Found by a fresh review of #786, which PASSED — these are non-blocking and were explicitly recommended as a follow-up rather than a round trip.
Every substantive claim at these sites is true at the pin (Lightricks/LTX-2 @ fd4ded7f). Only the line numbers are wrong, and nothing gates on them.
The mis-cites
| site |
cites |
correct |
src/vllm/multimodal/ltx2_video.cpp:1243, :1507 |
latent_cond.py:38-39 |
:40-41 — :38 is latent_state = latent_state.clone(), :39 is blank |
src/vllm/model_executor/models/ltx2_image_preprocess.cpp:145 |
resize.py:71 |
:68 — :71 is blank |
src/vllm/multimodal/ltx2_video.cpp:1202 |
distilled.py:212 feeding combined_image_conditionings |
:212 is resolve_crf; the feeding call is :252-262 / :282-292 |
src/vllm/model_executor/models/ltx2_video_vae_encoder_load.cpp:260 |
video_vae.py:66 for eps=1e-6 |
:67 — :66 is out_channels=out_channels, |
ltx2_video.cpp:1201 |
blocks.py:936-993 |
class ends at :991 |
ltx2_image_preprocess.cpp:128 |
resize.py:60-63 |
new_w is at :64 |
ltx2_video.cpp disagrees with itself: it cites latent_cond.py:41 correctly at :1306 and :1310, and :38-39 at :1243 and :1507.
Pre-existing, same class — not #786's, verified absent from its added lines
ltx2_video_vae_encoder.h:17 — attention.py:11-69 for AttnBlock3D, which starts at :26ltx2_video_vae_encoder.h:21 — model_configurator.py:37-70, where :70 is blank (that row's own new TU cites :37-69 correctly)ltx2_video.cpp:1363 — distilled.py:214-215 → :217-218ltx2_video.cpp:1612 — distilled.py:177-183 → :178-185, anchor :182
Why it matters more than tidiness
A reader checks resize.py:71, finds a blank line, and concludes the crop is unported or the pin has moved — the exact re-derivation these citations exist to prevent. AGENTS.md requires citing the file:line ported from precisely so a later reader can verify without re-deriving; an anchor that lands on nothing costs more than no anchor at all, because it looks checkable.
This is also the third consecutive campaign finding of this shape: recorded anchors went stale within a single PR, a repair pass claimed nine anchors "each re-verified against the pin" while this set survived it, and a checker's own line references drifted. The pattern is that anchors are written once and verified never.
What closing this looks like
Correct the nine occurrences, re-derived at the pin rather than shifted by the offsets above — and use a positive control in the same command, since a path filter is an absence claim too.
Then consider whether the class can be gated at all. A checker that resolves an upstream file:line citation would need the upstream checkout, which CI does not have (#673 records that LTX2_CHECKPOINT_ROOT is in no workflow either). A cheaper 80% guard: assert no cited line is blank in the pinned checkout, runnable by anyone with /home/mudler/_git/LTX-2 and enough to catch two of these six.
Related: #786, #673.
Found by a fresh review of #786, which PASSED — these are non-blocking and were explicitly recommended as a follow-up rather than a round trip.
Every substantive claim at these sites is true at the pin (
Lightricks/LTX-2 @ fd4ded7f). Only the line numbers are wrong, and nothing gates on them.The mis-cites
src/vllm/multimodal/ltx2_video.cpp:1243,:1507latent_cond.py:38-39:40-41—:38islatent_state = latent_state.clone(),:39is blanksrc/vllm/model_executor/models/ltx2_image_preprocess.cpp:145resize.py:71:68—:71is blanksrc/vllm/multimodal/ltx2_video.cpp:1202distilled.py:212feedingcombined_image_conditionings:212isresolve_crf; the feeding call is:252-262/:282-292src/vllm/model_executor/models/ltx2_video_vae_encoder_load.cpp:260video_vae.py:66foreps=1e-6:67—:66isout_channels=out_channels,ltx2_video.cpp:1201blocks.py:936-993:991ltx2_image_preprocess.cpp:128resize.py:60-63new_wis at:64ltx2_video.cppdisagrees with itself: it citeslatent_cond.py:41correctly at:1306and:1310, and:38-39at:1243and:1507.Pre-existing, same class — not #786's, verified absent from its added lines
ltx2_video_vae_encoder.h:17—attention.py:11-69forAttnBlock3D, which starts at:26ltx2_video_vae_encoder.h:21—model_configurator.py:37-70, where:70is blank (that row's own new TU cites:37-69correctly)ltx2_video.cpp:1363—distilled.py:214-215→:217-218ltx2_video.cpp:1612—distilled.py:177-183→:178-185, anchor:182Why it matters more than tidiness
A reader checks
resize.py:71, finds a blank line, and concludes the crop is unported or the pin has moved — the exact re-derivation these citations exist to prevent. AGENTS.md requires citing thefile:lineported from precisely so a later reader can verify without re-deriving; an anchor that lands on nothing costs more than no anchor at all, because it looks checkable.This is also the third consecutive campaign finding of this shape: recorded anchors went stale within a single PR, a repair pass claimed nine anchors "each re-verified against the pin" while this set survived it, and a checker's own line references drifted. The pattern is that anchors are written once and verified never.
What closing this looks like
Correct the nine occurrences, re-derived at the pin rather than shifted by the offsets above — and use a positive control in the same command, since a path filter is an absence claim too.
Then consider whether the class can be gated at all. A checker that resolves an upstream
file:linecitation would need the upstream checkout, which CI does not have (#673 records thatLTX2_CHECKPOINT_ROOTis in no workflow either). A cheaper 80% guard: assert no cited line is blank in the pinned checkout, runnable by anyone with/home/mudler/_git/LTX-2and enough to catch two of these six.Related: #786, #673.