A kept dry run was keyed on a stamp of every sequence row of the archive,
updated_at and verified_at included, so any upload, ingest status write,
verify or tag edit anywhere invalidated it and the confirm planned the whole
selection again. The plan was also bound to the request, so a proxy cutting
the request off cancelled it and a retry started over.
- The stamp (SequenceRepository.GetStampAsync) now hashes only what the
planner reads from the index: every row's id and folder (collisions and
what a moving folder carries), the planned rows' title, first capture,
camera, lens, photo count and frames version, and the titles of the rows
inside a planned folder. With all, every row is planned.
- ReorganizePlanCache makes a plan on the thread pool with the host's
stopping token, once per key for concurrent requests; a request cut off
only stops waiting, and the same request again waits for the plan or
takes it. Refusals and failures are still not kept.
- A kept plan is dropped by its own timer five minutes after it was made
(and never used past that), instead of staying until the next request.
- The protocol and endpoint docs say that plan_changed answers changes the
index has; one made on disk alone is found by the job, which fails and
changes nothing. The Capacity doc no longer claims a step per sequence.
- Tests: what the stamp follows and ignores, a kept plan surviving an upload
elsewhere and a verify, one plan for concurrent requests and after a
cut-off, the timer, an all:true apply carrying a nested sequence beside a
blocked-only one, the first sequence told before a pause. The
no-archive-module job_unavailable test expected a folder the planner no
longer plans without a sidecar; it now expects the unreadable step.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 7e01a4f85fe4b880ddd98a481fc8591649cec254)