GeezerVison

Summary

Group the related select-family conversion and timeout fixes around the VM's existing signed pytime conversion, rather than maintaining another conversion implementation.

• Share seconds/milliseconds scaling, index-before-float conversion, signed nanosecond bounds, and timeout rounding. Use checked platform timeval/timespec conversion and monotonic deadlines after descriptor/changelist callbacks.
• Reuse Fildes and derived raw argument binding. Preserve argument roots and callback/error ordering for select, poll, epoll, and kqueue.
• Accept all unsigned-short poll event bits, preserve returned high event bits, and report ENOENT for modifying an unregistered descriptor.
• Use ppoll on Linux/Android for nanosecond precision and waits beyond the signed-millisecond limit; retain checked poll fallback elsewhere and recompute remaining waits after EINTR.
• Align epoll registration keywords, maxevents checks, closed-state ordering, signed timeout limits, and sub-millisecond negative rounding.

No test or snippet edits. The overlaps with #9652's separate thread-timeout additions and #8954's C-API time helpers were checked by clean three-way compositions of pytime.rs.

Reference

Compared against CPython v3.15.0 Modules/selectmodule.c, its Clinic bindings, and Python/pytime.c.

Validation

• Focused VM, host_env, and stdlib threading cargo check --offline: passed.
• Matching strict cargo clippy --offline -- -D warnings: passed.
• Independent source review, formatting, git diff --check, and normal repository pre-commit hooks: passed.
• Bounded CPython 3.15 probes covered conversion errors, zero waits, a ready owned pipe with a large timeout, epoll keywords, and POLLRDHUP from an owned socket pair.
• Native RustPython execution, full workspace/C-API tests, and cross-platform builds were not run under the current task restrictions. Hosted CI remains required.

Platform-specific kqueue, Windows, and 32-bit paths are source-reviewed, not runtime-validated. The pre-existing epoll signal-handler-close locking behavior remains outside this conversion patch. A 32-bit ppoll time_t overflow retains a generic timeout-overflow diagnostic.

AI assistance

OpenAI Codex assisted with source comparison, implementation, review, and validation. The model version is not exposed by this environment.