unstd 1.0.5__tar.gz → 2.0.0__tar.gz

unstd-2.0.0/CHANGELOG.md

@@ -0,0 +1,385 @@
+# Changelog
+
+<!-- towncrier release notes start -->
+
+## [2.0.0] - 2026-09-27
+
+### Breaking
+
+This is a 2.0 surface pass: every change below renames or removes a public name,
+and each old name fails at import rather than changing meaning in place.
+
+- `serde.b64` is one verb per direction. `b64s` / `b64u_s` → `encode(data, url=,
+  pad=)`, `b64d` / `b64u_d` → `decode(s, url=)`, `b64text` / `b64u_text` →
+  `decode_text`, `b64_json` / `unb64_json` → `encode_json` / `decode_json`.
+  `pad=False` is new - the unpadded base64url JWTs and bearer tokens use, which
+  callers were `rstrip("=")`-ing by hand - and url-safe `decode` now accepts
+  padded and unpadded input alike. Standard `decode` stays strict.
+- `ids.hash` is gone; it was three second names for `crypto.digest`. Content
+  hashes are `digest.hex(data)`, short ids `digest.hex(data, n)`, raw bytes
+  `digest.raw(data)` (was `sum_` / `hex_` / `hex_n` / `content_hash`).
+  `digest.of_parts(parts)` is `digest.Stream(parts).raw()`; `Stream.update_all`
+  went with it, and `Stream.sum_` / `hex_` are `raw` / `hex(n=None)`.
+  `digest.KDF_BYTES` folded into `DIGEST_BYTES` (both were 32).
+- `crypto.token.mac` / `mac_bytes` → `mac_hex` / `mac_raw`, so the return type is
+  in the name as it is in `digest`. `verify_mac` takes a tag in either form.
+  `equal` lives in `token` only, and `token.opaque` is
+  `rand.crypto.token_urlsafe` - it was a pass-through.
+- `time.dateutil` is `time.zoned`, holding only what needs `whenever`:
+  `wall` / `parse` / `now` / `instant` / `to_utc` (were `zoned_wall` /
+  `parse_zoned` / `zoned_now` / `parse_instant` / `zoned_to_utc`). The pure-stdlib
+  half - `parse_iso`, `parse_utc`, `parse_date`, `from_epoch_s`, `from_epoch_ms` -
+  moved to `timeutil`, so none of it needs the `time` extra any more.
+- `timeutil.span(start, end=None, *, unit, assume_utc=False)` replaces
+  `span_hours`, `span_minutes`, `age_days`, `minutes_until`, `minutes_since`, and
+  `hours_between`. Those five disagreed about whether a floating stamp was UTC
+  depending on which unit you asked in; `span` is strict unless told otherwise.
+  `iso_range`, `lookback`, `days_range`, and `day_offset` are removed.
+- `timeutil.iso(dt=None, *, ms=False)` replaces `utcnow_iso`, `utcnow_iso_ms`,
+  and `iso_fmt`; `today(utc=True)` replaces `utcnow_date_iso`; `today_iso` and
+  `date_fmt` are removed (`.isoformat()`); `human(tz="")` replaces `human_now`
+  and `human_now_tz`.
+- `timeutil.parse_date` returns `None` on bad input like every other parser
+  here, instead of raising.
+- `serde.structs.codec(...)` is removed; construct `structs.Codec(...)`.
+- `audio.wavx` drops the stutter: `read_wav` / `write_wav` / `read_blocks` /
+  `WavWriter` → `read` / `write` / `blocks` / `Writer`.
+- `fs.read_bytes`, `fs.read_text`, and `fs.iter_dir` are removed. They measured
+  slower than the `pathlib` methods they wrapped.
+- `proc.TimeoutError` is removed. It aliased `subprocess.TimeoutExpired`, which
+  does not subclass the builtin of that name.
+- `iters.take(iterable, n)` takes the iterable first, like `chunked`, `windowed`,
+  and `nth` beside it. The old order fails loudly (`'int' object is not
+  iterable`), never silently.
+- `iters.batched_with_key` is `iters.runs` - it groups consecutive runs by key;
+  it never batched. Its module is `iters.consecutive`.
+- `iters.partition(iterable, pred)` takes the iterable first, like every other
+  combinator in `iters`. The old order raises `TypeError` at the call.
+- `serde.ndjson.read(fp)` / `write(fp, rows)` → `load(fp)` / `dump(rows, fp)`,
+  the same shape as `jsonx.dump(obj, fp)` and `toml.dump(obj, fp)`. `dump` still
+  returns the record count.
