LTX-2.5 VAE: upstream's bf16 noise draw is a DIFFERENT sequence, and mirroring it invalidates every captured render digest

Row: -

Owed by .agents/specs/ltx25-a24-video-vae-bf16.md (A24 wave 3), which is
drafted but not yet landed; this issue is filed ahead of it so the measurement is
not lost with the draft. Found while specifying wave 3 of gap A24.

This is a product decision, not a mirror question. It is filed for the
developer, and no row implements it until it is answered.

The measurement

Upstream's video VAE draws its noise from a different sequence at bfloat16
than at float32. Measured by executing Lightricks/LTX-2 @ fd4ded7f
(conv_video_decoder.py:288-294 into resnet.py:115):

• 1151 of 1152 drawn values differ between the two dtypes.
• Only 9 of 64 bf16 draws appear anywhere in the first 256 f32 draws.

So it is not a rounding of the f32 stream. It is a different stream.

Why it is a decision

Gap A24 mirrors upstream's single model dtype: distilled.py:109 sets
self.dtype = torch.bfloat16 and hands it to every component. Waves 1 and 2
(the text tower, 8e582a5f9, and the connector, #2753) mirrored it exactly,
because AGENTS.md says to mirror vLLM wherever it defines behaviour and never to
ask how a mirrored feature must behave.

Wave 3 cannot follow that rule silently. Mirroring the bf16 noise draw would
change every captured render digest in this repository — the goldens, the
committed frame digests, and the adherence-campaign renders that
.agents/specs/ltx25-adherence-rescore.md established are bit-deterministic and
whose control arm is byte-identical across 454 commits.

That is a scale of record invalidation a row should not choose on its own.

What a decision needs to answer

• (a) Mirror the bf16 draw, and re-capture every affected digest in the same
  change, recording the invalidation explicitly.
• (b) Keep the f32 noise stream as a named, recorded divergence from
  upstream, with the reason and the measurement above, so the digests stay
  comparable across the campaign.
• (c) Mirror it but gate the VAE arm on a fresh set of digests captured
  after the switch, leaving pre-existing renders anchored to the old stream and
  marked as such.

Unverified

Whether the shipped bench geometries produce more than one temporal group, which
decides whether the tiled-blend path (a separate 31%-of-values effect measured in
the same sweep) is reachable in practice. It needs a real-weights render and a
GPU lease; none was taken.

Read .agents/specs/ltx25-a24-video-vae-bf16.md, distilled.py:109, and the upstream conv_video_decoder.py:288-294 and resnet.py:115 measurements. The work is blocked on choosing and recording one of the three noise-stream policies; if mirroring is selected, affected render digests and the adherence-rescore records must be addressed, while the tiled-blend path remains unverified without a real-weights GPU render.