`PyMutex` is built on `RawCellMutex` on single-threaded targets (iOS,
Android, wasm32) and is not `Sync`, so a `static PyMutex<...>` fails to
compile. `PyTypeRef` is also not `Send` by default, which compounds the
issue.

Replace the static parameter-type registry with class-name comparison
(`"CompressionParameter"` / `"DecompressionParameter"`). The two names
originate from the `compression.zstd` module we ship, so identity vs.
name comparison is equivalent in practice. `set_parameter_types` now
just validates that its arguments are type objects.

Verified locally:
- `cargo build -p rustpython-stdlib` (default features): clean.
- `cargo check -p rustpython-stdlib --no-default-features --features host_env`
  (simulates iOS by dropping the `threading` feature): clean.
- `cargo run -- -m test test_zstd`: 119 tests pass, 1 expected failure.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

RustPython's clippy lint set rejects `std::` imports for items that also
live in `core::` or `alloc::`. Switch every such reference in `zstd.rs`:

- `std::ffi::{c_int, CStr}`        -> `core::ffi::{c_int, CStr}`
- `std::fmt::*`                    -> `core::fmt::*`
- `std::slice::from_raw_parts`     -> `core::slice::from_raw_parts`
- `std::borrow::Cow`               -> `alloc::borrow::Cow`

Also clear up other clippy findings under `-D warnings`:

- Drop the redundant `obj.clone()` on the last use of `obj` in
  `parse_zstd_dict_arg`.
- Collapse `.map(|n| n.get()).unwrap_or(0)` to `.map_or(0, |n| n.get())`
  in the two dict-id readers.
- Replace `&*work_data` with `&work_data` (auto-deref).
- Factor the `(Option<Digested>, Option<PyRef<ZstdDict>>)` return shape
  into a `DictLoadResult<D>` type alias to satisfy `type_complexity`.
- Replace the remaining `std::mem::transmute<u32, ZSTD_cParameter>` /
  `ZSTD_dParameter` in `get_param_bounds` with the existing safe
  `c_param_enum` / `d_param_enum` helpers; surfaces a clear "invalid
  parameter" `ValueError` instead of relying on UB-adjacent transmutes
  for unknown ints.

Verified locally:
- `cargo clippy -p rustpython-stdlib --no-deps -- -Dwarnings`: clean
  (only the pre-existing `socket.rs::sock_wait` warning remains).
- `cargo run -- -m test test_zstd`: 119 pass, 1 expected failure.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The previous run's `Lint` job failed with a `401 Unauthorized` from
reviewdog hitting the GitHub API (a transient CI/permissions issue,
unrelated to this PR's diff). Pushing an empty commit to re-run CI.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

{"subject": "Format zstd module and extend spell-check ignores", "body": "- Apply rustfmt formatting to zstd.rs\n- Add additional spell-checker ignore entries for zstd identifiers"}

Co-authored-by: Shahar Naveh <50263213+ShaharNaveh@users.noreply.github.com>

{"subject": "Remove rustdoc comments from zstd module", "body": "- Strip doc comments from internal items in _zstd module\n- Convert a few API-doc comments into regular implementation notes where appropriate"}

{"subject": "Mark zstd load_dict as unsafe and document invariant", "body": "- Add # Safety docs requiring PyRef<ZstdDict> to outlive the context\n- Wrap calls in load_compressor_dict and load_decompressor_dict with SAFETY comments"}

{
  "subject": "refactor(zstd): release GIL during (de)compress and tighten dict-load safety",
  "body": "- Wrap compress/decompress loops in vm.allow_threads to release the GIL\n- Replace load_*_dict helpers with build_*_state that assemble the full state, making the load_dict safety invariants structural\n- Switch constructor args from OptionalArg to OptionalOption and use try_to_value instead of the pyobj_to_i32/arg_or_none helpers\n- Add check_sample_sizes_match with overflow-safe summing for train_dict/finalize_dict"
}

