[jeff-grimes-pplx]

Fixes #1168.

Problem

babel.dates.parse_time() decides between AM and PM by searching the input for the English substrings "am" / "pm". Babel's own format_time() output for locales whose CLDR day-period names aren't English (Korean 오후, Spanish p.m., Greek μ.μ., Arabic م, Chinese 下午, Albanian m.d., Afar carra, ...) is therefore always parsed as AM, so every afternoon time comes back 12 hours early. No error is raised.

>>> from babel.dates import format_time, parse_time
>>> import datetime
>>> s = format_time(datetime.time(15, 30), 'short', locale='ko_KR')
>>> s
'오후 3:30'
>>> parse_time(s, locale='ko_KR', format='short')
datetime.time(3, 30)        # expected datetime.time(15, 30)

When I round-tripped format_time() through parse_time() for every CLDR locale whose short/medium time pattern uses the a period field, every hour 0–23, 4,038 of 17,040 round-trips were wrong, across 171 locales.

Root cause

parse_time() (babel/dates.py, the if 'a' in format_str: block) hardcodes English markers:

if 'pm' in string.lower(): ...
elif 'am' in string.lower() and hour == 12: ...

It never reads the locale's day_periods data, even though the same data drives format_time()'s a field.

Fix

A small helper, _parse_period(), works out the marker from the locale's own AM/PM day-period names:

• it checks every context (format / stand-alone) and every width (abbreviated / wide / narrow);
• it tries the longest name first, so a short name like p can't match inside a longer one (dop. in Slovenian, for example);
• it compares case- and whitespace-insensitively, because CLDR names contain U+00A0 / U+202F (Catalan p. m., for example);
• it falls back to the English am/pm, so existing callers that parse English markers with another locale keep working.

The 12 AM / 12 PM normalisation is unchanged.

Verification

• New regression test tests/test_dates.py::test_parse_time_localized_am_pm: it fails on master (assert datetime.time(3, 30) == datetime.time(15, 30)) and passes with the fix.
• Full suite: 7939 passed, 8 skipped, 2 xfailed.
• ruff check is clean.
• An exhaustive sweep over all locales, five 12-hour patterns (h:mm:ss a, h:mm:ss aaaa, h:mm:ss aaaaa, a h:mm, K:mm a) and all 24 hours: correct round-trips go from 99,075 to 134,476 out of 134,520, with 0 regressions. The 44 that still fail are all Anii (blo), whose period names contain digits (1ka / 2ja) and confuse the existing digit extraction. That's a separate, pre-existing limitation.
• Out of scope: flexible day periods (pattern B, used by zh_Hant, for example).

[akx]

I'm sorry, how does #1168 relate to this?