You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit bb5c125
Browse filesBrowse the repository at this point in the historyBrowse files
♻️ Give one owner the grids a terminal installation issued (#730)
The registry was a Set the authority searched. Two things have to meet before a
grid exists — the document submits the authored request and its pane work, a
provider presents a composite for that exact object — and nothing held both
sides or decided when they had met. The supervisor does: it keeps the entries,
converges them by object identity and installation generation, and is where
every refusal a presentation can meet now lives.
Each grid runs as a task the supervisor keeps, in a scope of its own beneath the
operation that submitted it. Beneath, and not beneath the supervisor, because a
pane's durable identity and the bindings its content reads are the expansion's:
parenting grids to the supervisor's own scope loses both, and the suite says so
loudly — pane durable order, pane bindings, the startup-failure paths and the
provider-failure path all fail together.
That parentage is also what makes a grid impossible to strand, which is worth
stating plainly: removing either finalizer the supervisor registers changes no
test, because the submitting operation unwinds whenever the call that routed it
does and takes the grid with it. They are kept as the supervisor saying what it
owns rather than as the thing that makes it true, and the comments say which.
Six rows drive the ownership directly, each one finite and signalled rather than
timed: a provider that answers without presenting opens nothing, one request
opens one grid however often it is presented, a settled grid is gone while the
next still opens and its composite was destroyed first, a presenting call that
is cancelled leaves no pane and no composite behind, an installation teardown
stops the grid still live and waits for it, and a second grid cannot be live
beside the first — which is why "every remaining grid" is one grid: the
foreground-terminal lease admits a single grid at a time.
Everything the refactor was meant to preserve is unchanged and still proven by
the rows that already existed: readiness and attachment, per-pane exclusivity,
close admission, reader close distinct from cancellation, completed and
incomplete replay, exact-request and generation refusals, and pane work that
still receives only its PaneTerminal.
No document names a registry, a live-grid holder or a supervisor, so no stated
contract moved with this.
0 commit comments