- Have set_parameter_types stash the CompressionParameter/DecompressionParameter classes as private _zstd module attributes instead of validating and discarding them
- check_wrong_param_kind now compares key classes by identity against the registered type, matching CPython's Py_TYPE check, and skips when unregistered
- Error message names the actual key type as an attribute

- Raise TypeError (not RuntimeError), matching CPython, when both
  `level` and `options` are passed to ZstdCompressor
- Parse set_pledged_input_size's argument before taking the state lock
  so a re-entrant __index__ can't deadlock, and fix the off-by-one in
  its bound message to match CPython

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Remove the remaining `// ===...` section separators from zstd.rs per
the project comment guidelines, drop the unused `zstd` crate
dependency (only zstd-safe/zstd-sys are used) and declare zstd-safe's
std feature explicitly instead of relying on the zstd crate's feature
unification, and convert a doc comment on #[extend_class] to a regular
comment.

Replace the C-library bindings (zstd-safe/zstd-sys) with Trifecta Tech
Foundation's libzstd-rs-sys (pinned git rev), a c2rust translation of
upstream libzstd. The module now calls its C-API-compatible functions
through small owning RAII wrappers (CCtx/DCtx/CDict/DDict) with the
same drop-order invariants as before.

The new backend builds zstdmt_compress in, so nb_workers bounds are
(0, 256) and test_zstd_multithread_compress passes for real; drop the
now-obsolete expectedFailureIf marker.

The zero-cap probe iteration handed libzstd a 1-byte buffer and the
post-loop truncation discarded whatever it emitted, while the input
counter had already advanced past the bytes that produced it — every
decompress(..., max_length=0) call with pending output silently dropped
one byte. The suite never caught it because max_length=0 was only
exercised against zero-output (skippable) frames.

Match CPython's mechanism instead: probe with a zero-size output
buffer, so libzstd consumes input without emitting and never reports
frame completion with un-emitted content. Add a regression snippet
covering lossless zero-cap probes (content, truncated, and skippable
frames), verified byte-exact against CPython.

- Derive the compression/decompression parameter ids from libzstd's own
  `ZSTD_cParameter`/`ZSTD_dParameter` constants instead of restating them
  as literals. The crate declares both as `#[repr(transparent)]` newtypes
  over a private `u32` with no accessor, `From` impl or `Deref`, so the id
  is read back through that guaranteed layout.
- Drop the redundant `name = ` on `#[pyattr(once)] fn zstd_version` and
  `fn zstd_version_number`; the function names already match.
- Use `CStr::to_str` for the version string. libzstd returns a fixed ASCII
  `MAJOR.MINOR.RELEASE` literal, so there is nothing for a lossy
  conversion to repair.
- Replace the `DICT_TYPE_*` int constants with a `DictType` enum carrying
  the pinned discriminants, decoded at the Python boundary by
  `DictType::from_marker`.
- Explain the non-trivial assertions: why `level_bounds`' `expect` cannot
  fire, and why a NULL context allocation panics rather than raising
  `MemoryError`.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`COMP_MODE_CONTINUE`/`FLUSH_BLOCK`/`FLUSH_FRAME` become a `CompressMode`
enum with pinned discriminants, decoded at the Python boundary by
`CompressMode::from_int` and mapped to libzstd by
`CompressMode::end_directive`. `CompressorState::last_mode` now holds the
enum, so the states `set_pledged_input_size` accepts are checked by the
type system rather than by comparing ints.

`flush()` keeps rejecting `CONTINUE` and out-of-range modes with its own
message rather than the one `compress()` uses, so the Python-visible
errors are unchanged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Verified each claim the previous two commits added, against the pinned
libzstd-rs-sys rev, CPython's `Modules/_zstd/` and `Lib/test/test_zstd.py`,
and corrected the ones that did not hold:

