fix(#1528): two comments name RetakePipeline as the caller that sets force_uncond_pass, and none of upstream's seven construction sites sets it

[localai-bot]

include/vllm/model_executor/models/ltx2_denoisers.h:144-145 and
src/vllm/model_executor/models/ltx2_denoisers.cpp:111-113 both told the reader
that RetakePipeline is the one upstream caller that enables
force_uncond_pass, citing retake.py:305-311. Both landed in daeff67f2
(row LTX25-GUIDED-VIDEO, #1092, PR #1102), and both are false at the pin.

Re-derived at Lightricks/LTX-2 fd4ded7f2d88d3da713abcdd4ad41ecc4a9314ca, the
revision .agents/specs/ltx25-guided-video.md:6 and
.agents/specs/ltx25-retake.md:12 both pin, read at the local checkout
/home/mudler/_git/LTX-2 whose git rev-parse HEAD is that SHA and whose
working tree is clean.

git grep -n force_uncond_pass over the whole upstream tree, not only *.py,
returns 11 lines and not one is an assignment by a caller. The filing counted 10
because it restricted the grep to *.py, so the two counts agree and name
different populations.
utils/denoisers.py:74 declares the _guided_denoise keyword parameter,
:102-103 are the two reads, :267,273,297 are GuidedDenoiser storing and
forwarding it, and :313,319,357 are FactoryGuidedDenoiser doing the same.
utils/samplers.py:692 is the substring force_uncond_pass=True inside a
raised error message. packages/ltx-pipelines/CLAUDE.md:76 documents it.

The filing checked retake.py alone. git grep -n 'GuidedDenoiser(' finds
seven construction sites, three GuidedDenoiser and four
FactoryGuidedDenoiser: a2vid_two_stage.py:230,
keyframe_interpolation.py:232, retake.py:305, t2a_one_stage.py:155,
ti2vid_one_stage.py:221, ti2vid_two_stages.py:248 and
ti2vid_two_stages_hq.py:271. None passes the keyword, so all seven take the
force_uncond_pass: bool = False default. The corrected comments state that
stronger fact rather than the retake-only one.

The flag is a CFG++ affordance, per CLAUDE.md:76: run the uncond pass even
when cfg_scale is 1.0, because CFG++ still needs the uncond prediction for the
ordinary differential equation derivative. Upstream ships the plumbing and turns
it on nowhere.

Two smaller errors rode along and are corrected in the same text. The header
cited the declaration as :76, which is a docstring line; the declaration is
:74. Both comments cited retake.py:305-311, whose last line is blank; the
construction expression is :305-310.

BEHAVIOR DOES NOT CHANGE, AND THE CODE WAS NOT THE DEFECT.
Ltx2GuidedDenoiseInputs::force_uncond_pass already defaults to false,
grep -rn force_uncond_pass src/ include/ tests/ examples/ returns five lines,
the two comments, the two reads and that default member initializer, and no
assignment anywhere, and the condition mirrors denoisers.py:102-103 operand for
operand. The struct has one construction site in the tree,
src/vllm/multimodal/ltx2_video.cpp:4306, which default-initializes and then
assigns ten fields by name, none of them this one, so no positional or designated
initializer can reach the field either. include/vllm.h names neither the field
nor the struct, so no external caller can set it across the application binary
interface. That is exactly upstream, where nothing enables it. The scope stayed
at comment text, and no wider repair was taken.

A comment-only change cannot be red-first tested, and this change does not
pretend otherwise. Two things stand in for that. The upstream re-derivation
above is cited by file:line at the pin, and a fresh reviewer re-derived every
anchor from the upstream checkout rather than reading it from this message. And
the comment-only claim is not asserted, it is measured: stripping comments from
each file at the base and at the head with gcc -fpreprocessed -dD -E -P gives
byte-identical output, e59efade7344a9bc... for the header and
c6188d0f6682b3ad... for the translation unit, so the token stream the compiler
sees did not move.

THE ANCHOR WAS LOAD-BEARING, which is why this is a bug and not a typo. A reader
who trusts it concludes that our retake path is missing a flag upstream sets,
and sets it. force_uncond_pass adds an uncond pass for every present modality,
so that repair would double retake's DiT forward count from 8 to 16 and change
its output. The comment invited the exact divergence it appeared to describe.

Two specs were checked and deliberately left alone.
.agents/specs/ltx25-retake.md never mentions the field.
.agents/specs/ltx25-guided-video.md:198 states the condition, uncond when
either guider asks or force_uncond_pass, without claiming a caller, which is
true at the pin.

WHY THIS TOOK TWO ATTEMPTS, because the issue thread still says it is blocked.
The first attempt wrote the same correction on row/LTX25-GUIDED-VIDEO-1528 at
dafde0dc4 and deliberately opened no pull request, because
scripts/check-doc-checkpoint.py classified any byte changed under
include/vllm/ as a user-usage change and demanded a docs/USAGE.md edit that a
comment fix had nothing true to write. That checker no longer exists. 1db7e59cf
(docs: retire shared status and split benchmark details, #1714) deleted it
along with docs/STATUS.md, so the obstacle the thread records is gone rather
than waived, and the full preflight is green here with no skip. The earlier
branch is superseded by this one.

Closes #1528.

FOLLOWING_AGENTS_PROTOCOL

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