Roll back plugin registration when historic replay raises

[xiehuanyi]

Related to #761.

If a historic hook implementation or its result callback raises during register(), the plugin currently remains registered with only some implementations installed. This prevents a clean retry and can leave hooks running from a failed plugin. Remove the failed registration before propagating the original exception. Preserve unrelated registered plugins and specifications added by user code during replay.

Regression coverage includes implementation and result-callback failures, RuntimeError and KeyboardInterrupt, retry after failure, validation against a specification introduced during replay, and preservation of a replacement plugin installed under the same name during failed replay.

Validation:

• Five regression cases fail on the original source.
• Full suite: 224 passed on CPython 3.15.
• All pre-commit hooks pass, including mypy.
• PluginManager line and branch coverage: 100%.

Tests use PYTHONPATH=src uv run --no-default-groups --with 'pytest>=8' --with 'pytest-benchmark>=4' pytest -q: the checked-in lock selects old test tools that cannot run on modern Python. The lock is unchanged. User callback side effects outside the failed registration are retained.

Prepared with OpenAI Codex.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.