[Yuren0312]

Problem

timeparse() mis-parses the output of datetime.timedelta.__str__ for negative durations, causing a round-trip drift (issue #23).

timedelta.__str__ renders negative durations as -N day[s], HH:MM:SS[.ffffff], where the leading - negates only the day count and the clock part is always non-negative:

>>> str(timedelta(seconds=-86395))
'-1 day, 0:00:05'   # -86400 + 5 = -86395

timeparse instead extracted the leading sign and negated the entire field sum, so -1 day, 0:00:05 became -(86400 + 5) = -86405. Repeating the round-trip drifts every iteration:

t = timedelta(days=-1, seconds=5)
for _ in range(3):
    t = timedelta(seconds=timeparse(str(t)))
# -86405 -> -259195 -> -259205 -> ...   (keeps drifting)

Fix

• Extracted the second-count computation into a small _sum_fields(mdict, sign) helper (behaviour-preserving refactor).
• For the day+clock form (pattern 2: weeks/days followed by an HH:MM:SS clock), a leading minus now negates only the week/day portion while the clock stays positive, matching timedelta semantics. The DAYCLOCK (DD-HH:MM:SS) and unit forms keep their existing whole-expression sign behaviour.

Now -1 day, 0:00:05 → -86395 and the round-trip is stable.

Tests

• Updated test_timeparse_11 / test_timeparse_12, whose old expected values encoded the previous (incorrect) whole-expression negation.
• Added test_negative_day_clock and test_negative_timedelta_roundtrip covering str(timedelta(...)) round-trips.

Verification: python -m unittest discover -s pytimeparse/tests → 62 tests, OK.