+- `timeutil.now_tz(tz)` requires its zone. Called bare it was a second spelling
+  of `utcnow()`.
+- `toml`'s write half is `tomlkit`'s own functions rather than wrappers around
+  them, so `dumps` / `dump` accept `sort_keys=` and each name carries tomlkit's
+  types. On a base install each is a placeholder that raises naming the extra,
+  decided at import rather than re-checked per call.
+- `unstd.crypto` no longer re-exports the `blake3` class; it was the backend,
+  not a surface. The guard moved into `digest` / `token`, so `crypto.tls` - pure
+  stdlib - now imports on a base install instead of failing on a missing BLAKE3.
+- `crypto.tls.client_context(certfile=...)` alone loads the chain, which is the
+  combined cert-plus-key PEM most bundles ship as; it used to be ignored and the
+  client dialed without mTLS. A `keyfile` with no `certfile` raises `ValueError`.
+- The `ids` extra no longer pulls `blake3`; nothing under `ids` has used it since
+  `ids.hash` folded into `crypto.digest`. Depend on `unstd[crypto]` for digests.
+- `unstd.containers.guard` is gone. Its placeholder and `toml`'s copy of it are
+  one private helper, and every missing-backend name now raises the same
+  `unstd.<module>.<name> requires the '<extra>' extra` message.
+
+### Performance
+
+- `jsonx.dumps` 416 → 264 ns, `dumpb` 375 → 200 ns, `loads` 251 → 212 ns on a
+  small dict (raw orjson: ~180 ns). The fast path no longer detours through a
+  checking wrapper, and `loads` names its return type with an annotation instead
+  of a runtime `typing.cast` call.
+- `ndjson.dumpb` / binary `write` let orjson append each newline
+  (`OPT_APPEND_NEWLINE`) instead of concatenating a second `bytes` per row.
+- `rand`'s scalar draws are stdlib `random` on every install. Routing
+  `random()` / `choice` / `shuffle` through numpy's `Generator` one value at a
+  time was 5-7x slower than the stdlib draw; numpy still fills the bulk arrays.
+- `timeutil.wall` / `mono` / `mono_ns` / `perf` and `rand.crypto.*` are the stdlib
+  functions bound directly, with no forwarding frame.
+- `timeutil.iso` formats through `isoformat` (~2.6x faster than `strftime`), and
+  `epoch_ms` is exact integer math on `time_ns()`.
+- `structs.Codec.encode` / `decode` are msgspec's own bound methods.
+- `digest.hex(data, n)` asks the XOF for `n` bytes rather than slicing a full
+  hex string, and checks the width inline.
+- `jsonx.canonical` reuses one prebuilt encoder instead of letting `json.dumps`
+  construct a fresh one per call: ~1.25 µs → ~0.9 µs on a small dict, same bytes.
+- `b64.encode(url=True)` rides pybase64's `altchars` straight to `str`: 293 →
+  192 ns for 64 bytes unpadded.
+- `token.mac_hex` / `mac_raw` check the key inline - 943 → 674 ns - and `verify`
+  follows (2.1 → 1.6 µs).
+- `ndjson.loads` parses a clean blob in one comprehension over the backend
+  decoder and strips a line only when a direct parse fails: 338 → 187 µs per
+  1k records, level with a bare orjson loop. A bad line still re-walks the slow
+  path, so the error and its line number are unchanged.
+- `iters.windowed(step=1)` is staggered `tee` views zipped together, a C-level
+  walk: 4x on 10k items. `unique_everseen` skips the key call when there is no
+  key, and `last` indexes exact builtin sequences without an ABC probe (2x).
+- Pure forwarders are now the function they forwarded to: `timeutil.utcnow`
+  (`partial(datetime.now, UTC)`, 132 → 96 ns), `uid.parse` / `uid.derive`
+  (`UUID` / `uuid5`), `iters.flatten` (`chain.from_iterable`, 2x), `zoned.now`
+  (`ZonedDateTime.now`), and - with the `text` extra - `fuzz.ratio` /
+  `distance` / `jaro_winkler` as the rapidfuzz kernels (~25-40% per call). Clock
+  freezers that patch at the C layer (`time-machine`) still see the bound
+  references. The pure-Python fallbacks are held to the kernels by a property
+  test, since with the extra installed nothing else runs them. The fuzz scorers
+  take their two strings positionally.
+- `wavx.wrap_pcm` packs the 44-byte header with one `struct` call - byte-identical
+  to `wave`, pinned across widths, channel counts, and partial frames - 2.7x
+  faster on 32 KB.
+- `iters.partition` is the current `itertools` recipe - verdicts teed beside the
+  values and split with `compress` - so `pred` runs once per element instead of
+  once per side: 533 → 459 µs on 10k items with a trivial predicate, and half the
+  predicate calls whatever it costs.
+- `ndjson.dumps` joins orjson's row bytes and decodes once, and a text-mode
+  `dump` decodes each row's bytes rather than re-entering `jsonx.dumps`: ~20%
+  on both (399 → 317 µs and 547 → 421 µs per 1k rows).
+- `clone.asdict` tests dict keys and values for atoms inline instead of paying a
+  call per leaf: 5.2 → 3.8 µs on a dict-heavy record.
+- `rand.choices` unboxes the drawn index array once instead of calling `int()`
+  per element: 379 → 209 µs for 10k draws.
+
+### Fixed
+
+- `text.fuzz.best_match` / `extract` / `cdist` keep a choice scoring exactly at
+  `score_cutoff`, as documented. rapidfuzz's own cutoff drops boundary pairs
+  through a float round-trip (about 15% of random short pairs on 1.1), so every
+  fast path now scores in full and applies `>=` itself - at no measurable cost.
+  `cdist` also returns float64, so its scores equal `ratio`'s exactly.
+- `fs.atomic_write` / `atomic_writer` no longer flip the process-wide umask to
+  read it when creating a new file. Setting it to 0 and back left a window in
+  which any other thread creating a file got mode `0o666`/`0o777`. The temp file
+  is now opened at `0o666` so the kernel applies the umask, and an existing
+  destination's mode is still copied over.
+
+- `rex` accepts its own compiled patterns. On the irgx path `rex.compile(p)` and
+  `rex.search(p, s)` raised `TypeError` for a `p` this module had just returned,
+  because only stdlib's `re.Pattern` was recognized; stdlib takes either.
+- `rex.purge` clears irgx's pattern cache too, not just stdlib's.
+- A pattern irgx declines (backreferences, lookaround) is no longer re-attempted
+  on every call. irgx caches what it compiles, not what it refuses, so each
+  `rex.search(r"(\w+)\s+\1", ...)` paid a full compile attempt plus a raise -
+  about 1.8 ms, against 0.2 µs for stdlib.
+
+### Changed
+
+- `rex` keeps a bounded front cache like stdlib's `re._cache`, and hands flags to
+  irgx as its keyword arguments instead of building an inline `(?imsx-u)` prefix
+  through `RegexFlag` arithmetic. A module-level `rex.search` drops from ~5 µs to
+  ~0.8 µs - now twice as fast as `re.search` - and a compiled pattern's
+  `.pattern` is the caller's source again rather than carrying the prefix.
+- `rex.findall` rewrites irgx's `None` for an absent group only when one is
+  actually present, found by a C-level containment scan instead of rebuilding
+  every row. A two-group `findall` over 10 KB drops from ~940 µs to ~240 µs.
+- The `rex` extra requires `irregex>=2.6.0`. Below it, a replacement template's
+  octal escapes (`\07`, `\101`) and unknown non-letter escapes rendered different
+  text than stdlib with no error; 2.6.0 reads them the way `re` does. It also
+  makes a group-reading `rex.sub` and a grouped `rex.split` one call into the
+  engine instead of one per match, and stops a capped `subn`/`split` walking past
+  its cap - on 10 KB, templated `sub` goes from ~7x slower than `re.sub` to
+  about even, and `subn(..., count=5)` from ~130x to ~1.7x.
+
+## [1.1.0] - 2026-09-27
+
+### Fixed
+
+- `clone.deep` returns the types it was given. Its msgspec round-trip decoded a
+  nested `StrEnum` as `str`, an `IntEnum` as `int`, and an `OrderedDict` or
+  `defaultdict` as a plain `dict` - losing the factory with it - and the equality
+  check meant to catch a lossy round-trip passed every one, because those values
+  compare equal to what came back. `deep` is now an exact-type walk that never
+  leaves Python objects, so there is nothing to lose.
+- `python3 -m bench update --group X` no longer erases every other group's floors.
+  A narrowed run now merges into the baseline instead of replacing it.
+
+### Changed
+
+- `clone.deep` is pure stdlib and faster where callers actually spend time:
+  4.1x over `copy.deepcopy` on a small dict (was 1.2x, slower than deepcopy on
+  CPython 3.14), 2.2x on pydantic models (was a failed encode then deepcopy), and
+  4.4x on large JSON trees (was 5.8x on CPython 3.13 - the price of the fix above;
+  on 3.14 the two tie). Dataclasses, pydantic models, and `msgspec.Struct` are
+  rebuilt the way `deepcopy` rebuilds them, with the walker standing in for its
+  recursion; any other node is handed to `deepcopy` on its own while its siblings
+  keep walking. The `clone` extra no longer installs anything and stays declared
+  so existing `unstd[clone]` requirements resolve without a warning.
+
+### Added
+
+- `clone.asdict`: `dataclasses.asdict`'s result on the same walker, 10x faster on
+  a dataclass tree with `datetime` leaves. A custom `dict_factory`, a
+  non-dataclass argument, or a namedtuple / container subclass in the tree hands
+  the call to the stdlib, so the answer is always the stdlib's.
+
+## [1.0.5] - 2026-09-01
+
+### Changed
+
+- Prose and fixtures name generic things rather than one deployment's. The
+  NDJSON docstring described the format by pointing at two artifacts nobody
+  outside their origin has ever seen, and a crypto fixture put a private service
+  in a JWT audience - so the docstring names what NDJSON is actually used for,
+  and the fixture says `example` like its sibling suite already did. No code
+  paths move.
+
+## [1.0.4] - 2026-08-31
+
+### Fixed
+
+- The extra floors sit at the minor series rather than the patch we happened to
+  develop against, and CI now proves them. 1.0.3 relaxed `==` to `>=` but left
+  each floor on an exact patch, so `unstd[model]` still rejected an application
+  one pydantic bugfix behind - a resolver error that reads as the application's
+  fault. The API a backend offers changes at the minor, which is where a floor
+  belongs.
+
+### Added
+
+- A `floors` CI job resolves `--resolution lowest-direct` and runs the whole
+  suite against the oldest version every floor admits. A floor is a promise that
+  the library works that far back, and nothing was testing it: the rest of the
+  matrix resolves the newest of everything, so the floors could have been fiction
+  indefinitely and only a downstream with an older pin would have found out.
+
+## [1.0.3] - 2026-08-31
+
+### Fixed
+
+- Every extra declares a floor rather than an exact `==` pin. The pins were
+  inherited from the monorepo this package was extracted from, where an exact
+  pin is correct because a monorepo is the application. Published, they meant an
+  application already on `pydantic==2.13.4` could not install `unstd[model]` at
+  all - the resolver reports the conflict against the application, not against
+  us - and two libraries pinning patches are simply not co-installable. The
+  floors are the versions each extra is tested against; a consumer that wants an
+  exact version has its own manifest and lockfile to say so in.
+
+## [1.0.2] - 2026-08-31
+
+### Fixed
+
+- The package ships `py.typed`. Every surface here is annotated and CI has been
+  running `mypy --strict` over all of it, and none of that reached a consumer:
+  PEP 561 says to ignore the annotations of a package with no marker, so a
+  strict downstream saw `jsonx.dumps` as `Any` and wrote a suppression per call
+  site to get its build green.
+
+## [1.0.1] - 2026-08-31
+
+### Added
+
+- `tls.client_context` takes a `minimum_version`, defaulting to the same TLS 1.3
+  it always asserted, and `tls.LEGACY_MINIMUM_VERSION` names the 1.2 floor for a
+  peer not yet shown to negotiate 1.3. Without it, a service that dials one such
+  peer had to hand-roll a second context — which is where the parameter this
+  module refuses to have gets added. Anything below 1.2 raises instead of
+  building a context.
+
+### Fixed
+
+- `unstd.__version__` is read from install metadata instead of being a literal
+  beside the one in `pyproject.toml`. The two drifted on the first patch release
+  and the gate caught its own package's version, so the tag was cut and nothing
+  shipped; there is now one place to edit.
+
+## [1.0.0] - 2026-08-31
+
+### Added
+
+- First public release of the full library: serde, ids, crypto, time, model,
+  rex, containers, rand, pack, clone, text, audio, toml, proc, fs, and iters.
+  Base install stays pure stdlib; every accelerated backend is an optional extra.
+- `crypto.tls` — the verifying TLS client context, with no parameter for
+  skipping verification, split out of the module that used to hold it. Its mTLS
+  and pinned-root tests now actually run: they mint a throwaway self-signed
+  certificate, which needs `cryptography`, which nothing had declared, so all
+  three had been skipping rather than passing.
