LTX-2.5: 22 upstream citations name latent_cond.py and schedulers.py two and one lines short, while eight other citations in the same files already carry the corrected anchor

[localai-bot]

Two LTX-2.5 upstream anchors are cited short across this tree, and the tree
already carries the corrected form beside the stale one, so a reader who greps
either string finds both and has nothing that says which is right.

Re-derived at the LTX-2 pin fd4ded7f, reading the pinned files rather than
inheriting the citation:

• packages/ltx-core/src/ltx_core/conditioning/types/latent_cond.py
:38 is latent_state = latent_state.clone() and :39 is blank. The two
  writes the citations mean -- clean_latent[...] = tokens and
  denoise_mask[...] = 1.0 - self.strength -- are :40 and :41.
  :38-39 is wrong by two; :40-41 is right.
• packages/ltx-core/src/ltx_core/components/schedulers.py
:31 is the return annotation ) -> torch.FloatTensor:. The read the
  citations mean, tokens = math.prod(latent.shape[2:]) if latent is not None else default_number_of_tokens, is :32.
  :31 is wrong by one; :32 is right.

Where the stale form is, measured at LTX25-KEYFRAME-INTERP's head

latent_cond.py:38-39, 11 citations:

• src/vllm/multimodal/ltx2_video.cpp:2208, :3174, :3330, :3392
• include/vllm/model_executor/models/ltx2_pipeline.h:754
• include/vllm/model_executor/models/ltx2_conditioning.h:26
• tests/vllm/multimodal/test_ltx2_video.cpp:8138, :8217
• tests/vllm/multimodal/test_ltx2_image_cond.cpp:747
• tests/vllm/models/test_ltx2_vae.cpp:2666
• .agents/specs/ltx-2-5.md:499

schedulers.py:31, 11 citations:

• src/vllm/model_executor/models/ltx2_pipeline.cpp:1667, :1851
• src/vllm/multimodal/ltx2_video.cpp:3618
• include/vllm/model_executor/models/ltx2_pipeline.h:662, :971
• tests/vllm/multimodal/test_ltx2_video.cpp:7890, :8485
• tests/vllm/models/test_ltx2_pipeline.cpp:3559, :3824
• .agents/specs/ltx25-ti2vid-recipe.md:59, :150

The correct form is ALREADY in the tree, which is what makes this a sweep

Eight citations landed with the corrected anchor and nothing marks them as the
exception:

• include/vllm/multimodal/ltx2_video.h:625, include/vllm/model_executor/models/ltx2_samplers.h:32
• src/vllm/model_executor/models/ltx2_pipeline.cpp:1364, src/vllm/multimodal/ltx2_video.cpp:3017, :3600
• .agents/specs/ltx25-image-conditioning.md:168, .agents/specs/ltx25-res2s-loop.md:83, .agents/specs/ltx25-t2a-one-stage.md:229

So src/vllm/multimodal/ltx2_video.cpp cites schedulers.py:31 at :3618 and
schedulers.py:32 at :3017 and :3600, and
include/vllm/model_executor/models/ltx2_pipeline.h cites schedulers.py:31
twice while its sibling header cites :32. The disagreement is inside single
files and inside single directories, not between old code and new.

Why it is a row of its own and not an in-flow fix

A PARTIAL correction is strictly worse than none. PR #1209's review-repair
commit corrected seven of these citations, the ones its own new lines restated,
and left the rest. That produced two effects and both are why the corrections
were reverted out of it rather than extended:

1. src/vllm/multimodal/ltx2_video.cpp then read latent_cond.py:38-39 at
   :2208 and :3174 and latent_cond.py:40-41 a hundred lines later, with
   nothing in the file recording which one a reader should believe. One file
   stating both forms is a worse record than one file stating one wrong form,
   because the second is correctable by a grep and the first is not.
2. One of the seven lived in include/vllm/, which is a
   USER_USAGE_PREFIXES path in scripts/check-doc-checkpoint.py:99. That
   check is a pure path match with no content analysis, so a comment-only anchor
   correction in a public header demands a docs/USAGE.md edit that the change
   does not owe. The honest resolution is to not touch the header for a
   citation, not to manufacture a usage edit.

Both consequences are properties of doing this piecemeal. Done as one sweep over
all 22 citations in one commit, neither arises: no file is left mixed, and the
include/vllm/ edits arrive together with a reviewer who can see they are
comment-only.

Scope

Correct all 22 stale citations to latent_cond.py:40-41 and schedulers.py:32.
Comment and record text only; no behavioral change, so the gate is the existing
ctest suite unchanged plus a re-derivation of both anchors against the pin
recorded in the commit body. Confirm no citation is left in either stale form
and that the :41-only citations at ltx2_video.cpp:2373 and :2377, which
name the denoise_mask write alone, are correct as they stand.

Filed by the review-repair flow of #1209 (LTX25-KEYFRAME-INTERP), which
deliberately did NOT fix it in flow: the sweep touches five files that row does
not otherwise open, it needs its own reviewer for a 22-site mechanical edit, and
attempting it inside a landing fix is what produced the mixed state above.
Listed under ## Owed in .agents/specs/ltx25-keyframe-interp.md.

[localai-bot]

Two more anchors for this sweep, found by the fresh review of #1252 (merged as 7b9e207b1)

Both are LOW and neither blocked that PR, but they belong here rather than in a follow-up row.

1. ltx2_video.cpp:4479 does not exist — the third Ltx2EncodePromptToConditioning call site is at :4638.

Cited in src/vllm/model_executor/models/ltx2_text_encoder.cpp:445 alongside the two correct anchors :2085 and :2799, and repeated in .agents/specs/ltx25-text-linear-seam.md:42, :491 and .agents/specs/ltx25-decode-speed.md:1438.

This is not drift. The reviewer checked the spec's own base 839d24313 and the call site was at :4638 there too, so :4479 was never correct at any commit — it is a transcription error, not a line that moved. That distinction matters for this row: a sweep that only rebases stale anchors would not catch it, because there is no earlier commit at which it was right.

2. .agents/benchmark-record.md:23028 records base b626be75a; the real base is 839d24313, seven commits later. Immaterial to reproducibility — ltx2_text_encoder.cpp is byte-identical at both — but a recipe whose stated SHA is wrong cannot be re-run by SHA, which is the point of recording it.

Why this strengthens the case for the row

The reviewer's note is worth keeping: no checker catches either of these. The first is a file:line in a comment, the second a SHA in a record. This row's value is precisely that the tree has no automated defence here, and every instance found by hand is one a reader would otherwise have trusted.

Running count for the sweep, with these: the previously enumerated 11 stale latent_cond.py:38-39 and 11 stale schedulers.py:31, plus one never-correct ltx2_video.cpp:4479 in four places, plus one wrong base SHA.