Mass passes need to say what they do now (photo x of y, sequence n of m) without
slowing down. This adds the contract and the engine side; handlers and displays
follow in their own parcels.
- Core: JobDetail (step, item, item x of y, part n of m with unit), JobStep and
JobPartUnit appended-only and stored by kebab name; Job.Detail.
- Protocol: JobDetailDto, appended as JobDto.Detail, JobProgressEvent.Detail and
NodeProgress.Detail, so older peers ignore it and newer ones fall back.
- Data: JobContext keeps a detail across reports (ReportAsync with a detail,
ReportItemAsync without allocating, CheckpointAsync with a detail) and publishes
at most once per ProgressInterval (500 ms on the engine's TimeProvider), at once
for the first report, a new step or a completed count; a dropped report costs no
allocation or lock, and the persist that falls due publishes what it saves.
Persist stays at most every 5 s. JobWire maps steps, units and details; an
unknown name reads as none.
- Migration 0011 adds jobs.progress_detail (nullable JSON text); an unreadable
cell reads as no detail. Lists return it.
- Server: ProgressGate (500 ms per job, step-change and completion bypass)
shared by the engine host and render nodes; node details are stored and
broadcast, an older node without one still works.
Refs #65
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>