argparse performance regression in 3.14+ due to colorization overhead

[savannahostrowski]

I spent some time running the argparse bm_subparsers benchmark from pyperformance (1000 optional arguments, 10 iterations)

Version	Time	Function Calls
3.10	0.585s	560K
3.11	0.207s	560K
3.12	0.229s	560K
3.13	0.246s	560K
3.14	0.934s	2,380K
3.15 (main)	0.883s	2,367K

For a "realistic" CLI with 10 arguments, parser creation is ~3x slower on main compared to 3.13.

The root cause is that _get_formatter() is called twice per add_argument() - once for metavar validation and once in _check_help() for help string validation (called at the end of add_argument ). Each _get_formatter() call creates a new HelpFormatter, which calls _set_color(), which calls can_colorize(), which checks 5 environment variables.

I think a viable fix is to cache the HelpFormatter on ArgumentParser for validation operations. The validation only performs read-only operations (_format_args, _expand_help) that don't modify formatter state. This preserves the existing _get_formatter() behavior while eliminating redundant _set_color() calls during argument setup.

Linked PRs

• GH-142267: Cache formatter to avoid repeated _set_color calls #142268
• [3.14] GH-142267: Cache formatter to avoid repeated _set_color calls (GH-142268) #142313

[savannahostrowski]

An alternative approach could be to lazily _set_color but that could also break the theme (and then I can't backport a fix)? I'll open the PR for discussion.