[farmer-data]

Fixes #706.

A descriptor that protoc did not produce can have two fields of one message with the same number, and codegen accepted it. With CodecStrategy::Table the consumer's crate failed to compile on a const panic in Table::new that did not identify the message or the field. With the default codec the crate compiled with an unreachable_patterns warning and lost data: with two int32 fields numbered 1, {a: 5, b: 7} encodes as [8, 5, 8, 7] and decodes as {a: 7, b: 0}.

generate now returns the new CodeGenError::DuplicateFieldNumber, which reports the message, both fields and the number, for every message of a generated file, whichever codec is selected. Files that are only imported, and packages under exclude_package, are not checked.

Of the issue's two outcomes this is the hard error: a fallback to the unrolled codec would have kept the data loss.

Not checked here: an extension's number against its extendee's fields (#626 covers the descriptors that codegen accepts and DescriptorPool rejects). dense_lookup's error for more than 255 low-numbered fields stays as a backstop that generate no longer reaches, so it has no test.

Two existing test fixtures had a duplicate number by accident and are corrected.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[iainmcgin]

[claude code] Thanks for offering the switch. The maintainer decision is the hard error: the fallback leaves unrolled code that compiles and loses data. With two int32 fields numbered 1, {a: 5, b: 7} encodes as [8, 5, 8, 7] and decodes as {a: 7, b: 0}, with an unreachable_patterns warning in the generated decoder.

Pushed one maintainer commit on top of yours (a471391):

• generate returns the new CodeGenError::DuplicateFieldNumber { message_name, first_field, second_field, number } for a message of a generated file that has two fields with one number, whichever codec is selected. It is checked before the table plan, so a message that also cannot use the table gets this error and not the plan advice to select the unrolled codec.
• The Ineligible check in table_fields and the guide bullet are removed (a debug_assert! keeps the invariant), and the dense_lookup doc points at the new error.
• Tests: the three fallback tests become four that expect the error: every strategy, including an exact-path rule for a message that is also table-ineligible; a nested message with a oneof member; 256 duplicates; a file that is not generated. Two existing fixtures had a duplicate number by accident (tests/naming.rs, tests/shared_corpus_context.rs) and are corrected.
• The fragment is rewritten and renamed.

cargo test --workspace passes on this head (4273 passed, 0 failed). The title and description will be updated to match.

[farmer-data]

Thanks for taking this over the line. The hard error is the better outcome. I've updated the title and description to match a471391.