Widen distribution matrix to Python 3.10-3.14t across all three OSes

[onlyxItachi]

Summary

Folds the full distribution-matrix expansion into v1.2.0 before its tag:
requires-python >= 3.10, per-version trove classifiers up to 3.14 plus the
official Free Threading classifier, a 3-OS x 6-version CI matrix
(3.10 / 3.11 / 3.12 / 3.13 / 3.14 / 3.14t), and a publish-gate job proving
the exact 3.10 floor claim. Version stays 1.2.0 — this rides into the
existing pre-release, no bump.

Local verification (both ends, full suite — not compileall)

leg	interpreter	result
floor	CPython 3.10.20	206 passed, 16 skipped
free-threaded	CPython 3.14.3t (Py_GIL_DISABLED confirmed True at runtime)	206 passed, 16 skipped
current	CPython 3.14.3	206 passed, 16 skipped

Zero code changes were needed for 3.10 — no stray 3.11-isms surfaced under
the real interpreter. Middle versions (3.11/3.12/3.13) ride on CI as
scheduled.

3.14t risk leg resolved clean: the whole dependency tree installs on the
free-threaded build, and the only compiled deps in the tree (pydantic-core,
rpds-py) ship cp314t wheels for manylinux, macOS, AND Windows — checked
against PyPI release files before widening the matrix. The 3.14t CI legs are
therefore strict (no continue-on-error escape hatch): there is no
blocking dependency to document.

Classifiers (validated, not guessed)

Every string was checked against the official trove-classifiers package
(895 entries) and the final pyproject re-validated programmatically — zero
invalid entries; uv build metadata passes. Settled list (additions):

• Programming Language :: Python :: 3.10 ... :: 3.14 (all five exist)
• Programming Language :: Python :: Free Threading — the official
  free-threading classifier family EXISTS (base + staged
  1 - Unstable/2 - Beta/3 - Stable/4 - Resilient variants). Claimed
  the base classifier only, and only because the suite genuinely passes on a
  GIL-disabled build and CI now runs that leg on every push — no maturity
  stage is claimed.

CI matrix (ci.yml)

os: [ubuntu-latest, macos-latest, windows-latest] x
python: ["3.10", "3.11", "3.12", "3.13", "3.14", "3.14t"], fail-fast: false. The repo's hard-won pinning lesson is preserved verbatim: uv run
pins the SAME --python/--extra as the sync on every step. uv accepting
the 3.14t key was verified locally (uv sync --python 3.14t resolves to
cpython-3.14.3+freethreaded), not assumed.

Publish gate (publish.yml) — additive only

New verify-floor job: full suite + uv build + clean 3.10 venv wheel
install + scripts check + tool-surface assert, every step pinned to 3.10 —
because requires-python >= 3.10 is the claim the package makes to the
world, the gate now proves it on the exact floor. The existing verify job
is byte-untouched; pypi.needs gains the second gate. The whole verify-floor
flow was simulated locally end-to-end before committing (wheel installed into
a clean 3.10 venv; tool surface asserted: 7 tools exact).

Scope discipline

• README: exactly ONE phrase added to the v1.2.0 status entry
  ("Python 3.10–3.14 including free-threaded 3.14t, Linux/macOS/Windows
  CI-verified"). Nothing else touched.
• uv.lock: re-resolved for the widened floor (adds < 3.11 resolution
  markers + floor-only backports) — the lock is committed by repo policy.
• No version bump, no CLAUDE.md edits, no code changes (none were needed —
  flagged explicitly: the pre-release-reviewed source is byte-identical).

🤖 Generated with Claude Code