docs: clarify code-only semantic extraction

[balloon72]

Summary

• Clarify that code files are handled by the local Tree-sitter/deterministic extraction path.
• State that code-only corpora skip the semantic LLM pass.
• Keep semantic extraction scoped to docs, papers, images, and transcripts.

Tests

• python -m pytest tests/test_install_strings.py

Notes

Closes #836

[safishamsi]

Thanks for addressing #836 — the second sentence (code-only corpora skip the LLM pass entirely) is accurate and worth documenting. The first sentence needs a correction though.

Code files are not sent to the LLM semantic extractor in the normal pipeline.

This is only true for code-only corpora. In mixed corpora (code + docs/papers/images), code files do go through the LLM — skill.md line 318 explicitly says: "Code files: focus on semantic edges AST cannot find (call relationships, shared data, arch patterns)." The current wording would mislead users with a mixed corpus.

Suggested replacement:

In a code-only corpus, the LLM semantic pass is skipped entirely — Tree-sitter handles all extraction. In mixed corpora (code alongside docs, papers, or images), code files also go through the LLM, but only to extract semantic edges that AST cannot find (call relationships, shared data, architectural patterns) — never to re-extract imports or structure.

Update the test to match and this is good to go.