feat(agent): wire ClaimVerifier + self-improvement write-path + subagent reflection into hot paths (#10602)

[mrveiss]

Thinking Path

Umbrella #10603 Task 6: mature components built with zero callers — wire each into its hot path (never delete), config-gated.

What Changed

• ClaimVerifier (services/grounded_agent.py) — _escalate_to_claim_verifier called from _classify_and_verify_claim for low-confidence/no-match claims; maps result into VerifiedClaim. Flag AUTOBOT_CLAIM_VERIFICATION_ENABLED (default False; OFF ⇒ UNKNOWN, zero overhead).
• Self-improvement write-path (orchestration/workflow_runner.py) — _record_outcome_for_learning fires on workflow success (asyncio.create_task, fire-and-forget): TaskOutcomeJudge.evaluate_task_outcome → persists, then learn_from_outcomes at threshold. Flag AUTOBOT_SELF_IMPROVEMENT_ENABLED (default True — pure plumbing). Closes the loop: orchestrator reads learned strategies at plan time, now writes them post-execution.
• Subagent reflection (execution_strategies/_parallel.py) — _execute_batch_with_reflection routes ready batches through get_subagent_dispatcher().spawn_parallel_tasks() with enable_reflection=True; falls back to direct exec on error. Flag AUTOBOT_SUBAGENT_REFLECTION_ENABLED (default False). Also fixed pre-existing task.timeout_seconds→task.timeout bug in subagent_dispatcher.py.
• FailurePatternDetector (6.2), NPU events (6.4), agent recommendation scores (6.6), lightweight short-circuit (6.7) confirmed already wired.

Verification

17 new tests pass (claim-verifier / self-improvement / subagent-reflection wiring); gated OFF ⇒ unchanged. Black clean. Merged latest Dev_new_gui (incl. #10986) cleanly.

Closes #10602

Model Used

Opus 4.8

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!