CheckSeamAnchor proves a seam anchor's POSITION and not its MAGNITUDE: a zero-width anchor passes green

[localai-bot]

CheckSeamAnchor proves a seam anchor's POSITION and not its MAGNITUDE, so a zero-width anchor passes it green.

Found while landing the three anchors #1668 owes, by mutating the gate that holds them. Filed by the change that introduced the gate, against its own work.

What the gate does prove

CheckSeamAnchor (tests/vllm/multimodal/test_ltx2_video.cpp) holds load.dit_config and artifacts.mux to a POSITION, deliberately carrying no constant, because .agents/specs/ltx25-phase-residue.md ## Design 3 withdrew the instrument-relative bound after three fresh reviews measured its tail. Its clauses:

• the anchor is emitted exactly once, is not a span, is not nested;
• it opens at or after its predecessor's end and closes at or before its successor's start;
• no other leaf lies inside the window it claims;
• for artifacts.mux, which has no successor, nothing follows it — it is the LAST leaf of its render.

Six mutations red against that, including both call-site deletions.

What it does NOT prove, measured

Mutation M7, NOT DETECTED. Close artifacts.mux immediately after opening it, at the head of the tail, and drop the late mux_phase.Close():

  phase::Scope mux_phase("artifacts.mux");
  mux_phase.Close();          // the whole tail is once again un-named
  result.frame_count = rendered_frames;

[doctest] assertions: 796 | 796 passed | 0 failed |
[doctest] Status: SUCCESS!

Every clause survives it. The count is 1; it is neither span nor nested; it still opens after artifacts.audio ends; nothing follows it, so it is still the last leaf; and a zero-width window [s, s] contains no other leaf vacuously, because every interval satisfies end <= s + 1e-9 || start >= s - 1e-9. So the anchor reports ~0 s, the region it was added to name is back in unaccounted_seconds, and the table reads green with a name on nothing.

The same argument applies to load.dit_config.

Why this is a class and not an oversight

It is the third appearance of one shape, and the file already records the other two:

• decode.audio.mel / decode.audio.vocoder: "A partial transfer -- an anchor covering half its call -- passes 0.50 and is not detected here. Closing that needs a scope INSIDE the callee."
• LTX-2.5: denoise.step and denoise.update are open to a seconds transfer that no assertion in the file can see #1568, denoise.step / denoise.update: the seconds transfer, where no floor separates the honest share from the transfer because the two distributions overlap.

A magnitude clause is what would close it, and one family of magnitude clause is what this row has already measured shut twice: any floor whose denominator is the INSTRUMENT, the wall, or a written-down number of seconds. ## Design 3: a seam's honest share is a property of the box, residue <= 2 * instrument went red 4 times in 45 at the table bound and 28 in 160 on a standalone probe, and a 20-run sample of that quantity does not see the tail that decides the gate. A floor of that family must not be added to make this issue go away.

That ban does NOT reach every magnitude clause, and an earlier revision of this issue said it did. Those three denominators all ask how many seconds a seam SHOULD hold, and the honest answer is a property of the box. A ratio whose denominator is the seam being named, measured in the same run, asks a different question. See the candidate below.

What would actually close it

The same two shapes the row already names, neither of which is a tolerance:

1. The seam-extent ratio, which is where to start. The anchor's own extent against the seam its two neighbours define, both measured in the same run: (end - start) against (next_start - previous_end). Honest that is near 1, because a correct anchor covers the seam except for two instrument boundaries; collapsed it is near 0. Wide separation, of the kind the decode.video.vae floor already exploits, and its denominator is the region being named rather than the instrument — so it is outside the banned family. It still needs a constant and a cross-box distribution, so it is a candidate and not a conclusion, and it needs its own red-first measurement and fresh review.
2. An anchor INSIDE the callee, so the window cannot be narrower than the work — the repair the decode.audio note asks for and the one LTX-2.5: denoise.step and denoise.update are open to a seconds transfer that no assertion in the file can see #1568 needs.
3. A bound on a quantity the scheduler cannot move, which is the open question of LTX-2.5: nothing bounds the instrument's own share of a leaf, so both residue gates can widen silently #1570 and, one level up, of The LTX-2.5 phase-table SUM gate's 95% floor sits inside its own measurement, and it is RED on main #1439.

Do not close this by widening, and do not close it with a floor from the banned family above. Read .agents/specs/ltx25-phase-residue.md ## Design 3 and ### M7 first.

