fix(plugins): evict failed sibling exec from sys.modules so later imports raise ModuleNotFoundError

[Mzzj114]

What

In plugins/plugin_loader.py::load_plugin_module, add the same sys.modules cleanup on failure for the sibling loop that already exists for the main module path (sys.modules.pop(full_sub_name, None) when _exec(sub_mod, ...) returns False).

Why

_new_module() inserts the module into sys.modules before execution; _exec() intentionally leaves the entry on failure ("callers needing a clean retry pop it themselves"), but the sibling loop is a caller that never retries — so a half-initialized module stayed cached. A later relative import of one of its names then raised ImportError: cannot import name ... (error.name == None, uncatchable as ModuleNotFoundError) exactly where the plugin's own handler expects ModuleNotFoundError. Repro and code evidence in #112096.

Fixes #112096

How to test

scripts/run_tests.sh tests/plugins/test_plugin_loader_sibling_cleanup.py -q — 4 tests:

• red on main, green here: failed sibling evicted from sys.modules
• red on main, green here: later importlib.import_module raises ModuleNotFoundError, not ImportError
• healthy sibling still loaded and bound (no behavior change)
• main-module failure path unchanged (pop + return None)

Platforms tested

Linux (Ubuntu 22.04, Python 3.11). Change is platform-neutral module bookkeeping.

Notes

• 5 unrelated tests/plugins/ files fail locally both with and without this change (pre-existing on this machine; likely side-dependent tests/optional deps) — not touched here.
• The issue's edge case (a sibling importing a not-yet-exec'd sibling whose entry then gets evicted) is strictly better than the status quo: the entry is re-created on re-exec, versus a permanently poisoned cache today.

[Mzzj114]

Duplicate of #112097, which was opened earlier for the same fix in plugins/plugin_loader.py — I should have searched open PRs per CONTRIBUTING.md before pushing; withdrawing this so the earlier PR can land.