fix(metadata): aggregate from_email validation errors - #1268

Pushed a follow-up that fixes the red CI issues from the first commit.

What changed:

• narrowed the ExceptionGroup.exceptions handling so mypy can prove InvalidMetadata.field access is safe
• preserved the lone InvalidMetadata compatibility path when there are no email-parser errors to aggregate
• added regression coverage for valid raw metadata with unparsed headers, non-InvalidMetadata grouped errors, single validation errors, validate=False wrapping, and description content-type validation without header defects
• removed the unreachable post-filter branch that left branch coverage below 100%

Checks run locally:

• PYTHONPATH=src python3 -m pytest tests/test_metadata.py -q
• uvx nox -s lint -- --show-diff-on-failure
• NOXFORCEPYTHON=3.11 uvx nox -s tests

Also ran the required adversarial review on the final diff; final verdict was approve.

🤖 AI text below 🤖

Thanks for the fix! I pushed 8c2f9bc, which simplifies the aggregation logic while keeping behavior identical:

• from_raw(validate=True) routes every error through _ErrorCollector.finalize(), so it can only raise ExceptionGroup, never a bare InvalidMetadata. The except InvalidMetadata branch was unreachable, so it is removed along with the two tests that monkeypatched from_raw to trigger it.
• The required-field dedup condition reduces to exc.field in unparsed and _EMAIL_TO_RAW_MAPPING.get(exc.field) not in raw: from_raw only reports fields that are in raw or required, so an error for a field absent from raw is necessarily a missing-required-field complaint. That makes the precomputed required_email_fields / unparsed_raw_fields sets unnecessary.
• The repeated message f-string in _process_description_content_type is hoisted into a variable.

Net -71 lines; all the new integration-style tests pass unchanged.

Claude Opus 4.8 took a look at the new Description-Content-Type handling; the first point below is a regression on input that parsed before this PR.

🤖 AI text below 🤖

Both new rejection paths in _process_description_content_type can raise a bare exception that isn't an InvalidMetadata, so it escapes from_email instead of joining the aggregated group.

Review comment:

• [P2] Braces in the value crash format_map (src/packaging/metadata.py:640-642)
  The .defects branch passes invalid_msg to _invalid_metadata, which runs msg.format_map({"field": ...}) (line 591). invalid_msg embeds {value!r}, so a { in the header value is read as a format field. Description-Content-Type: text/plain; {b} was accepted before this PR and now raises a bare KeyError('b'); text/plain; a}b raises ValueError. Fixing it in _invalid_metadata, e.g. msg.replace("{field}", repr(self.raw_name)), covers every message that interpolates a value.

• [P3] Assigning the content-type can raise IndexError, not only ValueError (src/packaging/metadata.py:636-638)
  An RFC 2231 parameter name ending in * triggers it: Description-Content-Type: text/plain; x* raises IndexError: string index out of range, which the new except ValueError misses, so it escapes from_email bare. except (ValueError, IndexError) catches it.

🤖 AI text below 🤖

Thanks! Both fixed in 8ab2fda:

• Braces in the value: _invalid_metadata now interpolates the field name with str.replace instead of format_map, as suggested, so braces in any interpolated value can no longer crash message formatting — this covers every field's error messages, not just Description-Content-Type. Note that text/plain; {b} is still rejected (it has header defects, which this PR now checks), but as a proper InvalidMetadata inside the aggregated group rather than a bare KeyError.
• IndexError from RFC 2231 parameters: the header assignment now catches (ValueError, IndexError), so text/plain; x* joins the group as well.

Regression tests for all three inputs were added on both the from_email aggregation path and the lazy validate=False path, and confirmed to fail before the fix.

GPT 5.6 Sol:

🤖 AI text below 🤖

The validation improvements generally work, but the new placeholder substitution corrupts diagnostics for metadata values containing the literal {field} token.

Review comment:

• [P2] Avoid replacing placeholders inside user values — /Users/henryfs/git/pypa/packaging/src/packaging/metadata.py:592-592
  When an interpolated metadata value itself contains the exact text {field}, this global replacement also modifies that user value. For example, an invalid version named {field} is reported as 'version' is invalid for 'version' instead of preserving the submitted value, making validation diagnostics inaccurate. Substitute the field placeholder before incorporating user-controlled values, or otherwise replace only the intended placeholder.

I can actually pull out that change from this PR, would be easier to review.

Dropping a Fable adversarial review here, but I'm going to see about the commit above being a separate PR first.

I pulled out the extra rejections to #1329, keeping this focused on just aggregations. That one is still needed to fix some of the leaking, though, so I think that should be considered first.

Pushed a small follow-up for the current Python 3.9 red jobs.

The failures were coverage-only: metadata.py still had an unreachable validate=False ExceptionGroup wrapping path after #1329 split the content-type tightening out of this PR. I removed that dead wrapper and the monkeypatch-only test that forced it.

Checks run locally:

• PYTHONPATH=src python3 -m pytest tests/test_metadata.py -q (327 passed)
• uvx nox -s lint -- --show-diff-on-failure
• NOXFORCEPYTHON=3.11 uvx nox -s tests (62123 passed, 1 skipped, 100% coverage)
• git diff --check

Pushed one more coverage-only follow-up for the Python 3.9 CI failures.

The failed 3.9 jobs were still reporting partial branch coverage in Metadata.from_email() after the previous cleanup. I reproduced that locally with the 3.9 nox session and marked the two structurally partial branches (ExceptionGroup.exceptions iteration and the collector-finalize fallthrough) with # pragma: no branch; there is no runtime behavior change in this commit.

Checks now passing locally:

• PYTHONPATH=src python3 -m pytest tests/test_metadata.py -q (327 passed)
• NOXFORCEPYTHON=3.9 uvx nox -s tests (62123 passed, 1 skipped, 100% coverage)
• uvx nox -s lint -- --show-diff-on-failure
• git diff --check + python3.9 -m compileall -q src/packaging/metadata.py