fix(orchestrator): allow an out-of-process deadline on the target agent

[shehral]

The problem

SIA bounds every subprocess except the one that matters most:

SHELL_TIMEOUT:  int = 30
EVAL_TIMEOUT:   int = 600
DOCKER_TIMEOUT: int = 3600

But _stream_to_log, the single place the target agent is launched, ends with:

for line in process.stdout:
    ...
return process.wait()          # no timeout

A bare wait() blocks indefinitely. timeout= appears exactly once in orchestrator.py, on the evaluator. So a target agent that hangs on a network call stalls the whole run, not just its own generation.

Why the agent can't bound itself

This came up in practice. One generation wrote an agent that set three internal timeouts and still deadlocked, because all three lived inside the event loop that had stopped. A deadline enforced by the thing being timed is not a deadline. It has to be held outside the process.

The change

An optional SIA_TARGET_TIMEOUT (seconds).

• Unset (default): behaviour is byte-for-byte unchanged. 129 passed, 1 skipped.
• Set: a daemon timer kills the process at the deadline and writes the reason into the run log, so the generation is attributable afterwards rather than just missing.
• Malformed value: ignored, no raise.

Roughly 25 lines, entirely inside _stream_to_log. The Popen call stays in this module's namespace, so it remains patchable in tests as the docstring promises.

Verification

scenario	result
unset + fast process	rc=0 in 0.0s, not killed
=3 + process sleeping 60s	killed at 3.0s, rc=-9
=3 + fast process	rc=0, not killed
="not-a-number"	ignored, no crash
upstream suite, unset	129 passed, 1 skipped

The reason is written to the log, so a killed generation is distinguishable from a crashed one.