PYTHON_LAZY_IMPORTS=all halves CLI startup on 3.15; 5 of 38 tools break
Measured on CPython 3.15.2 (PBS, Linux x86-64, warm cache, 30 interleaved rounds, writable cache dir): the PEP 810 global switch made 41 unchanged commands from 39 popular Python CLIs start in 0.54x the eager time (geometric mean; range 0.19-0.93). poetry --version 270->55 ms, datasette 324->61 ms, django-admin --help 97->30 ms; python -c pass unchanged. mypy only gets 0.92x because mypyc-compiled modules' imports stay eager. Five failure classes: (1) an import-time side effect never runs (black --no-cache / black.format_str: blib2to3 grammar not initialised); (2) a meta-path finder's own dependency is lazy, giving ImportCycleError (pytest assertion rewriter + fnmatch); (3) a submodule import overwrites a pending lazy name with the same name (hatch's from hatch.cli.env import env becomes the module; same as CPython gh-158086); (4) dict(mod.__dict__) copies unresolved proxies (coverage --version prints <lazy_import ...>); (5) plugin probing that catches import errors is defeated (stevedore loads bandit's sarif formatter without its dependency). Fix without patching the tool: a .pth that calls sys.set_lazy_imports_filter(lambda importer, name, fromlist: name != 'black.nodes') keeps one import eager. That restored identical output for 4 of 5 tools at 0.46-0.79x.
Follow-up: a broader sweep (console scripts of the top-1000 PyPI projects, --help only, binary wheels for 3.15) found real behaviour changes in 12 of the 153 projects it could test (7.8%, Wilson 95% CI 4.5-13.2%). New cases: protobuf-generated _pb2 modules (streamlit, tensorboard, modal), CLIs that register subcommands by importing modules and silently lose them from --help (spacy, cyclopts), and import cycles exposed by the changed order (torch, antlr4). Also, black only breaks with --no-cache or library use (black.format_str); with a writable cache, black --check works.
Update: rerunning the 57 projects that had no 3.15 binary wheels with source builds allowed (about 20 failed only because pyyaml has no cp315 wheel) added 8 testable projects. Only bandit changes, as already known. Overall: 13 of 161 (8.1%, Wilson 95% CI 4.8-13.3%). The antlr4 cycle case turned out to be CPython-side: a re-entrant lazy from-import raises ImportCycleError even when the name already exists, and it fails with the explicit 'lazy from' keyword too.