(1c)'s boundary sampler carries three terms and the gate can falsify only one: deleting the flushed write leaves the estimator inside its own noise

[localai-bot]

Assertion (1c) in tests/vllm/multimodal/test_ltx2_video.cpp bounds a leaf record's span slack at four times the worst boundary the run's own 1 kHz sampler observed inside that record. The sampler's Loop() claims to time the three things PhaseLog::Open and PhaseLog::Close both do — the process-wide mutex, SampleNow()'s /proc/self/statm read, and one flushed write, the last standing for EmitLocked, which the live lane performs on every boundary on the shipped default.

Only one of those three terms is falsifiable by the gate. A fresh review of PR #1712 deleted std::fputs(".", sink); std::fflush(sink); from the loop — the whole flushed-write term, and the one the derivation argues corresponds to EmitLocked — and the case stayed green 3 of 3 with assertions: 606 | 606 passed | 0 failed, compile_status=0, the ceiling reading 3.24-9.19 ms against an unmutated range of 0.94-17.5 ms. Nothing distinguishes the term's removal from run-to-run noise.

That measurement is the reason the write now goes to /dev/null rather than to fd 2: it established that the SINK is not what the term measures, which freed the sampler to stop corrupting the live [render] stream #1413 exists for (158 occurrences in one capture, only 116 beginning a line). But it also means the derivation carries a term nothing holds. If somebody deletes it, or replaces SampleNow() with something cheaper, or drops the mutex acquisition, the estimator moves inside its own scatter and no gate says so.

Two more halves of the estimator are unmeasured rather than ungated:

• The sanitizer lane. The third shape of this bound needed a separate constant under ASan and TSan because a sanitizer instruments the scope boundary this quantity IS. The sampler now inherits that automatically, because it compiles into the same binary — but the sample count also grows with the slowdown, so the maximum is drawn from a deeper quantile there. Nobody has run sanitize-cpu against this estimator; the review could not, because a second full build tree does not fit at 93% disk.
• The full-suite regime, at n > 1. #1576 measured this quantity at 3.223 ms inside the full 102-case ctest run against 18.8 us under a -tc= filter at comparable load — 171x, and that 171x is the whole reason the constant was declared unsafe. Both the row's own 120-run validation and the review's 24-run one were taken almost entirely under -tc=. One full-binary run (containment case as row/MODEL-KIMI-LINEAR-E2E: §8 oracle golden CAPTURED (STRICT) — e2e blocked on bf16-resident loader #40 of 102) was indistinguishable from the filtered ones — ceiling 2.31 ms, bound 9.22 ms, worst slack 86 us — so the 171x did not reproduce, at n = 1. Neither population establishes the bound in the configuration the gate actually runs.

NOT FIXED IN FLOW: gating a term of the estimator means asserting a relation between the sampler's own operations, which is a second instrument over the first, and the two unmeasured regimes each need a build this box cannot currently hold beside the one it has. What would settle them: one sanitize-cpu build with the containment case run 40+ times, and one 40-run population taken through the FULL binary rather than a -tc= filter, both reported as distributions.

Owned by .agents/specs/ltx25-device-residency.md ## Owed.

[localai-bot]

The estimator this issue is about was measured and WITHDRAWN; it is not on main. This issue stays open as a constraint on any re-proposal rather than as a live defect.

PR #1712's fifth shape (4 x the worst boundary a 1 kHz sampler saw across the whole case) and its sixth shape (4 x the worst inside the record's own window) were both built, both had their whole mutation set run against them, and both were withdrawn:

• the fifth lets one 200 ms descheduling of the sampler thread turn a real 20 ms un-named phase from red 9 of 9 into a green case 2 of 4;
• the sixth reds an unmutated tree 10 times in 45 consecutive runs at loadavg 21.8-61.5, five of them because the honest head-and-tail exceeds four of the record's own local worst boundary.

(1c) on main therefore still carries the fourth shape's 30 ms constant, and tests/vllm/multimodal/test_ltx2_video.cpp is byte-identical to main. The full account, with both distributions and the one hypothesis that has not been tried, is in .agents/specs/ltx25-device-residency.md ### The span-slack bound, FIFTH and SIXTH shapes.

Everything measured in the body above still stands and is still owed by anyone who rebuilds a sampler-based denominator: the flushed-write term is unfalsifiable by the gate, and the estimator is unmeasured under the sanitizers and in the full-suite configuration where #1576's 171x swing was measured.