record(LTX25-PROMPT-ADALN): the ~138x understatement stands, the keyframes claim next door is now main's, and nothing enters through the door (#644, #658, #900)

[localai-bot]

Claims-and-records repair for #644, reconciled onto main on 2026-08-15. This branch authors 8 files, +684/−40, none of them under src/ or include/ — git diff $(git merge-base origin/main HEAD) HEAD --name-only -- src/ include/ returns zero paths, against a positive control of 8 paths for the same command without the pathspec.

Read that against the merge base, not against origin/main. A bare git diff origin/main HEAD today lists 44 files, 30 of them under src/. Every one of those 30 is main moving ahead of this branch's merge base e8048ef63 — most recently #901, which this branch does not carry — and not a change this branch made. Measured at origin/main = b5f27c9a4; the count moves whenever main does, which is several times an hour, so re-measure rather than quote it. The runtime this repair talks about landed as 65e79eee5 (#654) and 98f8e046d (#658).

Three things happened between the review and this head, and each of them changed what the branch should say.

The branch was 120 commits behind, and half of what it asserts is now false

baa92ccf7 (spec) and 020381676 (implementation) reached main on 2026-08-13 as the squash 65e79eee5. Then 98f8e046d (#880, issue #658) ported keyframes_abs_pos_embedding and retired BOTH refusals this branch reasoned about. Ten files conflicted — five source, five keyed records.

Four of the five source conflicts were this branch re-asserting retired refusals, and each takes main's side wholesale. Taking HEAD on src/vllm/model_executor/models/ltx2.cpp would have re-added the use_keyframes_abs_pos_embedding VT_CHECK that #880 deleted, re-refusing BOTH shipped DiTs. ltx2.h's 57-line corrected bullet hung off an "unported modules" list entry that no longer exists — and main already carries this branch's own meta-device finding, in a better place, on the FIELD where it decides behaviour (ltx2.h:144-153: "upstream builds on the meta device (loader/helpers.py:84-95) and loads with strict=False, assign=True, so the absent parameter stays on meta and the add is never reached"). Same for ltx2_loader.h and ltx2_loader.cpp, whose text explained a tensor-keyed refusal that is gone.

tests/vllm/models/ltx2_goldens.inc is generated and was not hand-merged. Regenerated with the committed command, python3 scripts/gen-ltx2-goldens.py --ltx2 ~/_git/LTX-2 --out tests/vllm/models/ltx2_goldens.inc, against Lightricks/LTX-2 at the pinned fd4ded7f (checkout clean, REGEN_EXIT=0). It differs from main by 12 insertions and 7 deletions, all of them comment lines in the prompt-AdaLN magnitude block: every golden VALUE on both sides — including #880's new keyframe arrays — byte-identical.

The five keyed records were resolved the way AGENTS.md ## Records says — take the complete target-branch version, apply the scoped edit again, verify unrelated keys are byte-for-byte equal — and re-verified against origin/main a second time after main moved again mid-gate:

file	changed lines vs origin/main	key this row owns
.agents/model-matrix.md	1 +, 1 −	MODEL-DIFFUSION-ltx-2-5-ltx2-video-transformer-3d-model
.agents/roadmap_v1.md	0	none — the surface moved
docs/FEATURES.md	0	none
docs/STATUS.md	0	none
docs/USAGE.md	0	none

model-matrix.md is the one record main is stale on: its LTX-2.5 cell still read "timestep-INDEPENDENT prompt K/V (cacheable once per request)", which this campaign retracted on 2026-08-12 and #654 ported away on 2026-08-13. git diff --numstat origin/main reports exactly 1 1, so every other key in a 500-row file is byte-identical.

The four zero-line resolutions are the finding, not an omission. This branch changes no lifecycle state, so AGENTS.md ## Public documents owes it no STATUS / FEATURES / USAGE edit at all. Its edits there were length compressions plus an allow_unported-era sentence that #880 has since replaced with a truer one — "BOTH DiTs load, NO allow_unported". Reapplying them would have reintroduced a retired opt-in. roadmap_v1.md needs no reapplication because main retired its inline issue table into .agents/issue-index.md; the #673 row goes there instead.

#673 was never in the issue index — on this branch or on main

AGENTS.md requires the number in three places that agree: the index, the row's spec, and the pull request body. It was in two. Appended now, with #900 below it; nothing edited, nothing removed. scripts/check-issue-index-append-only.py --base origin/main --head HEAD exits 0, scripts/check-agent-record.py exits 0 at ENGINE=156 MODEL=377 QUANT=82 KERNEL=51 BACKEND=82, unowned high-water untouched.

