Skip to content

latent_cond.py:38-39 is cited across 7 files for writes that live at :40-41 — orphaned anchor drift with no owning row #959

Description

@localai-bot

Surfaced by the #948 repair, which found it, declined to silently widen scope, and handed the ownership question up. Confirmed independently before filing.

The drift

At Lightricks/LTX-2 pinned fd4ded7f2d88d3da713abcdd4ad41ecc4a9314ca, packages/ltx-core/src/ltx_core/conditioning/types/latent_cond.py:

36    stop_token = start_token + tokens.shape[1]
37
38    latent_state = latent_state.clone()
39
40    latent_state.clean_latent[:, start_token:stop_token] = tokens
41    latent_state.denoise_mask[:, start_token:stop_token] = 1.0 - self.strength

7 files in this repository cite latent_cond.py:38-39 for the claim that "only clean_latent and denoise_mask are written" (positive control: 9 files cite latent_cond.py at all). Those writes are at :40-41. :38-39 is clone() and a blank line.

Why this is worth an issue rather than a drive-by fix

The claim itself is true — the reading is right, only the line numbers are wrong. So nothing is broken today, and every gate is green. That is exactly what makes it worth tracking: a reader who follows the citation lands on clone(), which is real code that does something else entirely, and is misled rather than alerted. A wrong-but-plausible anchor is worse than a dangling one.

The sites span specs, golden-generator scripts and other rows' tests. They predate the row that found this, and each belongs to a different owner — which is precisely why it was not fixed in flow. AGENTS.md's in-flow rule covers a bug you find in work you are doing; it does not license one row rewriting seven other rows' records.

What it is an instance of

#911 — 4772 line-carrying citations across 315 .agents/specs/*.md that no checker examines. This is that class with seven concrete, confirmed instances and a known-correct replacement, so it is a good first target for whatever #911 produces.

Two things the fix should get right, both learned the hard way this week:

  • Derive the needle from the claim, not from the cited span. A validator that reads the expected text out of the file it is checking is circular — one reported 27/27 FRESH while five anchors pointed at unrelated code.
  • Cover bare :NN continuations. A sweep that resets file context at line boundaries misses them, and they have slipped four separate audits. One was resolved against the nearest filename above, landing on a real-but-unrelated line in a different file — the same failure mode as this issue.

Suggested disposition

Either SHA-anchor the seven sites (latent_cond.py:40-41 @ fd4ded7f) or correct them live and add them to whatever sweep #911 lands. Not urgent; no behaviour depends on it.

Related: #911, #930, #948.

FOLLOWING_AGENTS_PROTOCOL

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions