Implements the v3.1.0 epic (tbd strif-9nyk). The Windows/macOS CI matrix
bead (strif-dzxb) is intentionally deferred to a separate effort.

API additions (agent/IDE ergonomics):
- atomic_write_text() / atomic_write_bytes(): one-call wrappers around
  atomic_output_file() for the common whole-value write case.
- __version__ exposed on the package via importlib.metadata.
- Insertion / Replacement are now NamedTuples with named fields
  (offset/text and start/end/text). Runtime-compatible with positional
  tuples; see note below on static typing.
- HashAlgorithm Literal type on hash_string()/hash_file() params for
  autocomplete; default stays "sha1" (no behavior change).
- Removed the unreachable abbreviate_str/abbreviate_list aliases (they
  were never exported from the public package).

Tests (43 total, up from 26):
- New tests/test_atomic_var.py: AtomicVar set/swap/update, copy vs
  deepcopy independence, updates() context manager + immutable guard,
  truthiness, value_is_immutable, and a real concurrency test (10
  threads x 1000 increments) that would fail without the lock.
- hash_file: binary content + >8KB chunked-read path vs hashlib;
  unsupported-algorithm error.
- New tests/test_ids.py: new_uid charset/length-bits math; timestamped
  uid sorts by creation time.
- atomic_output_file {timestamp} backups don't clobber each other.
- atomic_write_text/bytes round-trip, make_parents, backup.
- Tightened test_string_replace.py: named-field construction, dropped
  redundant assert messages (per testing guidelines).

Tooling:
- Bump softprops/action-gh-release v2.6.2 -> v3.0.0 (Node 24 runtime;
  clears the Node 20 deprecation warning). SHA-pinned.
- Bump astral-sh/setup-uv v7.6.0 -> v8.1.0 in CI and publish. SHA-pinned.

README:
- Document atomic_write_text/bytes; add "Using strif with LLM Agents"
  subsection; note Insertion/Replacement named fields; version banner.

BACKWARD COMPATIBILITY NOTE: The NamedTuple change is fully
runtime-compatible (NamedTuple subclasses tuple, so positional
construction, unpacking, and indexing all still work). However, callers
who statically type a value as list[Insertion]/list[Replacement] and
assign bare tuple literals will now see a type error and should switch
to the named constructors. Removing the abbreviate_* aliases is a
breaking change only for code importing them from the private
strif.strif submodule.

From external review of PR #8:

- hash_file(): remove the `algorithm not in hashlib.algorithms_available`
  precheck. It rejected valid hashlib spellings like "SHA1" and "sha-1"
  (which are NOT in algorithms_available but DO work via hashlib.new),
  and was inconsistent with hash_string(). hashlib.new() still raises
  ValueError for genuinely unknown algorithms, so the contract is
  unchanged for callers.
- pyproject.toml: add [tool.hatch.build.targets.sdist] exclude so the
  published source distribution no longer ships .claude/.tbd/.github/
  .copier-answers.yml agent+tooling state.
- tests/test_ids.py: make the timestamped-uid test deterministic. Assert
  the sortable timestamp-prefix format and base36 suffix plus uniqueness,
  instead of comparing two rapidly-created ids (which could tie on
  timestamp precision and fall through to the random suffix).
- README.md: fix sf.write -> f.write typo in the backup_suffix example.

Not changed: kept the abbreviate_str/abbreviate_list removal (maintainer
confirmed); will be noted in the v3.1.0 release notes.