Let a parser's Grape errors through the formatter without re-raising

[ericproulx]

Summary

Formatter#read_rack_input answers a parser's StandardErrors with a 400, and let Grape's own errors go on to the error middleware by rescuing them first: rescue Grape::Exceptions::Base => e; raise e. Re-raising an exception makes Ruby read its backtrace (setup_exception calls rb_get_backtrace, which builds the lazily captured backtrace as Strings), and at request depth that is the dearest part of the error: about 25 µs. It was paid by every malformed body, since the built-in parsers report one as InvalidMessageBody.

A matcher module in the rescue clause now selects the errors to answer here, StandardErrors that are not Grape errors, so a Grape error is never rescued at this frame and travels on as it is. Unwinding past a frame that does not rescue it costs nothing extra.

Benchmarks

Median of 7 interleaved subprocess rounds, Ruby 4.0.6, no JIT, POST with a malformed JSON body:

API	compared against	delta
no rescue_from	master	+28.7%
rescue_from :all	master + #2931	+29.8%
rescue_from :all	master	−0.6%

The last row is not a contradiction. A backtrace is built once and then cached, so the cost is paid by whichever step reads it first. On an API with rescue_from :all, master's default handler also reads the backtrace eagerly (fixed by #2931), so either change alone just moves the cost to the other step, and the gain lands once both are in. Without rescue_from, nothing else reads it and this change alone gets it all.

Missing spec, added

A parser raising something that is not a StandardError, such as an Interrupt or a SystemExit, keeps propagating rather than being answered as a 400. Nothing pinned that: dropping the StandardError bound from the matcher passed the whole suite. The new spec raises a NotImplementedError from a custom parser; it passes before and after this change.

Behaviour

Byte-identical to master over a 30-case matrix of rescue_from :all APIs, with and without backtrace: true and original_exception: true, in JSON, txt and XML, including malformed bodies. The rendered backtraces of InvalidMessageBody are identical once paths are normalized, so re-raising never altered them.

Test plan

• Full RSpec suite passes locally.
• RuboCop clean.
• Mutation-checked: the matcher also taking Grape errors fails 4 specs; the matcher dropping its StandardError bound fails 1 (the new spec).
• CI green.

🤖 Generated with Claude Code

[github-actions]

Danger Report

No issues found.

View run

[ericproulx]

Closing in favour of #2936, which combines #2922–#2934 into one PR, re-benchmarked as a whole against master. The write-up here (behaviour matrix, mutation results) still describes this part of the change.