Harden AI plan parsing against truncated LLM JSON responses

[Copilot]

AI plan generation could fail when the LLM returned near-valid JSON with a truncated string/container (e.g., Unterminated string). This change makes parsing tolerant of common truncation patterns so valid plans can still be recovered instead of aborting planning.

• Parsing resilience

  • Added a targeted JSON repair path in AIPlannerClient._parse_json:
    • parse normally first (json.loads)
    • on JSONDecodeError, attempt truncation repair
    • accept repaired payload only if it parses to a top-level object
  • Kept existing strict failure behavior for non-recoverable or structurally invalid payloads.

• Truncation repair helper

  • Added _repair_truncated_json(text) to close:
    • unterminated string literals
    • unmatched { / [ containers
  • Rejects malformed mismatched closers instead of over-correcting.

• Regression coverage

  • Added focused test for a truncated warning-string response to ensure parser recovery returns a valid TestPlan with preserved warning content.

try:
    data = json.loads(stripped)
except json.JSONDecodeError:
    repaired = _repair_truncated_json(stripped)
    if repaired is not None:
        data = json.loads(repaired)  # accepted only when dict