Make a rejected plan diagnosable

[justinmclean]

What this changes

Two changes, both about the same failure: a decompose attempt is reported as "response is not valid JSON" and there is no way to see what the model actually said.

_extract_json now decodes the first complete object with JSONDecoder.raw_decode and ignores whatever follows it. json.loads is strict about trailing content, so a model that closes its object and then adds a sentence of commentary -- or leaves a ``` fence unclosed, which defeats the regex path -- was rejected even though a complete, usable plan was sitting right there.

The rejection journal entry gains agent_stdout_path. A 500-character stderr head is enough to see that a plan was rejected and useless for seeing why, because the malformed part of a plan is almost always near its end; the full response now goes to runs//artifacts/ so a rejection is diagnosable after the fact rather than only reproducible. Best effort: a journal without a path, an empty response, or an unwritable directory leaves the entry as it was.

Checklist

• One work item only — nothing unrelated bundled in
• make check is green locally
• Tests cover the change, including refusal paths (the negative cases
  are the product)
• No test, acceptance criterion, or check was weakened to get green
• No changes under specs/ (spec changes go through an issue first)
• Docs updated where behaviour changed (README / USAGE.md)
• Commit subjects are imperative; Generated-by: trailer present if an
  agent wrote the change