feat: expand scratch codegen and reduction evidence

[wesleysimplicio]

Summary

• add deterministic scratch codegen executors with LibCST/ts-morph paths and executor metrics
• add planner completion caching keyed by provider/model/prompt/settings/template version
• add cache telemetry (hits, misses, puts, hit_rate) for cold/warm evidence
• add rust-axum and php-laravel pilot stacks and include nested stacks/recipes in wheels
• add inline SkillOpt support through optional task required_skill
• add synthetic benchmark harnesses and generated evidence for codegen and static fixers

Refs #32
Refs #33
Refs #37

Evidence

• bench/results_scratch_codegen.md: 50/50 synthetic codegen cases, 100% executor match, 100% codegen share, 0 planner/LLM calls
• bench/results_static_fixers.md: 50/50 synthetic verify-loop cases, 80% fixed before retry, 40% retry-call reduction

Validation

• python -m pytest tests/python -q -> 157 passed, 1 skipped
• focused ruff check / ruff format --check on touched Python files -> passed
• git diff --check -> no whitespace errors; only Windows LF-to-CRLF warnings
• cargo test --manifest-path <rendered rust-axum>/Cargo.toml -> passed
• rendered php-laravel and ran php -l over all rendered PHP files -> passed
• python -m build --wheel --outdir <tmp> -> built successfully and included nested stack/recipe templates

Remaining Release Gates

This PR advances the local implementation and synthetic evidence, but the tracking issues still need external release evidence before closure:

• real 50-scratch corpus with planner/doer credentials
• real LLM baseline pass-rate and latency comparison
• real package-manager execution for Laravel requires Composer on PATH