- The `DictType` discriminants come from CPython's `dictionary_type` enum,
  not a type named `DictType`, and `test_zstd` never hand-builds a valid
  `(zdict, marker)` tuple — it pins the accepted range from the outside
  with `(zd, -1)` and `(zd, 3)` rejection cases.
- `#[repr(transparent)]` guarantees layout, not validity invariants. The
  transmute is sound because every bit pattern is a valid `u32`, which
  holds only in the enum-to-int direction; say so, and note that this is
  why `c_param_enum` decodes untrusted ids with an explicit match.
- The id is unreachable through the crate's public API only in the sense
  that no accessor exists; the derived `Debug` does print it.
- Undigested dictionary loading is not uniformly lazy:
  `ZSTD_DCtx_loadDictionary` digests eagerly and rejects corrupted content
  at construction time, while `ZSTD_CCtx_loadDictionary` accepts it.
- `CompressMode`'s discriminants are libzstd's `ZSTD_e_continue`/
  `ZSTD_e_flush`/`ZSTD_e_end` values.

`CCtx::create`/`DCtx::create` now return `PyResult` and raise `MemoryError`
like CPython's `_zstd` does, rather than panicking. The old justification
for the panic claimed a `vm` reference would have to be threaded through
every construction site; there are two, and both already had one.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Found by reviewing the module against CPython's `Modules/_zstd/` and the
pinned libzstd-rs-sys sources; each item below was reproduced before the
fix and re-checked after.

- `train_dict`/`finalize_dict` sized their output buffer with an
  infallible `Vec::with_capacity`/`vec![]` from a caller-supplied
  `dict_size`, so `train_dict([b'x'], 2**62)` aborted the interpreter
  instead of raising. Reserve fallibly and raise `MemoryError`, as CPython
  does.
- A digested dictionary was always built at `ZSTD_CLEVEL_DEFAULT`. Since
  `ZSTD_CCtx_refCDict` takes its parameters from the CDict, that silently
  discarded the requested level: `level=19` with `as_digested_dict`
  compressed a test corpus to 39817 bytes instead of 31951. Build the
  CDict at the compressor's effective level, tracked through `level=` and
  the `options` dict like CPython's `self->compression_level`.
- A bare `ZstdDict` was loaded as digested in both directions; CPython
  compresses with an undigested dictionary by default and only
  decompresses with a digested one. Give `DictLoader` a per-direction
  default.
- Neither compressor nor decompressor reset its session after an error,
  leaving the object permanently broken — every later call failed with
  "Operation not authorized at current processing stage" or re-reported
  the original corruption on valid input. Reset as CPython's error paths
  do, including restoring `last_mode` to `FLUSH_FRAME`.
- `apply_options` range-checked every parameter, rejecting values libzstd
  and CPython accept: 0 ("use the default") for `window_log`, `strategy`,
  `window_log_max` and friends, any non-zero value for the boolean flags,
  and the clamped `nb_workers`/`job_size`/`overlap_log`. Only
  `compression_level` needs the check — libzstd clamps that one silently,
  which is why `test_compress_parameters` requires a `ValueError` — so
  everything else now goes to libzstd, whose error code already maps to
  the same message.
- `ZstdDict` accepted content shorter than 8 bytes; CPython rejects it up
  front regardless of `is_raw`.
- `compress`/`decompress` refused `data` as a keyword and
  `get_param_bounds` required `is_compress` as one, both contrary to
  CPython's clinic signatures.
- The decompressor allocated a full 128 KiB scratch buffer even when
  `max_length` capped output far below that.

`test_zstd` (119 tests) and the snippet test pass; clippy is clean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The branch dropped the `skipIf(not SUPPORT_MULTITHREADING)` guard from
`test_zstd_multithread_compress`. That guard can never fire against this
backend: libzstd-rs-sys compiles multi-threaded compression in
unconditionally — its Cargo.toml has no `zstdmt`-style feature — so
`CompressionParameter.nb_workers.bounds()` is never `(0, 0)` and
`SUPPORT_MULTITHREADING` is always true.

Removing the decorator therefore changed nothing except to put a
modification of a vendored CPython test in the diff, and it would turn
into a hard failure rather than a skip if a future backend ever lacked
multi-threading. Restore the file so the PR touches no CPython test.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>