I2827 — A tool written a moment ago is run once its writer lets go

Give the crate one helper that starts a program and, while the program's file is still open for writing somewhere (ETXTBSY), waits briefly and asks again within a bound; use it at the spawn sites whose tests write a stand-in tool and run it at once.

DONE wave 0 · p1 · implementation profile · parallel safe

Part of failure-is-visible — Every failure the runtime can suffer is bounded, tested or stated, and changes the health it reports.

Objective

Why

CI run 39049547899, job rust (2), failed deploy::observed::tests::the_forge_is_asked_through_its_own_tool_and_every_refusal_is_said at apps/majordomus-cli/src/deploy/observed.rs:514 with 'the GitHub CLI (.../fake-gh) could not run: Text file busy (os error 26)'. The test writes an executable stand-in and runs it; another test thread forked a child while the write descriptor was open, the child holds it until it execs, and Linux refuses to exec a file open for writing. Writing a copy and renaming it does not help (the inherited descriptor holds the inode); waiting does. The same wait existed only test-side or inline (capability/builtin/health.rs, worktree/direnv.rs).

Current state

A product spawn of a freshly written stand-in fails at once on ETXTBSY and the failure is reported as the tool's, turning an unrelated CI job red. This issue covers the GitHub CLI calls of deploy::observed and monitor::github; five more spawn sites whose tests write a stand-in and run it are deferred until after v1.0.0-alpha.1 (listed under non_scope).

Desired state

apps/majordomus-cli/src/process.rs holds retrying_busy, which asks again 50 times 100 ms apart while the start fails with ExecutableFileBusy and returns the last answer; deploy::observed::run (the GitHub CLI call, also reached by the served capability's forge refresh) and monitor::github::run start gh through it.

Scope

apps/majordomus-cli/src/process.rs
apps/majordomus-cli/src/lib.rs
apps/majordomus-cli/src/deploy/observed.rs
apps/majordomus-cli/src/monitor/github.rs

Out of scope

Deferred until after v1.0.0-alpha.1: the spawns at apps/majordomus-cli/src/delivery/public.rs:279 (run_pages), apps/majordomus-cli/src/offload/remote.rs:139, :160 and :175 (Remote::run, run_to_file, spawn), apps/majordomus-cli/src/economics/runner.rs:644 (harness_version) and :370 (spawn_group), and apps/majordomus-cli/src/episodes.rs:358 (the hook tool), each of which a test reaches with a stand-in it has just written
Spawn sites reached only by integration tests through a separate majordomus process
Hooks git executes (continuity's pre-receive stand-ins), which no Majordomus spawn starts

Dependencies

None. This issue is a root of the graph.

Acceptance criteria

On Linux, a program held open for writing by this process is refused with ExecutableFileBusy, and retrying_busy starts it once the writer lets go
A program that cannot start for another reason is answered at once, with that reason
deploy::observed::run and monitor::github::run start the GitHub CLI through retrying_busy

Validation

bin/majordomus plan validate
cargo test --lib -- process:: deploy::observed:: monitor::github:: capability::builtin::served::

Evidence required

etxtbsy_retry_test

Evidence

covers	type	command	result	at commit
etxtbsy_retry_test	ci	validate run 39438896504 at master 9a5a9b77520: rust (1-3) jobs 118683336143 118683336062 118683336401 run cargo test, incl. process::tests::a_busy_start_is_asked_again_until_it_answers_or_the_bound_is_spent (8a82388e0af)	success (the run's only red job is pages, which runs no test)	573df96

Timeline

started: 2026-10-20T22:03:47Z
verified: 2026-10-20T22:04:26Z
completed: 2026-10-20T22:05:30Z

Those three fields, the evidence above and the state of the dependencies are all the status is made of. There is no status field to disagree with them.

Canonical record: .ai/repo/project/issues/I2827.yaml. Read it back with majordomus plan show I2827.