Reconcile the runtime cognition contract and retire dormant persistent-plan paths

[nicoechaniz]

Problem

DaemonCraft currently describes and executes conflicting cognition/plan contracts:

• agent_loop.py declares that persistent plans are disabled.
• The embodied profile describes each call as a bounded transaction.
• save_plan() and load_plan() are deliberately stubbed.
• The live loop still consumes plan_manifest.json, launches executor/orchestrator paths, and integrates a quest engine.
• Verification predicates use fields that do not match the current VerifySpec contract.

This split contract can contaminate autonomous context and recovery behavior, but it should not be addressed by reviving the 17 legacy plan cards independently.

Related to #24. The lab baseline in #23 remains the first live acceptance gate.

Legacy provenance: t_f22165d6, t_577fad38, t_6b326563, t_3c9dc272, t_f46d8d5b, t_aa9e0c69, t_d2c3ad57, t_bc5856d4, t_9d826e05, t_2dbc8ba5, t_a8426ce0, t_a76f5bba, t_aa90840b, t_95ae07fc.

Goal

Define one authoritative ownership contract for L4/Hermes, agent_loop, the embodied service, ephemeral manifests, the quest engine, and L2 reflexes; then remove or repair the contradictory runtime paths.

Acceptance criteria

• Document one ownership/lifecycle contract for L4 goals, bounded orchestration, quest phases, reflex preemption, verification, retry, and termination.
• Remove the obsolete persistent Plan/Step/save_plan/load_plan path rather than reactivating it.
• Decide whether PlanManifest/orchestrator wiring is removed or retained strictly as bounded, ephemeral orchestration.
• If retained, fix the BLOCK_PLACED and AREA_CLEAR verification-field mappings and cover real predicates with tests.
• Ensure quest phase changes invalidate incompatible active manifests with structured escalation.
• Prove retries and resumes are bounded and cannot become a second autonomous loop.
• Align agent_loop.py, SOUL/profile templates, architecture docs, README, and local setup documentation with the decision.
• Feed the resulting contract into the controlled lab↔autonomous comparison in Diagnose autonomous-mode context divergence and action thrash #24.

Non-goals

• Rebuilding the old persistent planning system.
• Creating one Issue per historical plan card.
• Increasing turn or retry limits as a substitute for a coherent lifecycle.
• Starting autonomous mode before Restore the known-good interactive lab embodiment baseline #23 establishes the known-good lab baseline.

[nicoechaniz]

Cross-link: semantic no-progress must be part of the cognition lifecycle contract

The TUI negative-control trace recorded in #24 shows that exact-repeat detection is insufficient when the agent varies tool arguments while testing the same failing hypothesis. The authoritative lifecycle contract defined here should therefore expose and own at least: active objective, exact success sensor, hypothesis family, observation budget, objective delta, mutation permission, and abort reason.

Evidence and measured trace: #24 (comment)

#24 remains the owner of the lab↔autonomous comparison and behavioral acceptance. #28 owns the runtime contract that prevents L4, orchestration, quest, and reflex layers from advancing or mutating incompatible objectives.