LTX-2.5: `denoise.step` and `denoise.update` are open to a seconds transfer that no assertion in the file can see

[localai-bot]

Found by the fresh review of #1536. denoise is now a two-part leaf and neither part carries a part_min_coverage floor.

The transfer. Leave denoise.step open across the post-process and the Euler step, and emit denoise.update as an empty scope after it. That preserves the alternation, both record counts, containment, non-overlap, exclusivity, assertion (1c) and assertion (2), and moves 100% of the leaf's decomposed seconds onto one name. It is R1b, the transfer shape test_ltx2_video.cpp already documents and closes for decode.audio with the vocoder's 0.50 floor.

Why (1b') does not close it. (1b') compares start_seconds only. It never reads a duration, so an empty sibling in the right position satisfies it.

Why no floor closes it either. The honest share of denoise.update runs 0.45% to 11.15% of the leaf across four boxes (the population 6b48edb2c recorded as 99.55% down to 88.85% coverage on an unanchored denoise), and a transfer puts it at ~0%. The honest distribution's bottom touches the defective value, so a floor would sit inside its own scatter — the defect #1439, #1494 and #1536 are all about.

Closing it needs an anchor INSIDE the callee, which is the third row of the anchor table in ltx25-device-residency.md ### Owed out of W0 for all six anchors.

FOLLOWING_AGENTS_PROTOCOL

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

[localai-bot]

Still open, and re-verified by running the mutation rather than by reading the assertions

#1888 landed denoise.update, which made this issue's transfer real instead of hypothetical. It does not close it. Confirmed on ced0ab639 by running the transfer.

Mutation T1, the empty-sibling form of this issue's R1b. denoise.update collapsed onto its own Close() at the same statement, line-count preserving, so the record is still emitted once per step and still sits between two denoise.step records while naming ~0 s:

::vllm::multimodal::phase::Scope update_phase("denoise.update"); update_phase.Close();

[doctest] assertions: 796 | 796 passed | 0 failed |    x5

Green 5 runs of 5 on the focused case, and reproduced independently at 3 of 3 by a fresh review that checked the mutated binary differed by hash from the unmutated one first. The anchor was asserted unique before it was applied, the application verified by hash, the build exit code read, and the tree restored byte-for-byte and re-hashed.

Everything this issue predicted survives it: the alternation, both record counts, containment, non-overlap, exclusivity, (1c) and (2). (2b) executes zero CHECKs on this leaf, because part_min_coverage is {0.0, 0.0} and the loop does if (c.part_min_coverage[i] <= 0.0) continue; — so decode.audio is still the only leaf (2b) asserts anything about.

What is still owed

An anchor INSIDE the callee, as this issue said. Not a floor: the honest share and the transfer overlap, and this box makes that worse rather than better — denoise.update's honest share measured 0.0006% and 0.0048% of the denoise leaf on two renders at loadavg 66, below the 0.45%–11.15% band recorded across four boxes, because contention dilates the DiT evaluations under denoise.step and not the sampler's own update. A floor that reds a transfer would red those renders.

FOLLOWING_AGENTS_PROTOCOL

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