feat: optimize EvidenceForge chat skills

[DavidJBianco]

Summary

• rewrite the pre-existing config, evaluate, generate, scenario, and validate development skills for reliable small-context chat use
• replace monolithic reference loading with focused conditional references and update ChatGPT/Codex installer mappings
• add stable JSON/read-only CLI contracts, source-aware validation diagnostics, config-family and storyline-schema inventories, non-writing composition explanation, and resolved-input replay protection
• add a typed project-overlay registry and close RSAT/TLS overlay validation/runtime gaps
• keep commands/eforge canonical; Git tracks no .agents files

User impact

AI chat workflows now make fewer speculative edits, distinguish mechanical/directly-implied/semantic repairs, preserve authoritative artifacts, and consume substantially less context while retaining focused detail on demand. Configuration and scenario diagnostics are machine-readable and point back to declaring include files.

Validation

• uv run ruff check .
• uv run ruff format --check .
• uv run pytest --no-cov -q: 5,749 passed, 21 skipped
• 100 integrated skill/CLI contract tests
• clean-room forward tests for config, validate, generate, and evaluate
• all 8 freshly installed skills passed skill-creator validation
• wheel and sdist built; all 36 canonical skill files matched source bytes and neither artifact contained .agents

No version bump is included; this targets dev.