+- `digest.legacy_blake2b` — the one deliberate exception to "one digest",
+  for an identifier already defined in terms of BLAKE2b that cannot change
+  algorithm without renaming every value derived from it. Keeping that case
+  inside the seam is what stops it becoming a reason to import `hashlib` next
+  door and grow a second digest for new work too.
+- `ruff` and `mypy --strict` now run in CI, and `release.yml` calls the CI
+  workflow as a gate rather than keeping its own copy of one — so the bar a tag
+  clears is the bar a pull request clears, by construction. Neither tool had ever
+  run over this package; the first pass found 26 type errors.
+- A `proto` extra. `timeutil`'s protobuf `Timestamp` bridge imported
+  `google.protobuf` with nothing declaring it, so the two `proto_timestamp`
+  functions raised `ModuleNotFoundError` on any install that had not happened to
+  pick protobuf up sideways. The import is guarded now and names the extra.
+- `bench` — a harness behind the speedup numbers the READMEs print. Each case is
+  one accelerated call against the stdlib call it stands in for, reported as a
+  ratio, with `bench/baseline.json` pinning a floor per case that `update` may
+  raise and never lower. It refuses to grade a missing backend as a pass, and
+  refuses to report a ratio between two functions that disagree on the answer.
+- The suite runs on a base install. It could not before: four surfaces have no
+  stdlib fallback and raise at import, which pytest reports as a collection
+  error, so `pip install unstd && pytest` aborted before running anything — and
+  the guarded-fallback paths, which are the reason the library exists, had never
+  been exercised by anything. `tests/conftest.py` now decides collectability by
+  probing importability, and the fallback CI leg runs the whole suite (477 tests)
+  instead of a hand-maintained subset.
+- Tests for `ndjson`, which had none and 0% coverage; for `timeutil`, which had
+  none; and for the package's own description of itself — the top-level docstring
+  listed four module groups while sixteen shipped, so the prose, the README table
+  and the directory listing are now held to each other, along with the
+  zero-dependency claim.
+
+### Changed
+
+- `crypto` no longer projects an external algorithm table. The widths it
+  enforces are BLAKE3's own — 32 bytes of digest, 32 of key, 32 of tag — so they
+  live beside the calls they bound (`digest.DIGEST_BYTES` / `KDF_BYTES`,
+  `token.KEY_BYTES` / `TAG_BYTES`) instead of being imported from a generated
+  registry. `derive_key` takes any context string and says plainly that
+  registering one is the caller's job; nothing about the derivation changed.
+- The crypto known-answer vectors are now the BLAKE3 team's own published
+  `test_vectors.json`, vendored verbatim and attributed in NOTICE. All three
+  modes — plain hash, keyed hash (the MAC), and `derive_key` — are checked
+  against the authors' answers across 35 input lengths spanning the chunk and
+  subtree boundaries, where the previous fixture carried seven hand-picked rows.
+- `rex` is now backed by [irregex](https://github.com/The-Billy-Company/irregex)
+  (`import irgx`) instead of `google-re2`; the extra installs `irregex>=2.3.0`.
+  The surface, the stdlib-`re` fidelity contract, and the degrade-to-`re`
+  fallback for backreference/lookaround patterns are unchanged.
+- The linear-time guarantee now covers three families that previously degraded to
+  backtracking `re`: `re.VERBOSE`, `re.ASCII`, and `bytes` patterns. `bytes`
+  patterns compile byte-oriented, so `rb"\w+"` stops at `caf` in `café` exactly
+  as stdlib does.
+- Two divergences that were always true of `rex` are now documented and pinned by
+  tests rather than left to be discovered: `$` matches the absolute end of the
+  text (not the position before a trailing newline, which is Perl's and stdlib's
+  rule), and POSIX bracket expressions like `[[:alpha:]]` are a real class to the
+  linear engine but a nested-set typo to stdlib.
+
+### Fixed
+
+- Nine `rex` signatures annotated their `pattern` and `string` parameters as
+  `DOTALL` — the regex flag — instead of `S`, the type parameter they declare.
+  Ruff's `FURB167` rewrites `re.S` to `re.DOTALL`, marks the fix safe, and had
+  spliced its replacement over every PEP 695 `S` in the file. The rule is now
+  disabled here, with the reason written down next to it.
+- `timeutil.human_now_tz` ignored the timezone it was passed. It built the right
+  instant and then called `.astimezone()` with no argument, which converts back
+  to local time, so every caller got local time under another zone's name.
