fix: allow thrift 0.25.x and clear additional Apache Thrift CVEs

[hannonpi1228]

Closes #969

Summary

• Widen the thrift constraint from ~=0.24.0 to >=0.24.0,<0.26.0 so 0.25.0 can be resolved.
• Bump the locked thrift to 0.25.0.
• Pass explicit string_length_limit=None, container_length_limit=None to TBinaryProtocol to preserve pre-0.25 unbounded read behavior for the connector's own already-bounded, trusted-server result stream.

Why this isn't a bare version bump

Apache Thrift 0.25.0 (released 2026-09-30) fixes 61 CVEs across language bindings (combined announcement); several affect the Python binding, including CVE-2026-66858 (skip() recursion-limit bypass in the Python accelerator) and CVE-2026-85494 (framed transport / binary protocol size a read buffer from a peer-declared length with no effective maximum). Snyk currently reports both as firing against this connector via the thrift==0.24.0 pin.

Fixing CVE-2026-85494 also changed TBinaryProtocol's default string_length_limit from unbounded (None) to DEFAULT_STRING_LENGTH_LIMIT (DEFAULT_MAX_FRAME_SIZE = 16,384,000 bytes ≈ 15.6 MiB). thrift_backend.py previously constructed the protocol with no kwargs:

protocol = thrift.protocol.TBinaryProtocol.TBinaryProtocol(self._transport)

so a bare version bump would make every readBinary/readString call -- including the inline Arrow result batches in TFetchResultsResp -- inherit the new ~15.6 MiB cap. This connector's own result-size knob, buffer_size_bytes (DEFAULT_RESULT_BUFFER_SIZE_BYTES = 100 MiB in client.py), is well above that, so this would be a real regression on any result batch between ~15.6 MiB and the connector's own 100 MiB default -- not a theoretical one. Confirmed empirically: a 20 MiB binary field reads fine under thrift 0.24.0, and under thrift 0.25.0's default construction raises TTransportException: Length exceeded max allowed: 16384000; passing string_length_limit=None (this PR) restores the read.

This connector already bounds and governs its own result stream (TLS-verified connection to a single trusted Databricks backend, server-negotiated maxBytes/buffer_size_bytes), so re-enabling the unbounded local read limit does not reintroduce the untrusted-peer DoS/amplification scenario the thrift default is meant to guard against for general-purpose Thrift servers.

DBR LTS packaging risk (the reason this pin existed at all, see #798/#840)

0.25.0 ships the same wheel matrix as 0.24.0 -- manylinux2014/macOS/musl/Windows for cp310-cp314, confirmed on PyPI -- so pip resolves a wheel and never runs setup.py on DBR LTS; the SEV0 build-time failure class that originally forced this pin does not apply here. The DBR LTS Install CI check remains the authoritative gate for this.

Other CVEs in the 0.25.0 batch with Python impact

Per #969, not all are mapped in GitHub Advisory Database / OSV yet for the PyPI thrift package, but Apache's announcement lists several more affecting Python bindings (e.g. TJSONProtocol size-limit issues, TZlibTransport decompressed-size enforcement). This connector doesn't use TJSONProtocol or TZlibTransport (thrift_backend.py only imports THttpClient, TBinaryProtocol, TSocket, TTransport), so those are moot for this codebase, but the version bump clears them for any downstream code path.

Validation

• Reproduced the regression directly against both versions with a standalone TMemoryBuffer round-trip of a 20 MiB binary field: fails under thrift 0.25.0's default TBinaryProtocol() construction, passes with the explicit string_length_limit=None this PR adds (same behavior as thrift 0.24.0).
• Added test_binary_protocol_preserves_unbounded_string_and_container_limits (asserts the connector constructs TBinaryProtocol with string_length_limit=None, container_length_limit=None) and test_binary_protocol_reads_result_larger_than_thrift_default_limit (round-trips a 20 MiB field through the real thrift package with the connector's exact kwargs) to tests/unit/test_thrift_backend.py.
• poetry run python -m pytest tests/unit with thrift 0.25.0 actually installed: 1020 passed, 5 skipped.
• black --check / mypy clean on changed files.

Signed-off-by: Paddy Hannon pih@ehukai.com