`WriteJson` reads `Elapsed()` and `Records()` under TWO separate mutex acquisitions, so `wall_seconds` and the record set are no longer one snapshot

[localai-bot]

PR #1711 moves PhaseLog::WriteJson's clock read ABOVE its ByStart(Records()), so that the writer's own copy and sort stop being charged to the render's wall and therefore to unaccounted_seconds. That is the defect #1569 tracks and the repair is right. It has a cost, and this issue is that cost recorded rather than hidden.

The two statements now take the process-wide mutex TWICE:

const double wall = Elapsed();                              // one acquisition
const std::vector<Record> records = ByStart(Records());     // another

so wall_seconds and the record set it is compared against come from two different instants. Under the previous order the pair was effectively one snapshot in the direction that matters — the clock was read last, so wall >= max(end_seconds) held by construction. It no longer does.

The observable if it ever broke is a NEGATIVE tail gap, which gaps reports and which ltx2 phase log: the emitted table DECOMPOSES its residue into the gaps between leaves refuses at CHECK(seconds >= 0.0). So the table would say so; nothing would pass quietly.

It is unreachable on the shipped path, and that is a property of the CALL SITE rather than of the function. Both WritePhaseLog calls in src/vllm/multimodal/ltx2_video.cpp run after generate_span.Close(); that span is the last live scope; and PhaseLog::Close stops and JOINS the sampler thread before it returns when nothing is left live. No thread can close a scope between those two lines on any path this project ships. A fresh review also failed to stage the inversion adversarially: 27,471 probes of a churn thread against a replica of the two statements produced zero inversions.

NOT FIXED IN FLOW: the real repair is to make the pair a single locked snapshot, and Elapsed() and Records() are separate public entry points on PhaseLog, so a combined one is a public API change. A public API change owes its own row, spec and red-first evidence rather than being smuggled into a repair commit for a different defect.

Owned by .agents/specs/ltx25-phase-instrument.md ## Owed.

[localai-bot]

One correction from the second fresh review of PR #1711, at head 56cfa0118: the title and body above name WriteJson, and that is now the smaller half.

PhaseLog::RenderText takes the process-wide mutex three times for one table — Records(), Elapsed(), and the Instrument() call this pull request adds. The first two are pre-existing and have the same two-snapshot property WriteJson does; the third arrives with #1711. The console copy is therefore assembled from three separate reads of a live instrument rather than one, and the repair is the same one: a single locked snapshot, which is a public API change on PhaseLog.

Nothing observable follows from it on the shipped path, for the same call-site reason WriteJson has — RenderText's callers run after the last scope closed — but the surface this issue owns is both functions, not one.