+- `timeutil.iso_fmt` stamped a literal `Z` on whatever it was handed. A
+  `datetime` in any other zone came out labelled UTC while still carrying its
+  original wall-clock digits — a silent offset, in the format used for the wire.
+  Aware datetimes are converted to UTC first now; naive ones are relabelled.
+- `timeutil`'s human format string used `%-d` and `%-I`, which are a glibc/BSD
+  extension. `strftime` raises `ValueError` on both under Windows, so every
+  human-facing timestamp in the module was a hard failure on one of the three
+  platforms the package claims to support.
+- A `bytearray` or `memoryview` pattern now raises `TypeError`, matching what
+  stdlib `re` does with one. The backends accept any buffer as a pattern, so
+  `rex` was quietly accepting patterns its own twin refuses — being a superset is
+  a divergence too, and code written against it broke the moment the extra was
+  absent and the `re` fallback took over.
+- `sub` and `subn` now type a callable `repl` as receiving either backend's match
+  object, which is what has always happened at runtime. The two are
+  interchangeable for everything a replacement function reads (`group`, `groups`,
+  `groupdict`, `span`, `expand`, `start`, `end`); only `.pos` and `.endpos` are
+  stdlib-only.
+- `findall` is typed `list[S | tuple[S, ...]]` rather than a union of two list
+  types, so the element type survives instead of collapsing (typeshed spells the
+  same fact `list[Any]`). Runtime results are unchanged.
+
+## [0.0.1] - 2026-08-31
+
+Parked a good name.
+
+### Added
+
+- Name reservation on PyPI.

{unstd-1.0.5 → unstd-2.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: unstd
-Version: 1.0.5
+Version: 2.0.0
 Summary: stdlib-faithful accelerated replacements behind optional extras
 Project-URL: Homepage, https://github.com/The-Billy-Company/unstd
 Project-URL: Repository, https://github.com/The-Billy-Company/unstd
@@ -24,14 +24,12 @@ Provides-Extra: audio
 Requires-Dist: numpy>=2.5; extra == 'audio'
 Requires-Dist: soundfile>=0.14; extra == 'audio'
 Provides-Extra: clone
-Requires-Dist: msgspec>=0.21; extra == 'clone'
 Provides-Extra: containers
 Requires-Dist: immutables>=0.21; extra == 'containers'
 Requires-Dist: sortedcontainers>=2.4; extra == 'containers'
 Provides-Extra: crypto
 Requires-Dist: blake3>=1.0; extra == 'crypto'
 Provides-Extra: ids
-Requires-Dist: blake3>=1.0; extra == 'ids'
 Requires-Dist: uuid-utils>=0.17; extra == 'ids'
 Provides-Extra: model
 Requires-Dist: pydantic-settings>=2.15; extra == 'model'
@@ -43,7 +41,7 @@ Requires-Dist: protobuf>=7.36; extra == 'proto'
 Provides-Extra: rand
 Requires-Dist: numpy>=2.5; extra == 'rand'
 Provides-Extra: rex
-Requires-Dist: irregex>=2.3.0; extra == 'rex'
+Requires-Dist: irregex>=2.6.0; extra == 'rex'
 Provides-Extra: serde
 Requires-Dist: msgspec>=0.21; extra == 'serde'
 Requires-Dist: orjson>=3.12; extra == 'serde'
@@ -85,15 +83,15 @@ python3 -m pip install 'unstd[serde,ids,crypto]'
 | Group | Extra | Backend / fallback |
 |---|---|---|
 | [`serde`](src/unstd/serde/README.md) | `unstd[serde]` | orjson / msgspec / pybase64 → stdlib `json`/`base64` |
-| [`ids`](src/unstd/ids/README.md) | `unstd[ids]` | blake3 / uuid-utils → `hash` requires the extra; `uid` falls back to stdlib UUIDv7 |
+| [`ids`](src/unstd/ids/README.md) | `unstd[ids]` | uuid-utils → `uid` falls back to stdlib UUIDv7 |
 | [`crypto`](src/unstd/crypto/README.md) | `unstd[crypto]` | blake3 (required; hashlib ships BLAKE2, a different algorithm) |
-| [`time`](src/unstd/time/README.md) | `unstd[time]` | `timeutil` is pure stdlib; `dateutil` requires `whenever`, and `timeutil`'s protobuf `Timestamp` bridge requires `unstd[proto]` |