Nothing lands dead: reached, but nothing enters through the door — #900

AGENTS.md ## Nothing lands dead and .agents/reachability.md landed in 8f49ac3be (#886) after this branch was written, so the question is answered rather than assumed. It ships no checker — 8f49ac3be touches seven files, all documents, and ls scripts/ | grep -iE 'reach|dead' returns nothing against a positive control of 37 check-* scripts — so there is no exit code to report, and the guide says why.

Reached: YES. Chain followed by hand, every anchor re-derived at this head and asserted UNIQUE (matches exactly once in its file) against a positive control that reports STALE on a deliberately wrong line:

include/vllm.h:962 (vllm_video_generate) → src/capi/vllm_c.cpp:1646 → src/vllm/multimodal/ltx2_video.cpp:1063 (Ltx2VideoEngine::Generate) → :1730/:1732 → src/vllm/model_executor/models/ltx2_dit.cpp:785 and ltx2_device.cpp:1152, where use_prompt_adaln_single defaults TRUE at include/vllm/model_executor/models/ltx2.h:133 → ltx2_dit.cpp:581 → :205 → :140. Loader half: ltx2_video.cpp:533 → :624. Default configuration, not an opt-in.

Entered by the smallest failing test: NO. The reachability mutation from the guide — delete the production call site, ltx2_dit.cpp:140 if (prompt_mod != nullptr) { → if (false) { — BUILT=YES, compile_err=NO, so this is a test result:

suite	enters through	result
test_ltx2	Ltx2DitForward, by hand	RED — 3 of 43 cases, 6 of 4388 assertions, exit 1
test_ltx2_video	vllm_video_generate / the ABI	GREEN — 37 of 37, 784 of 784, exit 0

test_ltx2_video's fixture sets the flag TRUE (ltx2_video_fixture.h:258) and every render it drives executes the path. It just asserts no value the term can move. Filed as #900, listed under ## Owed in the spec, and not fixed in flow: closing it needs an ABI-level value oracle designed red-first, which is a row with a spec rather than a record repair. The LOADER half passes both halves of the rule, and the same harness shows it — M3 takes test_ltx2_video to 29 of 37 cases failing.

Every number re-measured on the merged tree, and two of them moved

Nothing was carried forward. scripts/measure-ltx2-prompt-adaln.py against the shipped NVFP4 DiT (ltx_core fd4ded7f, __file__ asserted, exit 0):

	video (4096)	audio (2048)
rms|table|, 48 blocks	0.017553	0.021925
rms|term|, uniform σ∈[0,1]	0.236446	0.347171
term/table, RMS	1347%	1583%
max|term| / max|table|	7119%	2817%
term/table RMS, the file's own LinearQuadratic sampler, 8 steps	1275%	1492%
modulated context, unit-rms	1.0035 → 1.0915 (+8.8%)	1.0033 → 1.3170 (+31.3%)

Audio fixture denominator by importing the generator so no weight is re-drawn: max|term| 0.0201763 / max|table| 0.0496868 = 40.6068%, giving 2817 / 40.61 = 69.4x; video 0.0252012 / 0.0487142 = 51.7327%, 7119 / 51.73 = 137.6x. The ~138x headline stands exactly as written.

What DID move is the test_ltx2_video counts, and #658 moved them. 37 cases / 784 assertions with LTX2_CHECKPOINT_ROOT unset, 37 / 9031 with it set — where the row recorded 30 / 502 and 30 / 8734. So #673's figure is 784 of 9031, 8.7%, not 5.7%. Corrected in the spec, in the index row and at the skip site in test_ltx2_video.cpp, each now carrying its date. Note the CASE count is identical in both configurations, so the "unchanged case count" reading the original note relied on no longer distinguishes them.

The review's meta-device probe already exists on main. scripts/measure-ltx2-keyframes-meta.py, committed by #658, re-runs the whole observation. Executed here against BOTH shipped files, exit 0: it reproduces every line of the transcript this branch quoted, and it carries its own positive control — a neighbour (scale_shift_table) materialising device=cpu is_meta=False on the same load, so is_meta=True on keyframes_abs_pos_embedding is the parameter's state and not a loader that never ran. It settles the FP8 half by execution too: 2 of 6124 keys match, [1, 4096] F8_E4M3 with 4096 of 4096 bytes non-zero plus a rank-0 F32 scale, against __metadata__ keys : NONE. Writing a second probe here would have duplicated it.

