[rr-]

In a Numpydoc Raises or Warns section, an item written with its description on the head line made the whole line the type:

Raises
------
ValueError : if bad.
    Or worse.

gave a raises with type_name='ValueError : if bad.' and the description Or worse.. It now gives type_name='ValueError' and the description if bad.\nOr worse., as sphinx napoleon reads it. ValueError: if bad. without the space before the colon splits the same way, and ValueError : alone on its line takes its description from the indented lines below.

The line splits only when what comes before the colon is one type: a dotted name, maybe with a subscript such as pkg.Error[int]. That is the same type pattern Google's Raises: already uses, so it moves to _text for both styles to share. Prose with a colon, such as If the value is bad: an error, and several types, such as KeyError, TypeError : if missing., stay whole as before. Compose already writes the type and the description on separate lines, so the split items read back unchanged.

Parity against origin/main: 6 of 29,142 docstrings differ, 4 from the case files and 2 generated, none from the stdlib. The case ones are two pandas corpus docstrings (TypeError : on inequality comparisons and ValueError : `match` does not ...) plus the new numpydoc/raises_03 and numpydoc/warns_02 cases. The generated ones are a y: thing item and a Yields: item that ended up under a Raises heading, which now split into the type y with the description thing, and the type Yields. All of them differ in parse NUMPYDOC and parse AUTO and in their composes.

Bench (just bench): parse numpydoc 172.7 ms on main vs 171.8 ms (0.99x), and the other workloads 0.98x to 1.00x.