spec(LTX25-DEVICE-RESIDENCY): the levers were ranked and nothing staged them, so the instrument is stage zero (#1264)

[localai-bot]

ltx25-decode-speed.md is a SPIKE that ranked the LTX-2.5 device-residency
levers, read every oracle and filed thirteen issues it owns, and states in its
own stop conditions that it has no implementation authority and no fresh review.
This row takes that ownership. It cites the spike and restates none of its
evidence, oracle analysis or provenance.

.agents/specs/ltx25-device-residency.md is READY with the nine structured
sections and a ## Now. It stages fourteen issues into W0-W7 plus a parallel
oracle lane, each stage landing on its own branch, reachable from a production
entry point at its own merge commit, revertible without touching the next, and
owing an open issue before it starts.

The instrument is first, and the argument is not that instruments are virtuous

The ranking every later stage would inherit has three measurement defects. The
evidence is unretrievable (#1040 — the sampler CSVs and rung 2's run.log are on
a host that stopped answering, and the spike marks even its passing numbers
REMOTE_UNVERIFIED). The subject is unnamed (#1087 measures a 1731 s phase and
says in its own text "Do not guess it from the duration"). And the ranking is
STALE rather than merely unverified: #1009 landed as #1041 and #1208 landed as
LTX25-TEXT-LINEAR-SEAM, and neither was re-measured end to end, so the profile
the ranking describes is not the profile in this tree. W1 is therefore an
explicit gate on the order itself, and ## Now is amended with the re-derived
order before W2 starts — unless W1 itself reports BLOCKED, in which case W2 and
W3 proceed on their byte-compare and RSS gates and W4/W5 do not start. A lever
whose rank does not survive re-measurement is re-placed or dropped there, in
writing.

Three corrections to the framing, each with a file:line at this base

The denoise IS device-routed: ltx2_video.cpp:847 selects
Ltx2StreamDitToDevice and :3864-3866 selects Ltx2DitForwardDevice, which
calls CheckWeightsResident at ltx2_device.cpp:1144 and drives 48 blocks of
device ops. The spike's own record says the sampled window had written no frame,
which excludes the decode too. So #1024's GPU-zero window is neither the denoise
nor the decode, and the campaign's headline measurement does not identify its own
subject.

#1015 does NOT gate device residency. Ltx2StreamDitToDevice refuses
widen_to_f32 by name (ltx2_loader.cpp:759-765) and keeps one host buffer live
per tensor, so the double-hold at :739-754 is a host-arm defect. It stays early
because the host f32 arm is the reference W4's and W5's correctness gates compare
against. #1016 is the one that gates device headroom, because on GB10's unified
memory the device copy and the dead file cache come out of the same 119 GiB.

The conditioning stage is host-only BY TYPE and was in no lever table. Filed as
#1269 and staged as W4: ltx2_video.cpp:2085, :2799, :4638 hard-code a CPU
queue, ltx2_text_encoder.cpp:446 hard-codes a second one inside the extractor,
and Gemma4Weights is OwnedTensor over OwnedBytes with no device field.
LTX25-TEXT-LINEAR-SEAM found it, wrote "it needs its own issue", and filed
none.

Two levers this campaign does not carry as stages

#1011 becomes a rider on W5 rather than a stage, because once the decode is on
the device, upstream's memory-efficient HOST decode optimizes a path production
no longer takes; its portable half is the memory format W5's arm consumes. #1202
is deferred behind W1 with its measured number attached (2.3% of one pass), not
carried as debt with no number. Both conditions are written down so a later
session can see what would bring them back.

Records

Hazards are written as gates a run script asserts, not prose: the GB10 OOM
reboot, the ephemeral /mnt and root overlay against COS_PERSISTENT
/usr/local, the rc lease, and the lease that carries bytes rather than
executables. No docs/STATUS.md or docs/BENCHMARKS.md edit is owed: no
capability changes lifecycle state and no measurement is taken. No matrix row and
no claim file, argued in ## Risks/decisions — check-agent-record.py builds
its row-ID pattern from prefixes that admit no LTX25- id, so such a row would
be one no checker reads inside a file every future edit must write. The one
matrix cell this campaign genuinely invalidates is updated:
ENG-CUDAGRAPH-DIFFUSION now names the row that owns #1164's unblock order.

check-agent-record.py, check-public-doc-tables.py,
check-issue-index-append-only.py, check-commit-style.py and
check-commit-trailers.py pass. The record checker's green is weak evidence
about this spec, and the spec now says why: check_spec takes a ClaimRow, so
its structured-section requirement reaches only a spec linked from a matrix row,
and this one is linked from none. No checker in this tree reads a bare
path.cpp:123 written in prose either — check_links validates markdown link
targets and #L<n> fragments only. A person reading them is the whole
instrument, which ## Port map now discloses beside its per-stage
re-derivation obligation.

Applied from the fresh review

The review returned PASS with four LOW findings, all text inside the spec, all
applied on this branch. The central gate admitted two readings — ## Now said
the order is re-derived before W2 unconditionally while ## Risks let a
BLOCKED W1 proceed with no amendment — and both sections now carry the same
stop condition, with W3 named rather than inferred. Four anchors named a line
they do not sit on at the stated base 27d5432f9, re-derived here with
grep -n: CheckWeightsResident is :1144 not :1145, the bf16 buffer push is
:729 not :728, the f32 push is :752 not :753, and Ltx2RebindDitLoras
runs :808-912 not :806-830; the W7 row's capture primitive starts at
cuda_backend.cu:206, not :203. Two prose slips beside correct anchors:
Ltx2DitForwardDevice takes vt::Queue&, not vt::Backend&, and 14 other
files call MaybeReleaseSourcePages, not 15. And OwnedBytes "has no device
field" now cites owned_bytes.h:42, where the type is defined, beside the
qwen3_5_weights.h:47-56 that only holds one.

The issues this row stages

The thirteen ltx25-decode-speed.md owns under ## Owed, plus #1269 which this
row filed:

Stage	Issues
W0 the instrument	#1010, #1040 (artifact contract)
W1 re-derive the ranking	#1040 (closes), #1024 (re-measured, with the control it owes)
W2 the load path	#1016, #1015, #1210
W3 the staging path	#1021
W4 the conditioning stage	#1269
W5 the video VAE decode	#1007, with #1011 as a rider
W6 the residency question	#1014
W7 the graph decision	#1164 is decided here, not implemented
O1 the oracle lane	#1012
deferred behind W1	#1202, with its measured 2.3% attached
closed, listed so the count closes	#1009, landed as #1041

Closes #1264.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [Claude Code]