Mutations — all five, all BUILT, all restored byte-for-byte

#	Mutation	BUILT	Result
M1	host ModulateContext ignores prompt_mod	YES	RED — test_ltx2 3/43 cases, 6/4388 assertions, exit 1
M2	device path takes the static-only branch always	YES	RED — test_ltx2_device 1/18, 6/546, exit 1
M3	re-add use_prompt_adaln_single = false before the guard	YES	RED — test_ltx2_loader 1/28 cases exit 1; test_ltx2_video 29/37 exit 1
M4	prompt AdaLN driven by m.timesteps instead of m.sigma	YES	RED — test_ltx2 2/43, 4/4388, exit 1
M5	shift and scale rows swapped in the [2, width] row	YES	RED — test_ltx2 2/43, 4/4388, exit 1

M5's first attempt was a silent NO-OP and is reported rather than dropped. Its substitution spanned two lines and never matched, so the tree was unchanged, the build succeeded and test_ltx2 printed 43/43 SUCCESS — a green that says nothing about the guarantee. It was caught only because the harness prints git diff --stat for every mutation and that line was absent. Re-applied as a real edit (1 file changed, 2 insertions(+), 2 deletions(-)) it is red.

M3 shows the doctest counting trap. The case THROWS, so its assertions stop being counted and the total DROPS 4978 → 4967 while the summary prints 0 failed. The exit code and the test cases: line are the authority.

git status --porcelain on the three mutated source files was empty after each restore, and the final whole-tree status is clean.

Gate

Release, VLLM_CPP_CUDA=OFF, -j6, on the merged head.

CONFIGURE_EXIT=0
BUILD_EXIT=0
': error:' count = 0      (positive controls: Building CXX = 929, Linking CXX = 479 on the full build)
'No space left'  = 0      'BFD assertion' = 0      'warning:' = 0
ctest -N         = 477

Suite counts on this head, exit 0 each, against the 2026-08-13 figures:

Suite	2026-08-13	this head
test_ltx2	35 / 2435	43 / 4388
test_ltx2_loader	26 / 4826	28 / 4978
test_ltx2_device	15 / 523	18 / 546
test_ltx2_video	30 / 502 unset, 30 / 8734 set	37 / 784 unset, 37 / 9031 set

The growth is #658's, not this branch's: this branch authors no src/ and no include/ change at all — git diff $(git merge-base origin/main HEAD) HEAD --name-only -- src/ include/ returns zero paths. (Comparing against origin/main instead answers a different question, because main has advanced past the merge base; see the note at the top.)

Full ctest --test-dir build -j4 --output-on-failure on this head:

100% tests passed, 0 tests failed out of 477
Total Test time (real) = 234.44 sec
CTEST_EXIT=0

The following tests did not run:
	 19 - test_modelopt_mixed_precision_checkpoint (Skipped)
	199 - test_voxtral_e2e (Skipped)

Two of the three load-dependent tests passed in this run at a box load of 19–105: test_openai_conformance (0.77 s) and test_cpu_threadpool (11.18 s). Nothing needed a solo re-run. test_cpu_x86_llamacpp_floor is not registered in this configuration at all — ctest -N | grep -c llamacpp = 0, against positive controls of 1 for test_cpu_threadpool and 1 for test_openai_conformance in the same listing — so it was not run here and this head says nothing about it either way.

Known-red, proved pre-existing rather than asserted. For each, the checker's own inputs are byte-identical to origin/main on this head — git diff origin/main HEAD --name-only -- <inputs> returns zero paths — against a positive control on the paths this branch DOES change, which returns three:

tracked as	inputs changed by this branch
#873 — check-release-binary-contract, check-release-workflow, check-test-registration, test_check_release_binary_contract, test_release_manifest, test_release_pipeline, test_check_test_registration	0
#584 — windows-msvc-cpu, windows-msvc-vulkan	0
load-dependent — test_cpu_x86_llamacpp_floor, test_openai_conformance, test_cpu_threadpool	0
positive control — .agents/specs/ltx25-prompt-adaln.md, .agents/model-matrix.md, tests/vllm/models/ltx2_goldens.inc	3

Review repair — 7a6165dab: two anchors were dated, never re-derived

Review of d3f846768 returned FAIL on two file:line citations in the spec, both landing inside an unrelated CRF-refusal SUBCASE instead of the text they claim. :917-921 for the shipped-checkpoint skip-and-return becomes :1316-1319; :913-914 for the "differ in a TRAINED keyframes_abs_pos_embedding" comment becomes :1302-1303.

