vectorgrep v0.3.0

vectorgrep now holds up when several agents use it at the same time, and tells
agents itself when and how to use it.

Added

• Usage instructions for agents: the server sends instructions, and every
  tool description says when to use it — semantic search for questions about
  what code does, grep for exact text, init once per project, index_update
  for routine refreshes. These reach every agent, including Explore and Plan
  subagents, which don't receive a project's CLAUDE.md (#53).
• search_symbols can filter by enum and module (#41).

Changed

• Searching a project that was never indexed now answers "No index found …
  Run 'init'" instead of an empty result, and no longer creates empty tables
  (#52).
• Requests to Ollama and OpenAI time out after 60 seconds (Ollama's
  availability check after 5 seconds) and are retried, instead of waiting
  forever for a server that doesn't answer (#33).

Fixed

• Parallel tool calls, e.g. from subagents sharing one server: concurrent
  index_update calls added the same new files twice or failed with LanceDB
  commit conflicts, and a search started during init or reindex kept
  returning nothing afterwards. Index writes for a project now run one after
  another, and searches wait for them (#52).
• index_update indexed every file a second time when it couldn't read the
  existing file hashes (#32).
• index_status showed wrong file, chunk and symbol counts after
  index_update (#37).
• Projects using different transformers.js models in the same server process
  overwrote each other's model (#30).

Internal

• Test coverage is measured in CI and may not drop below a floor (currently
  84 % statements, 71 % branches, 90 % functions, 85 % lines). New tests cover
  input sanitizing, retries, the embedding cache, result formatting and all
  tool handlers (#54, #55).
• The README installs from npm instead of cloning (#51); CLAUDE.md now holds
  project guidance only, open work is tracked in GitHub issues (#67).

Upgrade notes

• No reindex required — the index format is unchanged since 0.2.0.
• Wrong counts in index_status from earlier versions are corrected the next
  time index_update finds changes, or right away with reindex.