io.github.MysterionRise/pubchem-finder

### Added

- Local similarity and substructure search work on first use: until an index is built,
  they search a built-in demo set of 1,932 FDA-approved drugs (PubChem's parent compounds
  of what it links to Drugs@FDA, public domain), built into the data directory in about a
  second on the first search. Results say so with `search_scope: "demo"` and a `note`
  that points to PubChem's live searches and the build command; `index_status` reports
  `search_scope` and `demo_compounds`. Once an index is built, searches use it without a
  restart. `index build --demo` builds the demo index ahead of time and
  `PUBCHEM_FINDER_DEMO=0` turns the fallback off. `scripts/build_demo_set.py` regenerates
  the data.
- Three MCP prompts start common workflows by name (in Claude Code,
  `/pubchem-finder:compound_dossier aspirin`): `compound_dossier` (a cited profile:
  identity, structure, safety, pharmacology, bioassay targets), `find_analogues` (close
  analogues grouped by scaffold and drawn with what changed, optional `threshold`) and
  `triage_structures` (medicinal-chemistry triage of comma-separated SMILES, ranked).
- Two MCP resources: `pubchem://compound/{cid}` (a PubChem compound record as JSON; in
  Claude Code, `@pubchem-finder:pubchem://compound/2244`) and `pubchem://index/status`.
  Errors say what went wrong: an unknown CID is "not found", other failures name their
  cause.
- The index manifest records an index format version. A pubchem-finder that meets an
  index written in a newer format says so and what to do, instead of misreading it.
  Indexes from 2.1 and earlier are format 1 and keep working.

### Changed

- When there is no local index, the message (and `index_status`) gives a command that
  runs from any terminal: `uvx --python 3.14 pubchem-finder==<version> index build --limit 1
  --with-substructure`, with `--home` when the server uses another data directory, and
  what it costs (PubChem's first chunk is now a 323 MB download). Plugin and MCP users
  have no `pubchem-finder` command on their PATH, so the old hint didn't run for them.
  `index build` shows download progress.
- `similarity_search`, `substructure_search` and `pubchem_annotations` ask Claude Code to
  show results of up to 100,000 characters inline (`anthropic/maxResultSizeChars`); by
  default it saves results over 50,000 characters to a file instead, which large searches
  can pass.
- The Claude plugin runs 2.1.1 and passes `--python=3.14` to uvx, like the other
  documented launch commands, so plugin users get the tool annotations and stay off
  Python 3.15.

### Fixed

- In-memory searches no longer risk failing with "Chunk N worker died: ...
  GenericSearch(): incompatible function arguments" or crashing the server. FPSim2 0.7.4
  updates the fingerprint arrays' reference counts unsafely when it splits a search
  across threads, and can free them mid-search; searches now run in the calling thread.
  Seen on Python 3.11, where it made long sessions and batch searches fail; reference
  counts drift on 3.12 too.
- The declared minimum versions work again: Pillow 10.1 (`draw_molecules` needs
  `ImageFont.load_default(size=...)`; with Pillow 10.0 it failed) and typer 0.16 (0.15
  crashed on usage errors with click 8.2 or later). CI now runs the tests against the
  lowest allowed version of every direct dependency.