+| [`time`](src/unstd/time/README.md) | `unstd[time]` | `timeutil` is pure stdlib; `zoned` requires `whenever`, and `timeutil`'s protobuf `Timestamp` bridge requires `unstd[proto]` |
 | [`model`](src/unstd/model/README.md) | `unstd[model]` | pydantic (required) |
 | [`rex`](src/unstd/rex/README.md) | `unstd[rex]` | irregex → stdlib `re` |
 | [`containers`](src/unstd/containers/README.md) | `unstd[containers]` | sortedcontainers + immutables (required) |
 | [`rand`](src/unstd/rand/README.md) | `unstd[rand]` | numpy PCG64 → stdlib `random`; `crypto` is always `secrets` |
 | [`pack`](src/unstd/pack/README.md) | `unstd[pack]` | numpy → stdlib `struct` |
-| [`clone`](src/unstd/clone/README.md) | `unstd[clone]` | msgspec → `copy.deepcopy` |
+| [`clone`](src/unstd/clone/README.md) | — | stdlib; exact-type walk, `copy.deepcopy` per foreign node |
 | [`text`](src/unstd/text/README.md) | `unstd[text]` | rapidfuzz → stdlib `difflib` |
 | [`audio`](src/unstd/audio/README.md) | `unstd[audio]` | soundfile + numpy → stdlib `wave` |
 | [`toml`](src/unstd/toml/README.md) | `unstd[toml]` | read is `tomllib`; write requires tomlkit |
@@ -101,7 +99,7 @@ python3 -m pip install 'unstd[serde,ids,crypto]'
 | [`fs`](src/unstd/fs/README.md) | — | stdlib `os` |
 | [`iters`](src/unstd/iters/README.md) | — | stdlib `itertools` |
 
-`proc`, `fs`, and `iters` ship in the base install.
+`clone`, `proc`, `fs`, and `iters` ship in the base install.
 
 ## Prior art
 

{unstd-1.0.5 → unstd-2.0.0}/README.md

@@ -27,15 +27,15 @@ python3 -m pip install 'unstd[serde,ids,crypto]'
 | Group | Extra | Backend / fallback |
 |---|---|---|
 | [`serde`](src/unstd/serde/README.md) | `unstd[serde]` | orjson / msgspec / pybase64 → stdlib `json`/`base64` |
-| [`ids`](src/unstd/ids/README.md) | `unstd[ids]` | blake3 / uuid-utils → `hash` requires the extra; `uid` falls back to stdlib UUIDv7 |
+| [`ids`](src/unstd/ids/README.md) | `unstd[ids]` | uuid-utils → `uid` falls back to stdlib UUIDv7 |
 | [`crypto`](src/unstd/crypto/README.md) | `unstd[crypto]` | blake3 (required; hashlib ships BLAKE2, a different algorithm) |
-| [`time`](src/unstd/time/README.md) | `unstd[time]` | `timeutil` is pure stdlib; `dateutil` requires `whenever`, and `timeutil`'s protobuf `Timestamp` bridge requires `unstd[proto]` |
+| [`time`](src/unstd/time/README.md) | `unstd[time]` | `timeutil` is pure stdlib; `zoned` requires `whenever`, and `timeutil`'s protobuf `Timestamp` bridge requires `unstd[proto]` |
 | [`model`](src/unstd/model/README.md) | `unstd[model]` | pydantic (required) |
 | [`rex`](src/unstd/rex/README.md) | `unstd[rex]` | irregex → stdlib `re` |
 | [`containers`](src/unstd/containers/README.md) | `unstd[containers]` | sortedcontainers + immutables (required) |
 | [`rand`](src/unstd/rand/README.md) | `unstd[rand]` | numpy PCG64 → stdlib `random`; `crypto` is always `secrets` |
 | [`pack`](src/unstd/pack/README.md) | `unstd[pack]` | numpy → stdlib `struct` |
-| [`clone`](src/unstd/clone/README.md) | `unstd[clone]` | msgspec → `copy.deepcopy` |
+| [`clone`](src/unstd/clone/README.md) | — | stdlib; exact-type walk, `copy.deepcopy` per foreign node |
 | [`text`](src/unstd/text/README.md) | `unstd[text]` | rapidfuzz → stdlib `difflib` |
 | [`audio`](src/unstd/audio/README.md) | `unstd[audio]` | soundfile + numpy → stdlib `wave` |
 | [`toml`](src/unstd/toml/README.md) | `unstd[toml]` | read is `tomllib`; write requires tomlkit |
@@ -43,7 +43,7 @@ python3 -m pip install 'unstd[serde,ids,crypto]'
 | [`fs`](src/unstd/fs/README.md) | — | stdlib `os` |
 | [`iters`](src/unstd/iters/README.md) | — | stdlib `itertools` |
 
-`proc`, `fs`, and `iters` ship in the base install.
