feat: short-chunk coalescing, rerank truncation, query vector cache

[Verdenroz]

Description

A fifth of finance-query's 16k chunks were one or two lines. The gap between two definitions, a doc comment, a #[cfg] line, a pub mod x;, got windowed into its own chunk and embedded on little more than its path header, and those chunks were the top sources for -a questions, which the answer route never reranked anyway. Leading comments and attributes now belong to their definition, spans under four lines merge into a neighbour, @generated files are skipped, and answers go through the same rerank as search. The reranker also reads at most 3000 chars per document, which measured better than full length and not merely faster, and the server keeps the last 256 query vectors so a repeated query skips HyDE and embedding.

Changes

• Extended each tree-sitter definition upward over adjacent comment, attribute, and decorator siblings.
• Added chunker/coalesce.rs: spans shorter than four lines merge forward into the next span, or back into the previous one at the end of a file. Sibling symbols merge under their common parent path.
• Skipped files carrying @generated in their first five lines.
• Added scry index --full, which re-uploads every file regardless of the manifest. A chunker change does not reach existing files otherwise.
• Routed the answer route's sources through the reranker with the same pool and budget as search, raised the per-source snippet to 1600 chars, and added rerank to AnswerRequest. The prompt asks the model to say what is missing instead of guessing.
• Added max_chars to [rerank] (default 3000); RerankClient::document cuts each embedding input there on a char boundary.
• Added scry-server/src/query_cache.rs, a 256-entry FIFO of query vectors used by the search and answer routes.
• Changed changelog.yml to open an auto-merging PR from bot/changelog-update instead of committing to main, minting its token from the same GitHub App soothfast and finance-query use (CHANGELOG_APP_CLIENT_ID variable, CHANGELOG_APP_PRIVATE_KEY secret).
• Documented chunking, truncation, and the cache in docs/search.md and the re-chunk requirement in CLAUDE.md.
• Bumped the workspace version to 0.2.0.

Testing

• cargo test --workspace: 57 core unit tests plus the integration suites pass; cargo clippy --workspace --all-targets -- -D warnings clean. New tests cover trivia attachment, detached comments, one-line mod runs, coalescing rules, @generated skipping, rerank document truncation, and cache eviction.
• finance-query chunk count after scry index --full: 16,331 to 10,438; chunks of one or two lines 3,267 to 33.
• scry eval eval/finance-query.toml, 60 cases, HyDE on, bge-reranker-v2-m3 top 20, each configuration run alone on the iGPU: start of branch 0.717 recall@10 / 0.577 MRR / p50 1804ms; chunker only 0.817 / 0.570 / 3232ms; with 3000-char truncation 0.833 / 0.572 / 2336ms, reproduced twice; second run with the cache warm 0.833 / 0.572 / 1459ms.
• Truncation sweep on the same setup: 1200 chars 0.783 / 0.575 / 1990ms, 2000 chars 0.817 / 0.567 / 2149ms, full length 0.817 / 0.570 / 3232ms.
• Rejected on the same eval and not included: a shorter HyDE prompt (0.817), HyDE on qwen3.5:0.8b (0.767), and jina-reranker-v1-tiny-en (0.767, below no rerank at 0.783).
• The changelog workflow change is untested until it runs on main; it needs the App installed on this repo, the CHANGELOG_APP_PRIVATE_KEY secret, and auto-merge enabled in the repo settings.
• cargo soothfast gate -p scry-core --against-ref origin/main --filter chunk: bench_chunk_rust instructions +1.7% after replacing the per-definition prev_named_sibling walk, which had pushed it to +9.6%.