GeezerVison

Summary

Make CSV dialect construction collect raw attributes and keyword overrides before converting fields, using Python 3.14's lookup/conversion order.

• Preserve exceptions from strict, doublequote, and skipinitialspace truth conversion for instances, classes, native _csv.Dialect, and dialect registration.
• Read dialect classes directly rather than instantiating them, and skip overridden attributes before lookup or conversion.
• Continue treating failed attribute lookup as an absent attribute and applying defaults, including the default-quoting interaction with quotechar=None.
• Use the same ordered conversion path for keyword options and native dialect construction, retaining original exception objects.

This is a coherent replacement of the existing duplicated dialect-conversion paths. A one-line strict propagation change is insufficient: the old parser also replaces exceptions, instantiates class inputs, and converts overridden attributes before applying keyword options. The shared path removes the duplicated conversion code.

The strict-conversion issue was identified while auditing #8954. That PR's attribute-lookup propagation was intentionally not copied because CPython 3.14 ignores attribute lookup failures here.

Validation

Base: bd8fdb7e5ca3e4e9f78f4f34b122a84c9dd2e2d0.
Head: 62a324e740cf0f19a219e358304bce24bd0212de.
Reference: CPython 3.14.7, Linux x86_64.

• The new regression in the existing extra_tests/snippets/stdlib_csv.py fails the separately built, immutable current-main executable because a strict.__bool__ exception is suppressed.
• The complete CSV snippet passes patched RustPython and CPython 3.14.7 through the actual pytest snippet runner.
• Unmodified canonical -m test test_csv passes on baseline and patched RustPython: 134 tests, 6 skips. Existing expected-failure marking is unchanged.
• Observation tests match CPython's field lookup/conversion order, lookup failures, class handling, keyword override behavior, and exception precedence.
• Independent 432-case constructor review covers _csv.Dialect, reader, and writer, with 8 fields and 18 values per field. Baseline differs on 242 cases; patched semantic outcomes match all 432. The remaining 21 text-only differences are 18 preexisting validation-error wordings and 3 generic Rust integer-overflow messages versus CPython's C-int wording.
• Unknown keyword and excess reader/writer positional arguments fail before truth conversion.

This does not change the preexisting non-ASCII delimiter limitation, native-dialect identity behavior, or unrelated CSV parsing/writing paths. Broad Rust runtime suites were not executed; only focused CSV runtime checks and compile-only aggregates are used.

Additional checks passed:

• cargo clippy --locked --release -p rustpython-stdlib --all-targets -- -D warnings
• cargo test --locked --release --workspace --exclude rustpython_wasm --exclude rustpython-venvlauncher --exclude rustpython-capi --no-run
• From crates/capi: cargo test --locked --release --no-run
• Normal commit hooks passed without changing the tested source or regression file.

AI assistance

OpenAI Codex assisted with the audit, implementation, regression tests, and validation. The runtime checks above were actually executed in the Linux environment; no untested platform behavior is claimed.