, / at 87-90%
full, with ~60 other Claude sessions sharing the machine.

Why this is filed rather than dismissed

The fifth review of #1441 predicted exactly this class and recorded it as F12:
the denoise coverage floor is 0.95 at 9 frames and 0.90 at 81, with roughly
six points of margin, and the per-boundary sampling cost is fixed while the leaf
shrinks. It is explicitly a false-RED risk, never a false pass. A contended
box narrowing a coverage ratio below its floor is the most likely explanation,
and this is the first observed occurrence rather than a predicted one.

An intermittent red on the instrument that gates the whole
LTX25-DEVICE-RESIDENCY campaign will be re-hit by W1, and a reader who sees it
once will not know whether it is the box or the code.

The instrumentation gap, which is the operator's

The run script captured only doctest's summary lines, so the failing case's
identity was not recorded and cannot be recovered. Four green runs establish
that it does not reproduce readily; they do not establish what failed. Anyone
reproducing this should keep the full output.

Note the line MESSAGE: captured 483 bytes of stderr on a FAILED file write
appears in the passing logs too, so it is an intentional case exercising a
write-failure path and is not the failure.

What would settle it

• Run the suite N times on an idle box and on a loaded one, keeping full output,
  and record which case fails and its measured coverage ratio against its floor.
• If it is the coverage floor, decide between widening the margin and making the
  assertion insensitive to sampling jitter. Do not simply loosen the floor: the
  same review established the floor is not a mute switch because the honest
  coverage distribution is bimodal.

Owner: row LTX25-DEVICE-RESIDENCY, .agents/specs/ltx25-device-residency.md
## Owed. Related: #1010, #1440, #1446.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [Claude Code]