Scope

Not fixed in flow: closing it needs an anchor inside a callee or a scheduler-independent bound, each of which is its own row with its own red-first tests and fresh review. The anchors themselves are correct and are where the gap decomposition said the time was; what is unheld is that they stay wrapped around it.

Owned by row LTX25-PHASE-RESIDUE, listed under ## Owed in ltx25-phase-residue.md.

[localai-bot]

A fresh review found this independently, and measured what it costs

Confirmed on the same head by a reviewer who did not write the gate, in a stronger form than the original filing: both anchors collapse, not just artifacts.mux, and the collapse is green on the FULL suite as well as the focused case.

The mutation is not a contrived one. It moves each phase::Scope declaration down to sit on its own Close() — the exact shape of a refactor that hoists work out from under an anchor:

R1  src/vllm/multimodal/ltx2_video.cpp   load.dit_config opened at its Close
R2  src/vllm/multimodal/ltx2_video.cpp   artifacts.mux   opened at its Close

mut sha e0d0a53460f20187  lines preserved: True
build rc 0
FOCUSED rc 0 | [doctest] assertions: 796 | 796 passed | 0 failed |
FULL    rc 0 | [doctest] test cases: 102 | 102 passed | 0 failed | 0 skipped
             | [doctest] assertions: 4668 | 4668 passed | 0 failed |
             | [doctest] Status: SUCCESS!
restored True

Line-count preserving, so the reader-anchor ledger case is not what fails.

What it costs, from attribution_multichunk/phase-log.json

quantity	head	collapsed
load.dit_config duration	0.000109239 s	0.000016543 s
gap load.dit -> load.dit_config	0.000003296 s	0.000141781 s
gap load.dit_config -> load.video_vae	0.000001042 s	0.000087407 s
gap artifacts.audio -> artifacts.mux, r1 / r2	0.000000983 / 0.000003287 s	0.000036901 / 0.000066776 s

About 229 us handed back to the residue on a fixture whose whole residue is about 1.2 ms — roughly a fifth of what this row exists to name — while the anchor that was supposed to name it measures 16.5 us, which is instrument cost.

The (4b) gap gate cannot see it either. gap_count == leaf_records + 1 and gap_total == unaccounted_seconds both still hold, because the collapse only SPLITS one gap into two rather than changing their sum.

The next traceable hypothesis

The same review proposes a discriminator whose denominator is not the instrument, which is what separates it from the bound ## Design 3 withdrew: the anchor's own extent against the seam its two neighbours define, both measured in the same run.

(end - start)  against  (next_start - previous_end)

Honest, that ratio is near 1, because a correctly placed anchor covers the seam except for two instrument boundaries. Collapsed, it is near 0. That is the wide separation the decode.video.vae floor already exploits, and crucially it never asks how many SECONDS a seam should hold — which is the question that made residue <= 2 * instrument a property of the box rather than of the tree.

It still needs a constant, so it is a candidate and not a conclusion. It needs its own red-first measurement across boxes and its own fresh review. Recorded so the next attempt starts here instead of re-deriving it.

This does not change the warning in the issue body: do not close this with a share floor against the instrument, and do not widen anything.

[localai-bot]

The candidate discriminator this issue names was measured, and it does not separate

This issue and .agents/specs/ltx25-phase-residue.md ### M7 both name one non-banned candidate and describe it as "near 1 honest, near 0 collapsed": the anchor's own extent against the seam its two neighbours define, (end - start) against (next_start - previous_end). It was measured before anything was built on it. It is shut as proposed.

Method: test_ltx2_video -tc="ltx2 video: the three carrying phases contain their work and the load keeps its order" under VLLM_KEEP_TEST_ARTIFACTS=1, reading attribution_multichunk/phase-log.json. HONEST = unmutated ced0ab639; COLLAPSED = M7 on BOTH anchors, line-count preserving, same tree, same box, one binary pair. artifacts.mux has no successor, so its far side is the enclosing generate span's end — the tightest available, which is the reading most favourable to the candidate.

arm	anchor	n	min	p05	median	max
honest, 67 runs	load.dit_config	67	0.0626	0.7366	0.8057	0.9646
collapsed, 25 runs	load.dit_config	25	0.0428	0.0473	0.1089	0.2122
honest, 67 runs	artifacts.mux	134	0.0156	0.3440	0.5977	0.8779
collapsed, 25 runs	artifacts.mux	50	0.1414	0.1474	0.2989	0.4766

For artifacts.mux no constant exists. The collapsed range lies entirely inside the honest range. Held to the collapsed maximum the honest arm reds 11 of 44 in the quieter sub-population and 15 of 90 in the busier one — both regimes, so this is a property of the anchor, not of the box's load.

