Design 47 H2: H-A-468, the grants part makes a seat's one shared worker

This commit is contained in:
2026-11-22 11:29:29 +01:00
parent 1ea623bf1e
commit 0dbcb115f9
@@ -817,3 +817,14 @@ The lead's one-line answers to the questions H2's develop and review raise (impl
outcome) in the same change (the seat's form raised once, regenerated, read.go's row writer), so a page's list is one
read of the short form (one-question-one-tool, P3); the full form stays as it is. The view's developer makes the
runner's change in the same branch, since the view is its first reader. P3.
- **H-A-468 (the grants part makes a seat's one shared worker, 11:35)** A fault of H2 found at its delivery: the H2
builder pulls build steps from the seat mesh-builds' work queue through the durable `SEAT_MESH_BUILDS_worker` and
waits for its creation ("the broker derives it"), but no part creates it: the grants part composes shared workers only
for seats pulled by kind, and the old builder subscribed to the accept subject directly (its queue piled up unread:
2,246 asks). The whole test hid it by creating the worker itself. (1) The grants part composes, for every seat with
an accept that is not pulled by kind, the one shared worker buscompose.SeatQueue names, and creates it where the bus
lacks it, as it creates the seat's streams (P3: one derivation, in the broker's part). (2) The builder binds the
worker again after a deletion (the advisory subscription outlives the first bind; no sync.Once). (3) The builder's
whole test proves the pull with the worker the grants composer makes, not one the test makes. Delivered as a small fix
on main; the grants part and the builder are placed by hand from that head (the pipelines cannot pull before it).
fix-a-bug-met-twice (the kind: a whole test that makes what the broker should), P-failure-loud, P3.