Source COMMAND_STARTED from the actor output, not the dispatched event

[tobyhede]

Parent

#799. Design record: .scratch/render-execution-unit/spec.md.

What to build

Build the COMMAND_STARTED observation from the command actor's output rather than from the dispatched event.

Its two sibling effects — command-completed and policy-denied — already spread the actor output. COMMAND_STARTED is the only one of the three still reaching into the event for data the output carries. Both arms of the actor's output union carry the rendered command and the display command, the latter documented "for observations".

This is a precondition for emptying the EXECUTE_COMMAND event later, and it stands on its own as consistency.

Acceptance criteria

• The effect is built from the actor output; the event's command fields have no remaining reader in the observation path.
• Emission order is unchanged — started still precedes completed on the stream — and that order is asserted by a test rather than guaranteed by statement position.
• The absent-output case is decided and tested. Today an EXECUTE_COMMAND that produces no actor output still emits a started event for a command that never ran; sourcing from the output emits nothing. Whichever is chosen, a test states it.

Blocked by

None — can start immediately.

[tobyhede]

Sharpening the third acceptance criterion — it was written as "decide the absent-output case", which is a decision handed to the implementer rather than a specification.

The question is narrower than it looked. An EXECUTE_COMMAND that yields no actor output can mean two different things:

• the machine never reached the command-execution state, so no command ever started — a started event here is a lie, and emitting nothing is correct;
• the actor was invoked and threw, so a command may have started and then failed — suppressing the started event loses the only record that it ran.

Establish which of the two the absent-output path actually corresponds to before choosing. If both are reachable they need different handling, and the guard cannot be a bare commandOutput === undefined.

Note the existing code already anticipates absence: both sibling effect pushes are guarded on the output's discriminant, which is what makes this path reachable rather than theoretical.