[Sskift]

Summary

round() without ndigits (or with None) should return an integer for a float, but Float.__round__ defaults to zero digits and returns another Float. A value read from TOML therefore fails where the equivalent rounded native float works:

import tomlkit

count = tomlkit.parse("count = 2.5")["count"]
range(round(count))  # TypeError before this change

Distinguish omitted/None digits from an explicit digit count, wrapping the native integral result as Integer with copied trivia. Explicit digit counts still produce Float, including zero. This also restores the native errors for rounding infinities and NaN without digits.

Validation on Windows / CPython 3.12.10:

• The new tests produce 16 failures and 4 passing controls on the original source.
• All 132 item tests and all 1,244 tests, including the pinned TOML conformance corpus, pass with this change.
• Ruff 0.17.2 lint/format and git diff --check pass.
• Mypy 1.19.1 reports the same 33 existing errors in eight files on both versions; diagnostic messages are identical after normalizing shifted line numbers.
• A separate Codex agent reviewed the exact patch and independently reproduced the red/green item tests. Its 2,812-case native-Python comparison found no mismatches after the fix, and strict result-type fixtures passed in Python 3.9 and 3.12 type-checking modes. Runtime testing used Python 3.12.

Tests cover ties-to-even, both signs, omitted/None/direct invocation, explicit digit counts, nonfinite errors, and comment/sign preservation.

Duplicate Check

• I searched the open PRs and found none that already fixes this issue.
• Refreshed the open-PR list and targeted __round__ searches before submission. Related open PRs Preserve floating-point results of negative integer powers #640 and Forward modulus arguments in Float power methods #691 concern power methods, not rounding; this change does not modify them.

Agent Drafting Metadata

• Agent: OpenAI Codex
• Model: Exact model identifier was not exposed to this worker.
• Notes: Codex assisted investigation, implementation, tests, and this description. The separate review described above was performed by another AI agent; no human review or maintainer acceptance is claimed.