refactor(orchestration): retire last bare AgentRegistry + canonical capability lookup (#6828)

[mrveiss]

Closes #6828

Thinking Path

Owner green-lit this design refactor today. Wiring map established first (the issue's table had gone stale): the 4-way class rename already landed via #8082/#11430, and the "orphan" claim on orchestration/agent_registry.py is false — it feeds the REAL production tool-dispatch seam (chat_workflow/tool_handler.py) plus orchestrator/API callers. This PR delivers the reopened scope's remaining substance, not docs.

What Changed

• Retired the LAST bare AgentRegistry name (zero-importer compat alias in agent_evolution.py); post-grep exactly 2 class defs remain (AgentRegistryService, AgentRegistryProtocol) — issue AC met.
• Canonical "find an agent that can do X": new runtime-checkable AgentCapabilityLookup Protocol in autobot_shared/agent_registry_protocol.py; AgentCapabilityRegistry is the canonical implementer and now concretely implements AgentRegistryProtocol.update_specializations — the Protocol's first real implementer, verified end-to-end.
• New get_default_capability_registry() singleton accessor — folded 3 ad-hoc registry constructions onto one instance.
• GET /api/agent/agents/available now merges AI Stack agents with the canonical local registry, optional capability filter, and degrades to local-only 200 instead of 503 when AI Stack is down (the one deliberate behavior change; follows bug(agents): /api/agents/status 503 in single_user — agent registries are PG-gated, no unified runtime source #10511 precedent; no frontend consumers — grep-verified; covered by test).
• Stale protocol-module docstring fixed (pre-rename names, false orphan claim).

Verification

• 14 new tests (9 protocol/consolidation, 5 endpoint) pass — agent + independent re-run.
• Touched + adjacent suites: 42 passed; mesh_brain 5; orchestration 708 passed; startup imports 344. 5 pre-existing failures reproduced identically on clean base → filed discovery(tests): orchestration suite has 5 pre-existing failures — checkpoint test rot (_checkpoint_manager rename) + workflow_memory isolation pollution #11754 (not dropped).
• py_compile + flake8 clean.

Follow-up (orphans left tracked, not deleted): AgentEvolutionTracker has a real registry seam now but no scheduler — wire-in issue text provided in #6828 closure; AgentClient/get_agent_client zero external callers.

Model Used

Claude Fable 5 (subagent: general-purpose)

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!