+`clone`, `proc`, `fs`, and `iters` ship in the base install.
 
 ## Prior art
 

{unstd-1.0.5 → unstd-2.0.0}/bench/README.md

@@ -92,8 +92,8 @@ The harness earned its keep on the first run:
   real value is the shape beside it: handed an ndarray it already owns,
   `pack_array` is a memcpy and wins **159×**, and `unpack_array` returns a
   zero-copy view for **1667×**. Hold your data in an array, not a list.
-- **`hash.sum_`'s honest twin is `blake2b`, not `sha256`.** BLAKE2 is what
-  `hashlib` ships from the BLAKE family, so it is the thing `sum_` stands in for,
+- **`digest.raw`'s honest twin is `blake2b`, not `sha256`.** BLAKE2 is what
+  `hashlib` ships from the BLAKE family, so it is the thing `digest.raw` stands in for,
   and BLAKE3 beats it 1.7×. `sha256` is not a fair comparison in either
   direction: every ARMv8 and modern x86 core implements it in silicon, so
   `hashlib.sha256` is a hardware instruction and beats single-threaded BLAKE3
@@ -101,6 +101,15 @@ The harness earned its keep on the first run:
   the algorithm, and it is not something a caller can act on.
 - **`clone.deep` on a small dict is 1.19×, not the 1.8× the README claimed.**
   The floor case is real and the README now says what the harness measures.
+- **The 1.0.5 `clone.deep` floors were set against a clone that dropped types.**
+  Its msgspec round-trip turned a nested `StrEnum` into a `str` and a
+  `defaultdict` into a `dict`, and its equality guard could not tell. 1.1 replaced
+  it with a type-exact walker (4.4× on the large trees instead of 5.8×, 4.1× on
+  the small dict instead of 1.2×), and the two large-tree floors were re-pinned
+  by hand at the usual 70% of the new measurement. That is the one legitimate way
+  a floor comes down: the old number was bought with a bug.
+- **`update --group X` used to erase every other group's floors.** It rewrote the
+  baseline from the narrowed run alone. It now merges, and a test pins that.
 
 ## Adding a case
 

{unstd-1.0.5 → unstd-2.0.0}/bench/baseline.json

@@ -1,14 +1,14 @@
 {
   "_": "Speed floors, ratio of stdlib cost to unstd cost. Raised by `python3 -m bench update`, never lowered by it. See bench/README.md.",
-  "recorded": "2026-09-01",
+  "recorded": "2026-09-27",
   "host": "Darwin arm64 python3.13",
   "cases": {
-    "b64.b64d · 128 KiB": {
+    "b64.decode · 128 KiB": {
       "floor": 6.42,
       "measured": 8.94,
       "group": "serde"
     },
-    "b64.b64s · 128 KiB": {
+    "b64.encode · 128 KiB": {
       "floor": 11.77,
       "measured": 16.52,
       "group": "serde"
@@ -28,19 +28,29 @@
       "measured": 1767.1,
       "group": "pack"
     },
+    "clone.asdict · dataclass tree": {
+      "floor": 7.55,
+      "measured": 10.78,
+      "group": "clone"
+    },
+    "clone.deep · 40 pydantic models": {
+      "floor": 1.58,
+      "measured": 2.25,
+      "group": "clone"
+    },
     "clone.deep · 500 records": {
-      "floor": 4.09,
-      "measured": 5.85,
+      "floor": 3.16,
+      "measured": 4.51,
       "group": "clone"
     },
     "clone.deep · nested (200 items)": {
-      "floor": 4.13,
-      "measured": 5.67,
+      "floor": 3.02,
+      "measured": 4.22,
       "group": "clone"
     },
     "clone.deep · small dict": {
-      "floor": 0.85,
-      "measured": 1.21,
+      "floor": 2.93,
+      "measured": 4.16,
       "group": "clone"
     },
     "draw.floats · 100k": {
@@ -68,7 +78,7 @@
       "measured": 41.87,
       "group": "text"
     },
-    "hash.sum_ · 1 MiB": {
+    "digest.raw · 1 MiB": {
       "floor": 1.21,
       "measured": 1.72,
       "group": "ids"