LTX-2.5: nothing bounds the instrument's own share of a leaf, so both residue gates can widen silently

[localai-bot]

Row LTX25-PHASE-RESIDUE (#1536) replaced two wall-clock ratios with residue <= 2 * instrument, where instrument is the wall PhaseLog spends inside its own entry points.

That is stricter than what it replaced only while instrument stays small, and nothing bounds it from above. Tick charges to the innermost live record, so moving the DiT tick out of Evaluate would charge about 110 flushed writes to denoise and buy a budget larger than the 0.75 share floor the row deletes. The ratio is emitted as a MESSAGE and not asserted, so a change that made the instrument ten times more expensive would widen both gates and print a small number.

What would close it is a bound on instrument_seconds / duration_seconds per record — a second gate, on the instrument's cost rather than on the residue. Found by the fresh review of #1536 as F5.

FOLLOWING_AGENTS_PROTOCOL

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

[localai-bot]

Closed as superseded: the bound this issue guards was measured and withdrawn.

The residue <= 2 * instrument shape needs Record::instrument_seconds, which does not exist on main — git grep instrument_seconds origin/main -- src include tests returns nothing. It existed only on refs/pull/1556/head, closed unmerged (superseded by 519303d15).

#1668 settles it in terms: three fresh reviews measured the proposed bound and found a heavy right tail, and the recorded direction is "Do NOT re-propose the bound... Both ratios therefore STAY."

A gate cannot widen against a quantity that was never introduced.

[localai-bot]

Correction to my close comment above, posted because one of its two stated reasons has since become false.

I wrote that the residue <= 2 * instrument shape "needs Record::instrument_seconds, which does not exist on main". That was true when I wrote it. It stopped being true about an hour later: be432e8e3 (#1711) introduced it, and instrument_seconds now has 25 occurrences across src/, include/ and tests/. I merged that PR myself, so I invalidated my own reason.

The closure still stands, on the other reason. The bound was measured and rejected, not merely unimplemented:

• tests/vllm/multimodal/test_render_phase_log.cpp:24 records three fresh reviews measuring residue <= 2 * instrument RED 4 times.
• include/vllm/multimodal/render_phase_log.h:148 names it as the "obvious use" and does not adopt it.
• LTX-2.5: three phase anchors and the instrument's own cost accounting are measured and reviewed, and none of them is on main #1668's recorded direction is "Do NOT re-propose the bound ... Both ratios therefore STAY."

So: the quantity exists now, the bound over it was tried and found to have a heavy right tail, and re-proposing it is what #1668 forbids. Reopen only with a measurement that contradicts those four red runs, not with the observation that instrument_seconds is now available.

Leaving the stale sentence in place rather than editing it, so the record shows what was claimed and when.