Design 47 H2: H-A-413 a dry run's kept stores; H-A-410 second addendum

This commit is contained in:
2026-11-22 01:01:10 +01:00
parent 4d8c853f46
commit 6daefe666f
@@ -360,3 +360,13 @@ The lead's one-line answers to the questions H2's develop and review raise (impl
Settled: a nak'd ask is waiting again (the stream redelivers it), so the runtime lists a message whose delivery count
is above zero and that is not held by a worker as waiting too, with `again: true`; a one-row change in T4, by the
mesh-tools developer, before the review. P-failure-loud.
- **H-A-413 (a dry run keeps on the bus only the stores it writes to the bus, 01:02)** §804, I3. The state `dry-runs`
keeps, under `<branch>.<call>.<kind>.<name>`, only the two stores a dry run writes to the bus today (packages and
manifests, passed through to the trunk's states, which is §804's fault); the working copies held in memory (candidates,
modules, graph, taken, predefined, old sections, the branch view) stay in memory. Writing every working copy would put a
whole catalog into a 16 MiB state at every branch take-in and re-pin for nothing a reader needs. P3.
- **H-A-410 (second addendum, 01:02)** The bus cannot say per message whether a worker holds it, so mesh/queue.waiting
lists every message still in the stream after the worker's ack floor and marks `again: true` each one at or below the
worker's delivered sequence (handed at least once and not acked, whether held now or handed back); the holder's asks
tool tells held from waiting-again through its mesh-wide state `taken`. The bus says what it knows truthfully, the
owner's state says the rest. P-mesh-owns-the-truth, P-NATS-native.