Module container 

One authority for docker container lifecycle mechanics ([[RFC-0003:C-RUNTIME-WORKFLOWS]]): bounded client invocation with concurrently drained pipes, and confirmed container removal on the launch machine. Callers own their argv assembly, evidence shapes, and operator-facing messages — this module owns only the mechanics that have actually failed in the field: the pipe-capacity deadlock, the deadline kill, and the container that outlives its docker client.

Structs

CommandCleanupEvidence

Enums

BoundedError

Where a bounded invocation failed before producing a wait outcome.

BoundedWait

A bounded invocation that ran to an observable end.

CommandCleanupTrigger

Removal

The outcome of a confirmed-removal attempt.

RemovalFailure

Why a removal stayed unconfirmed; callers format their own messages.

Constants

REMOVAL_TIMEOUT

How long a removal may take before it is reported as unconfirmed: a wedged daemon is a plausible cause of the original deadline miss, and a cleanup path must not hang on its own cleanup.

Functions

docker_device_args

The --gpus argv pair for a device spec (a single index or a comma-joined list). The literal quotes are part of the value: docker’s --gpus parser splits an unquoted device=0,1 on the comma and exposes only the first device (verified on real hardware in v1.5).

remove_container

docker rm -f on the container’s launch machine, always under REMOVAL_TIMEOUT: the SSH client itself runs through the bounded invocation, because the transport’s connect and keepalive options detect dead connections but cannot bound a live connection to a wedged remote daemon. A container the daemon no longer knows is the confirmation, not a failure ([[RFC-0003:C-RUNTIME-WORKFLOWS]]).

run_cleanup_with_bound

run_with_bound

Spawn argv, optionally write stdin_payload, and wait under the operation’s remaining budget and optional attempt cap while draining stdout and stderr concurrently: a child whose output exceeds the pipe capacity would otherwise block writing and never exit, turning a large response or traceback into a timeout.