Chore/improve agentic setup

[Jonasb8]

---

Changelog

v0.3.0 — 2026-04-10

Features

• Structural tagging — knowledge records are now automatically tagged based on the files changed in a PR (e.g. migration, schema). Tags are written to the tags: frontmatter field.
• Improved PR comment capture — the action now collects all three types of GitHub comments: top-level review bodies, inline line-level review comments, and general PR thread comments. Previously only top-level review submissions were captured, causing inline comments to be silently ignored.
• Incremental indexing — memex index now skips records whose content hasn't changed since the last run, using a SHA-256 hash of the embed text. Only modified or new records are re-embedded.
• Improved query excerpts — memex query results now show a structured preview drawn from the ## Context and ## Decision sections instead of the first raw paragraph.
• Smarter embed text — embedding uses a cleaned, structured representation (title + context + decision + alternatives + constraints) rather than raw markdown, improving semantic search quality.

Fixes

• changed_files now passed to extractor — the list of files changed in a PR is forwarded to extract(), enabling file-aware extraction.
• test_action.sh stub fix — the gh stub in the local test script was recursively calling itself when passing through --json reviews calls, causing review comments to silently time out and return empty. Fixed by capturing the real gh path before modifying $PATH.

Tests

• Full test suite for cli.py — index/query cycle, embed text building, excerpt extraction, score filtering, and query expansion.
• New TestGetReviewComments tests covering all three comment sources, empty string filtering, partial failures, and full failure fallback.
• Expanded TestChangedFilesWiring — verifies structural tags are derived from changed files and forwarded to write_record.

[github-actions]

Memex detected an architectural decision in this PR but couldn't capture the rationale fully. One sentence: what was the main reason for this approach?