feat(ai): add recursive agent composition analysis

[rizumita]

Summary

• Adds recursive fsl-ai agent parsing and IR for scoped nested agents, explicit authority/context grants, output visibility, orchestration edges, failure policy, and lightweight contract rule metadata.
• Adds fslc ai check structural analysis for agent graphs with deterministic agent_ir and graph summaries.
• Preserves existing ai_component hard-contract lowering/replay behavior and extends docs, examples, schemas, skill reference, LSP indexing, and corpus snapshots.

Why

• Goal: Implement Implement recursive agent composition for fsl-ai #157 without introducing a separate sub_agent concept.
• Reason: Nesting should define lexical scope and grant boundaries, while runtime collaboration stays in a separate orchestration/delegation graph.
• Alternative: Lower recursive agents into the kernel immediately. Rejected for this PR because agent composition is structural graph evidence, not a formal proof of LLM semantics.

What Changed

• Parser/IR: top-level and nested agent blocks, bracketed lists, grants, visibility, orchestration, failure policy, trust labels, review gates, and contract { hard { rule ... } } metadata.
• Analysis: validates grant subset boundaries and reports agent_structural_violation findings for authority exceedance, low-trust reachability to high-authority tools, irreversible tools without human approval, review bypass, and sibling visibility leaks.
• CLI: fslc ai check now handles both ai_component and agent; regular fslc check accepts agent files for corpus parse/semantic sweeps without claiming kernel proof.

Blast Radius

• Affected: src/fslc/ai_*, CLI AI commands, LSP raw-tree parser selection, AI finding schema, examples/docs/skills, corpus snapshot.
• Compatibility: Existing ai_component syntax, check result, replay result, and focused tests remain unchanged.

Invariants

Condition	Evidence	If Broken
ai_component remains backward-compatible	tests/test_ai_hard_contract.py	Existing AI hard-contract users regress
Recursive agents do not implicitly inherit parent authority/context	grant rejection and exceedance tests	Nested agents gain unsafe capabilities silently
Agent analysis does not claim formal proof	formal_result: "not_run" tests/docs	Structural findings could be mistaken for verifier proof

Failure Modes

Type	Pattern	Detection
Likely	Parser ambiguity between component output and agent output visibility	recursive agent focused tests
Severe	Child grant exceeds parent boundary but is accepted	invalid grant regression test
Subtle	Visibility/delegation graph conflated with lexical nesting	graph summary and visibility leak tests

Evidence

• .venv/bin/python -m pytest tests/test_ai_recursive_agent.py tests/test_ai_hard_contract.py -q -> 18 passed
• .venv/bin/python -m pytest tests/test_corpus_snapshot.py -q -> 176 passed, 1 skipped
• .venv/bin/python -m pytest -q -> 1004 passed, 78 skipped
• git diff --check

Rollout & Rollback

• Rollout: Normal release; new agent syntax is additive.
• Rollback: Revert this commit. Existing ai_component behavior is isolated from the new structural analyzer.

Review Focus

• src/fslc/ai_parser.py: grammar remains unambiguous and preserves ai_component syntax.
• src/fslc/ai_agent.py: grant subset and graph findings match Implement recursive agent composition for fsl-ai #157 semantics.
• docs/DESIGN-ai-hard.md / docs/LANGUAGE.md: proof boundary is clear.

Related

Closes #157