Both were correct when written at 2990ed81b. The row then re-measured its suite counts and dated them, but re-derived none of its own anchors into test_ltx2_video.cpp — the one file whose line numbers #658 moved most. 98f8e046d inserted 169 lines above the cited text and this branch added 9 more, so the citations drifted 389 lines while the row was auditing the numbers beside them.

Verified by uniqueness, not existence, with every clean result paired to a control that must report STALE:

check	rule	result
:1316-1319 skip-and-return	enclosing TEST_CASE is ltx2 video: the SHIPPED Lightricks checkpoints parse and load, declared exactly once	FRESH
:1302-1303 keyframes comment	text occurs exactly once in the file	FRESH
control — old :917-921	same rule	STALE
control — old :913-914	same rule	STALE
control — :1451-1454, the identical VAE-case skip block	same rule	STALE
control — off-by-two span :1304-1305	same rule	STALE

The skip block is not uniquely identifiable by its text: :1316-1319 is byte-identical to :1451-1454 in ltx2 video: the SHIPPED Lightricks VAEs and upsampler load, so a grep-based re-derivation gets two hits and cannot choose. The sentence now names its enclosing case, which is what makes the anchor re-derivable rather than merely trusted.

Spec-only edit; git diff HEAD~1 HEAD --name-only lists .agents/specs/ltx25-prompt-adaln.md alone. check-doc-checkpoint, check-commit-trailers, check-commit-style, check-issue-index-append-only and check-agent-record each exit 0 (ENGINE=156 MODEL=377 QUANT=82 KERNEL=51 BACKEND=82), against a positive control — check-commit-style on a reversed range — that exits 1.

mergeable reads CONFLICTING / DIRTY, and it is a GitHub artefact, not a tree conflict. A real git merge origin/main at this head reports "Automatic merge went well" with zero paths in git diff --diff-filter=U, in both directions, at origin/main = b5f27c9a4 (confirmed identical to GitHub's main tip). The single overlapping file is .agents/issue-index.md, which both sides appended to and which carries merge=union in .gitattributes; GitHub's mergeability check does not apply .gitattributes merge drivers, so it sees the append-vs-append as a conflict. A plain 3-way git merge-file on the three blobs exits 1 with 2 conflict markers, and does so identically at the previous head d3f846768 — so this predates 7a6165dab, which does not touch issue-index.md. Negative control: the same 3-way merge on .agents/model-matrix.md exits 0.

Spec: .agents/specs/ltx25-prompt-adaln.md. .agents/specs/ltx-2-5.md is operator-owned and untouched. Issues: #644 (campaign), #673 and #900 (owed).

FOLLOWING_AGENTS_PROTOCOL

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

[localai-bot]

Operator call on the one question the review left open, recorded explicitly rather than by omission, because the reviewer was right that it should not be settled by silence.

Question. AGENTS.md ## Nothing lands dead and .agents/reachability.md say a disclosed slice must name "the row ID that owns the wiring". The commit body, the PR body, and ## Owed all name the issue (#900) and the shape of the work, but no row ID, because no such row exists.

Ruling: the staged-slice clause is not triggered, so the landing is permitted — and on stronger grounds than the clause.

That clause governs landing a layer before the wave that wires it. This PR lands no product code at all: 8 files, none under src/ or include/. The code in question is already on main via 65e79eee5 (#654) and 98f8e046d (#880), where it is reached from vllm_video_generate on the default configuration — the review re-derived that chain by hand, 14 anchors, each unique in its file.

What is unreached is a test entry, not a slice. #900 is therefore an honest disclosure of a pre-existing hole in the gate around already-landed code, and filing it as owed debt is the correct disposition. Requiring a row ID here would mean opening a row whose only purpose is to be named by a document, which is the kind of record-for-its-own-sake this protocol keeps trying to delete.

What makes me confident rather than merely persuaded is the reviewer's second mutation. Deleting the production call site (ltx2_dit.cpp:140 → if (false)) takes test_ltx2 red and leaves the ABI suite green at 37/37. On its own that could be an artifact of deleting one if. So they zeroed out.prompt_modulation at a different altitude (ltx2_dit.cpp:582, straight after PrepareTimestep) — BUILT, 1 file changed, 1 + — and got the same split: test_ltx2 exit 1, test_ltx2_video exit 0 at 37/37, 784/784.

Two independent injections, one verdict: the ABI suite cannot see this term at all. That is a class gap, not a one-line artifact, and #900 states it accurately rather than understating it. An issue that describes the gap correctly is worth more than a row opened to satisfy a sentence.

If a row is later opened for the ABI-level value oracle, it should name #900 — not the other way round.

FOLLOWING_AGENTS_PROTOCOL