Fix two silent no-ops in LTXLikenessAnchor / LTXLatentAnchorAware

[jjdejong]

Two bugs that make the anchors silently do nothing. Both fail without raising, without logging, and while still returning a valid MODEL — so the node looks like it is working. Found while debugging LTX-AV i2v runs on ComfyUI.

Two commits, reviewable independently.

---

1. The video latent is located by guessing tensor rank

Both anchors discover their grid dimensions in the backbone pre-hook by scanning forward()'s args/kwargs for "the first 5D tensor", commented as "typically the latent input".

It never is. LTX-AV passes x as a video+audio container, not a Tensor — separate_audio_and_video_latents indexes it as x[0]/x[1], and recombine_audio_and_video_latents returns a plain [vx, ax] list — so isinstance(..., torch.Tensor) skips it and the scan never descends.

What it actually finds is denoise_mask, reshaped to (B, 1, F, H, W) by model_base. Its F/H/W coincidentally equal the latent's, so the dims come out right by accident:

graph	denoise_mask	result
i2v / inpainting	present, 5D	works — via the mask, not the latent
no mask supplied	None	nothing captured

In the second case LTXLikenessAnchor's latent_frame_0 mode returns its input unmodified at every block, and LTXLatentAnchorAware's blend never runs.

Fix: _extract_video_latent() resolves the x argument (keyword or first positional), unwraps a .tensors wrapper or plain list to reach the video element, and only then falls back to a rank-5 scan that excludes denoise_mask/concat_mask.

Verified against the real argument shapes:

list [video, audio] + denoise_mask  -> (1, 128, 31, 17, 17)   # was returning the 1-channel mask
NestedTensor-style wrapper          -> (1, 128, 31, 17, 17)
bare 5D tensor                      -> (1, 128, 31, 17, 17)
x as keyword                        -> (1, 128, 31, 17, 17)
denoise_mask alone                  -> None

The channel count is the tell: 128 is the video latent, 1 was the mask. Confirmed live — LTXLatentAnchorAware now logs captured video latent: (1, 128, 31, 17, 17).

---

2. The backbone pre-hook is never re-registered

if not getattr(backbone, HOOK_ATTR_BACKBONE, False):
    backbone.register_forward_pre_hook(backbone_pre_hook, with_kwargs=True)
    setattr(backbone, HOOK_ATTR_BACKBONE, True)

That sentinel lives on the backbone module, which model.clone() shares by reference. From the second apply onward in a single ComfyUI process the gate is already set, so the new pre-hook is never attached and the first apply's closure stays live — writing captured_latent_shape and current_sigma into a state dict nothing else can read.

For every apply after the first, for the process lifetime:

• latent_frame_0 silently returns unmodified output — captured_latent_shape stays None, so the hook bails at every block.
• skip_when_sigma_above stops working in all modes, including guide — current_sigma comes from the same dead closure.

Only the first prompt after a restart ever worked, which makes it easy to mistake for a functioning node.

The attn1 hooks in this same node already remove-and-re-register correctly, and LTXLatentAnchorAware does so for both its hook types — this path was the outlier.

Fix: track the handle, remove any prior registration, re-register unconditionally. The bypass path now detaches the pre-hook too, rather than only dropping the sentinel — which previously left it attached and firing.

Verified by applying three times against a shared module:

apply 1: state='run1' live_hooks=1
apply 2: state='run2' live_hooks=1
apply 3: state='run3' live_hooks=1

Confirmed live: before the fix a second apply logged video latent shape was never captured; after it, the same second apply reports HOOK ACTIVE.

---

Tested on LTX-2.3 22B i2v, ComfyUI on ROCm/gfx1151. No API, node-name or default-value changes; behaviour is unchanged on graphs that were already working (i2v with a mask, first apply after restart).