A18: reference-audio delivery is a product decision, not a port — its only upstream caller is DubIt, which this tree does not ship

Row: LTX25-REF-AUDIO-DELIVERY

Derivation: .agents/specs/ltx25-ref-audio-delivery.md, landed with this issue.
Split from the A18 row of .agents/specs/ltx25-completion-scope.md §4, whose
sizing this issue contradicts.

A18 of the LTX-2.5 completion scope: reference-AUDIO conditioning delivery.
Refused at src/vllm/multimodal/ltx2_video.cpp:2934-2947.

This is a decision, not a port

The scope document sizes A18 M, "the op is ported and undriven", and orders
it at 10 behind A19. Reading the item's consumer rather than its operator
contradicts that.

At the pinned Lightricks/LTX-2 @ fd4ded7f, over packages/:

git grep -n 'AudioConditionByReferenceLatent(' -- packages/
  packages/ltx-pipelines/src/ltx_pipelines/dubit.py:272

One caller. The only inference-side delivery of this conditioning item lives
inside DubItPipeline — A3, which this tree does not ship
(ltx2_pipeline.h:876). Every other upstream pipeline hands its audio
ModalitySpec no conditionings at all: a2vid_two_stage.py:251-256 and
:291-296 freeze on an initial_latent, which is the arm this tree already
serves.

The refusal names the wrong mechanism as its own blocker

The message cites ltx-pipelines/utils/helpers.py:264-269 as what it needs.
That is audio_latent_from_file, whose only inference caller is
retake.py:250 and which ends in _conform_latent_length — it produces
retake's initial latent and never touches the conditioning item. This message
had already been edited twice to retract earlier false reasons.

There is no geometry to mirror

Two upstream consumers disagree, and neither is a default:

So "mirror vLLM" cannot arbitrate it.

Local cost the sizing did not count

StreamState (ltx2_video.cpp:181-192) is a different type from
Ltx2LatentState (ltx2_conditioning.h:94+); bridging them is already called
"its own change" by ltx25-a2v-audio-input.md's ## Owed. And there is no
audio-side trim at all: target_tokens is video-only (:3753),
Ltx2ClearConditioning runs on video alone (:5045), the invariant at
:5071-5072 checks video only, and :5090 unpatchifies the live audio buffer
with the target ashape.

What a decision needs to answer

• (a) Withdraw A18 as an independent row and fold the delivery into A3,
  recording Ltx2ConditionAudioByReference as owed by that row. Recommended in
  the spec.
• (b) Fund it as an engine-defined extension with no upstream referent. Needs
  a ratified product decision, a chosen position convention, a divergence note,
  and reclassification out of class A.
• (c) Repair the refusal text and the scope record only.

The operator-owned scope document is not edited from here; the contradiction
is recorded in the row spec instead, as #2584 did for A8.

Unverified

Whether the LTX-2.5 checkpoints were trained with a reference-audio condition on
non-DubIt pipelines at all is unmeasured. It needs a real-checkpoint render
and a GPU lease.

Start with .agents/specs/ltx25-ref-audio-delivery.md and the A18 section of .agents/specs/ltx25-completion-scope.md, then inspect the refusal at src/vllm/multimodal/ltx2_video.cpp:2934-2947 and the cited upstream callers. Compare the DubIt and trainer position conventions and the local state and trimming paths. Done means recording a ratified product decision, including the chosen convention and scope outcome, or correcting only the refusal and scope records.