Improve language quality prompting

[scawful]

Summary

• supplement task-based skill matching with bounded C++, Python, and TypeScript codebase signals
• strengthen the bundled language-quality rules for lifetimes, typed boundaries, runtime validation, async behavior, and repository-native checks
• remove broad cross-language triggers and recognize CMake and tsconfig.json during codebase discovery
• document the matching behavior and add regression coverage for prompt delivery, trigger precision, and body budgets

Why

Language-quality skills previously depended on the task text naming a language. Generic coding requests could miss relevant guidance, while broad terms such as header, ownership, react, node, and npm could select the wrong skill.

Validation

• 93 passed across skill discovery, session bootstrap, session harness, and agent-context tests on the features base
• 2544 passed, 4 skipped in the full suite with the identical patch on the current main base
• ruff check src/ tests/
• mypy -p afs — 364 source files
• mkdocs build --strict
• matcher smoke ranked cpp-quality, python-quality, and typescript-quality first for representative prompts