Files
orca/docs
Brennan Benson 3316614da6 fix(agent-status): one revision rule for every teardown marker
The store's child marker was checked with a plain `has`, so a torn-down
child id stayed refused only until compaction evicted the marker and then
silently started re-admitting. The verdict depended on store volume, not on
the work. The parent side already moved to a revision compare; this brings
child, alias and fact onto the same rule.

`agentStatusTombstoneFences` is now the only spelling of that rule, called
by both admission and state validation for all four entities: a marker
fences only writes at or below its own revision, so an absent marker and an
older one answer the same and compaction can never change a verdict.

What actually fences a delayed observation survives compaction on its own:
the lock-step revision envelope refuses any out-of-order mutation, and the
live-row lookup answers `unknown-child` for a removed id. The new test
ablates the marker and shows both still hold.

This also stops a marker latching a reusable provider name. A child work id
is minted; a `task_id` is a name later work will use again. The next
legitimate occupant of a name is admitted and gets a fresh identity, so it
can never be confused with the work that held that name before.

No wire or persistence change: snapshot version, mutation shape and
tombstone record are untouched, and the same markers are written at the same
revisions.
2026-09-15 22:45:52 -07:00
..
2026-09-15 18:29:06 +00:00
2026-08-28 00:59:21 -07:00