fix(skills): enforce capabilities.skills in conductor validate

[jrob5756]

Follow-up to #215 (merged). Closes the validator gap flagged in review there.

Problem

ProviderCapabilities.skills documents this contract:

Workflows that declare runtime.skills or per-agent skills: against a provider with skills=False fail validation.

That check was never implemented. Every other capability has a cross-check in config/validator.py — mcp_tools, workflow_tools_passthrough, reasoning_effort, max_session_seconds, working_dir, concurrent_safe — but skills had none.

Verified on main before this change, using the aca provider (which declares skills=False because skill directories are host paths its in-sandbox runner cannot read):

workflow:
  runtime:
    provider: { name: aca, pool_endpoint: https://example.../sessions }
    skills: [conductor]
agents:
  - name: worker
    skills: [conductor]

╭─── Validation Successful ───╮

It passed, then silently dropped the skill content at run time — precisely the regression the descriptor exists to prevent. The blast radius is wider than aca: any third-party or future provider that omits skills=True inherits the safe-looking False default and hits the same silent drop.

After this change:

- Agent 'worker' declares skills=['conductor'] but provider 'aca' does not
  support skills (capabilities.skills=False). Remove the skills, opt out with
  'skills: []', or override the agent to a skill-aware provider.

Changes

• Per-agent check in _check_agent_capabilities. This covers top-level agents and for_each inline agents, since the inline path already routes through the same helper (Extend per-agent capability validation to for_each inline agents #270) — no separate for-each branch needed. An empty list is an explicit opt-out, so only a non-empty skills: errors.
• Workflow-level check over all_llm_agents for inherited runtime.skills, skipping any agent whose own skills: (including []) overrides it. Mirrors the existing max_session_seconds / working_dir inheritance checks.
• AGENTS.md: record that skills is not an allowed experimental carve-out. A provider reaches skills=True either natively (supports_native_skills=True) or through AgentExecutor's eager preamble injection, which is provider-agnostic — so False is only correct when neither path can work.

Tests

11 new tests in tests/test_config/test_validator_capabilities.py:

• per-agent skills: against a skills=False provider → error
• inherited runtime.skills against a skills=False provider → error
• skills: [] opt-out → no error, both standalone and overriding a runtime default
• per-agent provider override in both directions (to and from a skill-blind provider)
• for_each inline agent, both explicit and inherited
• cross-checks against the real AcaRuntimeProvider descriptor, so a future accidental widening of its capabilities is caught here rather than passing silently

_caps() gains skills: True in its defaults so the helper keeps describing a fully-capable stable provider.

Full suite: 4598 passed, 27 skipped. make check clean.

Related

• feat: Conductor Expert — opt-in knowledge base for Conductor-aware agents (#180) #215 — merged; introduced capabilities.skills and the descriptor text this enforces
• Skills: discover user-installed, project, and standalone skills (follow-up to #215) #350 — follow-up covering skill discovery (user-installed, project, and standalone skills), a separate concern from this validation gap