v0.2.0

The GMAT integration release. Adds GMAT mission execution, parameter
sweeps, and script validation behind an optional [gmat] extra;
credential passthrough for Space-Track and ESA DISCOSweb; a
satellite_metadata tool; and a wider eval suite and CI matrix. Drops
the Smithery publishing path.

Added

• GMAT integration tools behind an optional [gmat] extra —
  gmat_run_mission, gmat_sweep, gmat_execute_script,
  gmat_validate_script, and gmat_read_run_artefact, wrapping
  gmat-run and gmat-sweep. The tools register only when gmat-run
  is importable (pip install astrodynamics-mcp[gmat]), so the base
  install stays GMAT-free. Each call runs gmatpy in an isolated
  subprocess: gmatpy bootstraps a single global Moderator that
  successive runs in one interpreter would corrupt, and its C++ run loop
  holds the GIL, so isolation keeps runs from leaking state into each
  other or blocking the server's event loop (#70, #71, #72, #73, #82,
  #91).
• GMAT script skeletons shipped as MCP resources under
  gmat-skeleton:// URIs — 20 vetted mission archetypes (LEO / GEO /
  lunar transfers, Hohmann, B-plane targeting, finite- and
  electric-propulsion burns, LEO and lunar station-keeping, contact and
  eclipse location, attitude pointing, constellations, control flow) to
  seed LLM-assisted script authoring (#83).
• Credential passthrough for credentialed data sources — environment
  variables for the stdio transport, session-init _meta metadata for
  HTTP. A tool that needs a credential it cannot find raises a typed
  CredentialRequiredError (stable credential_required.<source> code,
  naming the missing fields) rather than failing silently (#67).
• Space-Track as an alternate source for tle_lookup via
  source="space-track", for recent launches and as a CelesTrak
  fallback. Requires Space-Track credentials; CelesTrak stays the
  no-auth default (#74).
• satellite_metadata — physical and provenance metadata (mass,
  bounding-box dimensions, COSPAR ID, launch date and site, operator,
  mission type, decay status) for a NORAD ID, backed by ESA DISCOSweb.
  Requires a DISCOSweb bearer token (#75).
• Eval suite expanded to 40 prompts with GMAT-tool and credentialed-tool
  coverage. GMAT and credentialed prompts skip — rather than fail — when
  their prerequisites are absent, so the suite runs without secrets or a
  GMAT install (#76).
• Multi-model offline eval workflow for release-cut comparison across
  the GitHub Models catalogue, separate from the per-PR gate (#69).
• macOS added to the CI test matrix — now Ubuntu, Windows, and macOS ×
  Python 3.10 / 3.11 / 3.12 (#68).
• Documentation for the v0.2 surface: a GMAT-integration design page
  recording the tool-granularity, HTTP-transport, and [gmat]-extra
  decisions, plus tool-reference and recipe updates (#66, #77).

Changed

• Typed AstroMCPError envelopes now reach the wire on every tool — the
  code, message, and data fields serialise into the MCP error
  response so the LLM consumer can branch on a stable code, and
  non-finite (NaN / Inf) values are rejected at the Quantity
  boundary instead of leaking into a response (#108).
• Data-layer hardening across the adapters: per-event-loop client
  lifecycle with clean shutdown, stale-cache fallback on upstream
  outage, correct Space-Track query encoding, and caching of Horizons
  in-band errors (#110).

Fixed

• Contained the basename fallback in gmat_read_run_artefact so a
  crafted name cannot escape the run's output directory — artefact
  reads are confined to direct children of the run workspace (#109).

Removed

• Smithery publishing infrastructure. The canonical MCPB bundle ships
  with an empty tools[] (clients discover tools over the MCP protocol
  at runtime), but Smithery requires a populated tools[] to List a
  server — an irreconcilable conflict — so the Smithery-specific bundle
  and its publish job are removed. PyPI, the Official MCP Registry, and
  the GitHub Release MCPB asset remain the distribution path (#64).