codexloop completion detection

CompletionEvaluator in domain/completion.py maps TurnSignals +
CapacityState into a CompletionVerdict.

CompletionVerdict variants

Done  # task fully complete
Continue  # remaining: list[str]
Blocked  # reason: str — human intervention needed

Three-layer evaluation

Capacity rejection always outranks a completion claim. If the turn is
not Available, return Continue regardless of structured output or
markers.

For an Available turn, the layers are:

1. Structured output (StructuredOutput tool call result):

   • blocked_on set → Blocked(reason=blocked_on).
   • complete: true and remaining_work: [] → Done.
   • complete: true but remaining_work not empty → Continue.
   • complete: false → Continue(remaining=remaining_work).

2. Marker on own line: if CODEXLOOP_TASK_FULLY_COMPLETE appears on
   its own line in the final message → Done.

3. Fallback: Continue(remaining=[]).

The marker check uses _marker_on_own_line(message, done_marker): the
string must be the only non-whitespace on the line. Partial matches or
inline occurrences do not count.

Never treat a non-Available turn as Done. Rate-limit windows and
transient errors are waitable; completion is deferred until capacity
returns.