contextpull 0.2.0

Added

• Office ingest: docx, xlsx, pptx with the standard library (Word heading styles, lists and tables; Excel sheets as pipe tables; PowerPoint slides as sections).
• PDF ingest through the pdf extra, headings inferred from font size.
• Hybrid search: ingest --embed-model stores per-section vectors; search(mode="hybrid") fuses lexical and dense by reciprocal rank.
• Streamable HTTP transport: serve --transport http.
• Go SDK (sdk/go): store-native reader and single static server binary, stdio and HTTP, 100% conformance.
• TypeScript SDK on npm as contextpull: store-native reader and server.
• contextpull.eval: ragbisect adapters for the direct-API tool loop and Claude Code headless, with strict and surfaced-ids variants.
• Scenario test layer: a six-format corpus, ten complex use-case tests, a 27-case scenario conformance suite run by Python, Node and Go, and a live agent scenario runner.
• Embedding guide with an access-control example; claude-md command; CONTEXTPULL_EVAL_DEBUG tracing.
• Benchmarks on two corpora under benchmarks/, including the retraction of the first agentic rows.

Fixed

• Eval adapter used one SQLite connection across threads; every tool call failed and the failure was fed to the model as tool output. Runs whose tool calls all fail are now marked as errors and never cached.
• Fence info strings with attributes broke code-block detection.
• - and . as token characters made --no-cache and sentence-final words unsearchable; indexed columns are normalised.
• Grep pre-filtered with the word index and could not find substrings inside tokens.
• Index budget enforced in every mode; compact mode added between flat and hierarchical.
• Sectioner: word-boundary splitting, headings always followed by content, merges re-pack and cannot loop.

Changed

• Default index budget 3,000 tokens. search runs three passes: phrase, all words, any word.