[CI] Add Windows and macOS to the test matrix and publish a platform support statement

Summary

CI currently runs on ubuntu-latest only (all jobs in ci.yml), while
the primary adoption paths — Claude Desktop, Cursor, VS Code recipes —
run heavily on macOS and Windows. Add both platforms to the test matrix
(scoped to keep CI cost sane), fix whatever falls out, and publish an
explicit platform support statement in the docs.

Why this matters

• The gateway's stdio lifecycle (SIGINT/EOF handling), path handling in
  stores (JsonFileArtifactStore, SQLite paths), and console behavior
  (Rich on legacy Windows terminals) are exactly the things that differ
  across platforms — and exactly what a Claude Desktop user on Windows
  hits first.
• "Tested on Windows/macOS" removes a silent adoption objection for
  desktop users and enterprises; an explicit support statement belongs
  next to the Python-version policy in docs/stability.md.
• Recipes that say "works with Claude Desktop" implicitly promise
  macOS/Windows support today, without CI backing.

Proposed scope

• Extend the test job matrix: full Python matrix on Ubuntu (as today),
  plus windows-latest and macos-latest on the newest and oldest
  supported Python (4 added cells, not 8) — fail-fast: false
  preserved.
• Keep the heavier jobs (drift checks, benchmarks, floor-deps,
  examples/demo) Ubuntu-only; run make test (+ make demo if cheap)
  on the new cells.
• Fix discovered issues: likely suspects are path separators in stores
  and fixtures, signal handling differences in _mcp_cli.py
  (SIGINT vs CTRL_BREAK_EVENT), tmpfile semantics in store tests,
  and line-ending assumptions in golden files (.gitattributes may be
  needed).
• docs/stability.md: add a platform support table (Tier 1: Linux;
  Tier 1 after this issue: macOS/Windows for library + stdio gateway).
• README badge/bullet update once green.

Implementation notes

• Workflow: .github/workflows/ci.yml test job (currently
  runs-on: ubuntu-latest with a 3.10–3.13 matrix); convert to an
  os/python-version matrix with include/exclude to get the
  4-cell expansion.
• Golden/byte-identical tests (cards, scorecard fixtures) are the most
  likely cross-platform failures — normalize newlines at comparison
  time rather than weakening the determinism claims.
• tiktoken wheels exist for win/mac; the offline fallback covers any
  gap.
• Windows runners are slow: keep their job to pytest -q without
  coverage to bound wall time.

Acceptance criteria

• CI green on windows-latest and macos-latest for oldest and
  newest supported Python.
• Any platform-specific fixes covered by tests (not just skipped
  on the failing platform — skips need a written rationale).
• Platform support statement published in docs/stability.md.
• CI wall-time impact measured and documented on the PR.

Out of scope

• Platform-specific packaging (MSI/Homebrew).
• Windows-specific gateway features (e.g. named-pipe transports).
• Running the full drift-check suite on non-Linux.

References

• GitHub Actions matrix docs:
  https://docs.github.com/en/actions/using-jobs/using-a-matrix-for-your-jobs
• Current workflow: .github/workflows/ci.yml
• Related: #359 (Python 3.14), docs/stability.md Python policy

Priority: P1 · Effort: M · Impact: Medium

Start with .github/workflows/ci.yml and its test job, then run make test on the oldest and newest supported Python versions for the added Windows and macOS cells. Check platform-specific failures in the store, CLI, signal, fixture, and golden-file tests; done means CI is green, fixes have test coverage, docs/stability.md has the support table, and the README and CI timing are updated.