For load.dit_config the overlap is contention, and it is the withdrawn bound's contention. A k in (0.2122, 0.7187] catches every collapse and reds 2 of 67 honest renders: 0 of 22 quieter, 2 of 45 busier. Clean when the box is calm and a tail when it is not — 3.0%, the same order as the 4-in-45 that withdrew residue <= 2 * instrument.

The premise fails too, without any tail. artifacts.mux's honest median extent is 0.5977 because the anchor is tiny: median duration 14.5 us against a median instrument_seconds of 7.0 us on the same records, with per-record instrument_seconds / duration running 0.394 to 0.997, median 0.495. Half of what it measures is this instrument's own charge inside it. ### 1 already said it was "close to naming its own cost"; this is that as a distribution.

And the mechanism is the withdrawn bound's mechanism, visible in the table. The honest outliers are ordinary anchors beside a preempted boundary:

run	anchor	its duration	the flanking gap BEFORE it	ratio
h2-5	load.dit_config	234 us	load.dit -> it, 3473 us	0.0626
h2-1	artifacts.mux r2	39 us	artifacts.audio -> it, 2446 us	0.0156

Both gaps are ONE instrument boundary — a Close stamp then an Open stamp — and both are ~3 us on a quiet run. ## Design 3 already says the un-instrumented remainder of a boundary dilates faster than the instrumented part under contention. The seam-extent ratio puts that remainder in its denominator, so anchor / (anchor + gaps) is a monotone function of gaps / anchor and inherits the same tail. It asks a different QUESTION than the withdrawn bound and is conditioned on the same QUANTITY, and the conditioning is what decided the earlier one. Ask of any successor what its denominator dilates with, not only what it means.

What the same data says to try next

A fresh review mutated the CONCLUSION rather than only checking the arithmetic, and narrowed it: the negative result belongs to the definition, not to the family. Subtract the record's own instrument_seconds from the numerator — net = duration - instrument_seconds — and:

discriminator	load.dit_config honest / collapsed	artifacts.mux honest / collapsed	honest reds at the collapsed max
extent / seam, as proposed	[0.0626, 0.9646] / [0.0428, 0.2122]	[0.0156, 0.8779] / [0.1414, 0.4766]	2/67 and 26/134
net / seam	[0.0460, 0.9187] / [0.0024, 0.0041]	[0.0030, 0.4637] / [0.0039, 0.0298]	0/67 and 2/134
net / (net + gap_after)	[0.7578, 0.9922] / [0.0028, 0.0060]	[0.0252, 0.7355] / [0.0183, 0.0677]	0/67 and 1/134

The last row separates load.dit_config by two orders of magnitude with no honest red in 67 and leaves artifacts.mux with one in 134. That red is named rather than averaged away: h2-38 render 2, where the anchor's whole 2968.0 us extent was its own 2957.8 us instrument charge — the boundary itself was descheduled, so net is 10 us of a 2968 us window. Same mechanism, one level in, exactly as the warning above predicts.

This is one measurement, not a conclusion. It owes its own constant, its own cross-box distribution, its own red-first mutation and its own fresh review, which is what this issue asks of any candidate. It is recorded so the next attempt starts from the definition that worked rather than the one that did not.

Correcting this issue on one point

The body says the repairs are "the seam-extent ratio", "an anchor INSIDE the callee", and "a bound on a quantity the scheduler cannot move (#1570, #1439)". #1570 was CLOSED by be432e8e3 on 2026-08-23, so only #1439 stands above.

And a non-sequitur to avoid, which an earlier draft of the spec correction contained and the review caught: CheckSeamAnchor as it stands proves a POSITION, and a position clause cannot bound a magnitude — but that does not mean no clause added to CheckSeamAnchor can. A MAGNITUDE clause there is not forbidden. It owes the measurement above.

M7 also reproduces here: both anchors collapsed onto their own Close(), green in 24 of 25 focused runs. The single red is decode.audio's coverage floor, which is the #1470 flake class and not an anchor clause.

Record correction in #1907.

FOLLOWING_AGENTS_PROTOCOL

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