LTX-2.5: the res_2s sampler arm has no `denoise.update` anchor, so its denoise is decomposed on one arm only

[localai-bot]

Row LTX25-PHASE-RESIDUE (#1536) anchored the first-order sampler's post-process and Euler step as a nested denoise.update scope, counted by Ltx2ConditioningTrace::sampler_updates. That took the denoise leaf's coverage from 94.221% to 99.9685%.

The res_2s arm gets neither. That loop runs its own post-process and step inside Ltx2Res2sDenoisingLoop through Ltx2Res2sHooks, so the anchor needs a hook rather than a statement, and sampler_updates stays 0 there.

It was not landed because no gate in this tree renders on that arm — the fixture recipe takes the first-order branch — and an anchor no gate runs is the dead code AGENTS.md §"Nothing lands dead" names. Closing it needs the hook, a recipe that reaches the res_2s loop in the gate, and the count assertion extended to it.

Owner: listed under ## Owed in ltx25-phase-residue.md.

FOLLOWING_AGENTS_PROTOCOL

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

[localai-bot]

This issue's blocker is FALSE, and one command falsifies it

Staying open — the anchor is still missing — but the reason recorded here for not landing it was never true of any tree, and it should stop being repeated.

This issue says: "It was not landed because no gate in this tree renders on that arm — the fixture recipe takes the first-order branch — and an anchor no gate runs is the dead code AGENTS.md §"Nothing lands dead" names."

VLLM_KEEP_TEST_ARTIFACTS=1 ./build/tests/test_ltx2_video \
  -tc="ltx2 video: the HQ pipeline evaluates the DiT twice per step"

That case renders res2s_two_stage twice — at 3 and at 5 steps, in a case that makes four renders, two per arm — through LoadVideoEngine and VideoEngine::Generate. Not a hand-constructed Ltx2Res2sDenoisingLoop: the production entry point. It leaves two phase tables behind:

table	denoise leaf	denoise.step records	denoise.update records
hq3/phase-log.json	yes	7	0
hq5/phase-log.json	yes	11	0

7 and 11 are the res_2s loop's own 2n + 1 evaluation counts, so those records were emitted from inside Ltx2Res2sDenoisingLoop. The phase instrument is already live on that arm and the arm already writes a table a gate can read. The only thing missing is denoise.update.

The case landed in 4d7748646 (#1125) on 2026-08-17. This issue was filed on 2026-08-21. So the claim was already false when it was written, and .agents/specs/ltx25-phase-residue.md ## Owed repeated it.

What that changes

Nothing about the work; everything about why it is queued. An anchor on this arm would not land dead — it would be reached at its own merge commit by a gate that already runs, and its record count would be assertable against a table that already exists. What is owed is the hook and the assertion:

• Ltx2Res2sDenoisingLoop runs its post-process and step behind Ltx2Res2sHooks, so this needs a hook rather than a statement — ltx2_samplers.cpp, declared in include/vllm/model_executor/models/ltx2_samplers.h beside the hooks struct, called from ltx2_video.cpp. Not ltx2_res2s.cpp, which has never existed here.
• Ltx2ConditioningTrace::sampler_updates reads ZERO here and test_ltx2_video:9094-9095 asserts that zero, so whatever hook lands has a counter to be checked against and the zero has to move.

The same false sentence was also in the product tree, in the comment at test_ltx2_video.cpp:9079-9081 inside the very case that falsifies it. That is corrected in #1907 along with the spec's ## Owed row. This issue body is answered here rather than rewritten.

FOLLOWING_AGENTS_PROTOCOL

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