Ignore invalid numeric GML polygon position lists

[guocn]

Summary

Addresses the malformed GML polygon position-list case in #599 (not all cases in that umbrella issue).

_parse_georss_line() returns None for invalid numeric coordinates, but the polygon branch of _parse_poslist() unconditionally indexes that result. Consequently an invalid numeric token such as . raises TypeError through public feedparser.parse().

Return None when the ring parser fails, matching the existing geometry-parser failure contract. The existing _end_gml_poslist() already ignores an unsuccessful geometry result. Valid polygon coordinates, coordinate ordering and unrelated feed data are preserved. This does not change polygon validity/closure rules or broaden exception catching.

Added 11 regressions/controls: three helper inputs, the same three complete Atom/GML feeds through both strict and loose parsers, the issue's minimized malformed input, and a valid ring control. Included a Fixed changelog fragment.

Checked #600/#602/#603/#604 and current open PRs: they address other #599 failures; none handles this invalid ring result.

Verification

Linux x86_64, baseline a22c5521cbb109871f1a2318948581901bd47e26:

• Before fix: 10 new tests fail with TypeError, 1 valid control passes.
• Python 3.13.5 with chardet 5.2.0: full suite 4316 passed.
• Python 3.11.16 without optional chardet: full suite 4308 passed, 8 skipped.
• black --check, isort --check-only, flake8 on changed Python files: pass.
• mypy: success, 33 source files.
• git diff --check: pass.

With newest chardet 7.6.0, six existing encoding-expectation tests fail on both baseline (4299 passed, 6 failed) and patched source (4310 passed, 6 failed); new geometry tests pass. Those dependency-specific failures were not altered or hidden. The green full-suite run uses chardet 5.2.0 explicitly.

Tests ran with networking disabled and a clean environment; HTTP tests use the project's mocked responses. Windows/macOS and other supported Python versions were not tested locally. Upstream CI/merge success and performance claims are not made.

AI disclosure: this investigation, patch, regressions and PR text were prepared by an OpenAI coding assistant (gpt-6.1-sol) under guocn's explicit authorization. The account owner has not yet independently reviewed the changes.