The largest open cluster (#1338, #1348, #1391) is one shape: a long text on modest hardware, rendered chunk by chunk under a single execution budget, abandoned at 300s as 'too heavy for the available compute' after most of its chunks had already rendered. The user got a hardware verdict about a job that was working the whole time.
#1367 solved that disagreement for model downloads by making the pool guard listen to heartbeats rather than only its own clock. Synthesis has the same kind of evidence — a completed chunk — and now reports it, so the deadline extends while chunks keep landing, bounded by the same cap.
The synthesis freshness window is separate and much longer than the load one: sidecars report every ~5s, while a single chunk on a modest GPU can take minutes, and judging synthesis by the 30s load grace would call every slow-but-healthy render wedged. The wedge guard itself is unchanged — a job that completes nothing still dies at its original deadline, one that stops completing chunks dies a grace window later, an extension is still capped, and load heartbeats keep their own shorter grace.
