[codex] add AI stochastic evidence commands

[rizumita]

Summary

Adds the implemented fsl-ai stochastic evidence surface: project-level AI declarations, precomputed eval checks, migration regression comparison, runtime drift checks, and AI compatibility profile generation. Removes proposal docs whose content is now reflected in implementation/design docs.

Why

• Goal: make docs/fsl_ai_stochastic_proposal.md real rather than a stale proposal.
• Reason: stochastic AI quality, AI migration, drift, and compatibility need explicit external-evidence commands that never claim kernel proof.
• Alternatives: putting probability semantics into fslc verify was rejected; the implemented boundary keeps the kernel deterministic and reports formal_result:"not_run" for statistical and observed evidence.

What Changed

• Added src/fslc/ai_project.py and src/fslc/ai_stochastic.py for project-level parsing and evidence evaluation.
• Added CLI commands: fslc ai eval, ai regress, ai compare, ai drift, ai compat, and compat check --include-ai integration.
• Extended ai replay to select a component from project files and detect artifact metadata mismatches plus tool_calls[] events.
• Updated schemas, examples, snapshots, docs, website pages, and skills/fsl/reference.md.
• Renamed AI/domain v0 result identifiers away from *-mvp.v0 wording and deleted superseded proposal documents.

Blast Radius

• Affected: fsl-ai parsing/check/replay CLI, AI result schemas, docs/site, skill reference, corpus snapshot, and domain finding version labels.
• Compatibility: AI result fsl identifiers now use fsl-ai-hard.v0, fsl-ai-agent.v0, and fsl-stochastic.v0. Consumers pinned to the older *-mvp.v0 labels need to accept the new v0 names.

Invariants

Condition	Evidence	Symptom if broken
Statistical and observed AI evidence must not be reported as kernel proof	formal_result:"not_run", schema updates, tests	Users could confuse eval/drift evidence with proved
AI hard-contract kernel checks remain deterministic	existing and extended AI hard-contract tests	fslc verify semantics would change unexpectedly
Proposal content must live in design/reference docs after proposal deletion	docs/DESIGN-ai-hard.md, docs/DESIGN-stochastic.md, docs/LANGUAGE.md, skills/fsl/reference.md	Deleted proposal would remove required user-facing guidance

Failure Modes

Type	Pattern	Detection
Likely	project-level block scanner accepts or drops the wrong proposal construct	tests/test_ai_stochastic.py, fslc ai check examples/ai/support_answer_quality.fsl
Severe	unsupported statistical result is treated as success	ai eval/regress/drift tests and CLI smoke commands
Subtle	schema/result version drift between code, examples, and schemas	schema tests, corpus snapshot, full pytest

Evidence

• .venv/bin/python -m py_compile src/fslc/ai_check.py src/fslc/cli.py src/fslc/ai_expand.py src/fslc/ai_agent.py src/fslc/ai_stochastic.py src/fslc/ai_project.py
• .venv/bin/python -m pytest tests/test_ai_stochastic.py tests/test_ai_hard_contract.py tests/test_ai_recursive_agent.py -q
• .venv/bin/python -m pytest tests/test_corpus_snapshot.py tests/test_ai_stochastic.py tests/test_ai_hard_contract.py tests/test_ai_recursive_agent.py tests/test_db_dialect.py tests/test_domain_dialect.py tests/test_analysis_schema.py tests/test_lsp_index.py -q
• .venv/bin/python -m pytest -q => 1035 passed, 78 skipped
• git diff --check
• CLI smoke: ai check, ai eval, ai replay, ai compat, ai regress, ai drift, ai compare, compat check --include-ai

Rollout & Rollback

• Rollout: merge normally; commands are additive except for documented result identifier cleanup and proposal doc deletion.
• Rollback: revert this commit if downstream consumers require the old labels or proposal files.
• Cannot rollback if: no known data migration or external state change is required.

Review Focus

• src/fslc/ai_project.py: lenient project-level parser boundary.
• src/fslc/ai_stochastic.py: statistical/regression/drift status semantics.
• src/fslc/ai_check.py and src/fslc/cli.py: replay/CLI integration and exit behavior.
• docs/DESIGN-stochastic.md, docs/DESIGN-ai-hard.md, skills/fsl/reference.md: proof vs external-evidence wording.

Unknowns

Item	Resolution	Blocker?
Full evaluator execution and provider sampling are still external to fslc	documented as outside this layer	No
Paired migration statistics are stricter than aggregate comparison	documented as optional external evidence	No

Related

• Follow-up to issues design(stochastic): statistical evidence layer の MVP semantics を固定する #139/feat(compat): AI component を artifact/environment compatibility に統合する #140 and the fsl-ai stochastic proposal cleanup.