Changed the domain root from showing a latest‑date archive like the site’s main page to a fixed page that reads the README introduction, so search results no longer display the archive as the primary page.
Organized the site‑name structured data, canonical link, hreflang per language, and favicon at the domain root so the source and identity of the site appear clearer in search results.
Excluded the date archive from the search index and sitemap, and configured individual posts to automatically generate search metadata from the document title, category, and summary.
Added a search‑metadata audit script that checks all individual posts and the sitemap for nine languages without manually registering each document path.
Separated the Vercel production deployment from the self‑hosted Mac translation workflow and moved it to run in GitHub‑hosted Actions.
Configured main commits that only change code to be deployed immediately, while changes to README or docs trigger the multilingual translation pipeline first, then automatically build and deploy the latest main.
Re‑checked the SHA‑256 of the committed SQLite and the original 1,570 Markdown files, three tables, and FTS5 row counts, and verified actual JSONL search using queries generated automatically from record contents.
All 60 docs_chunks, 60 api_mapping entries, and 2 label_prototypes dynamic samples were re‑searched and ranked #1 by BM25; the single api_mapping entry with an incorrect SHA‑256 was excluded from search candidates.
Imported the 2d/hexagonal_map project from the MIT‑licensed godotengine/godot-demo-projects commit cae8dc567a56d3e7936f171bcb85f0ccb9634ad0, split its six actual files into 134 chunks, and inspected SQLite JSONL search via the Source Flow Debugger GUI.
Verified that the “Manually editing project.godot” chunk in project settings, the 2D movement chunk with CharacterBody2D, and the scene chunk with TileMapLayer appear in the top results.
Refined the structure table to return records only when the original chunk directly matches a source_api, target_api, or prototype evidence, excluding records that only share negative patterns or decomposed common words.
The Qwen endpoint was temporarily unavailable, so this run only inspected SQLite search results; the Qwen version re‑evaluation will resume after the endpoint is restored.
Since record_embeddings currently has 0 entries, I validated up to SQLite lexical/BM25 search. The full F‑strategy E2E that connects embedding, reranker, and Qwen validator will be checked after the actual index is created.