Answer the formatter's common Accept headers from a table built once

[ericproulx]

Summary

• An API that pins no format (content negotiation, e.g. default_format :json) negotiated one from the Accept header on every request that sent one, through Rack::Utils.best_q_match: about 1.7 µs for Accept: application/json. The answer depends only on the header and the middleware's mime_types.
• Formatter::FormatForAcceptCache computes it once for every registered media type, */* and no Accept header, and format_from_header looks the header up. Any other header misses and is negotiated as before. Only registered media types are keys, so client input cannot grow a table.
• Shared per mime_types, which Grape::ContentTypes already shares per content-type registry, since every endpoint builds its own formatter. An API that pins a format never negotiates from the header and builds no table.
• The negotiation moved into the class method Formatter.format_for_accept (@api private) so the cache and the miss path run the same function. The keys are valid strings that scrubbing leaves alone, so only a miss scrubs the header.

Benchmarks

Median of 7 interleaved subprocess rounds against master, Ruby 4.0.6:

request, API with no pinned format	no JIT	YJIT
Accept: application/json	+40.3%	+54.1%
Accept: */*	+5.3%
pinned-format GET (control)	+0.2% (noise)

Boot, 500 negotiating endpoints: retained memory and compile time unchanged.

Behaviour

Byte-identical to master over a 525-case matrix: 7 APIs (defaults, default_format, a custom text/csv type, a mixed-case registered type, a registered application/*, a parameterized application/json; charset=utf-8, a pinned format) × 25 Accept values (a browser Accept, axios's default, q-value lists, other casings, invalid bytes, binary-encoded headers) × 3 paths (plain, .json extension, ?format=xml).

Test plan

• Full RSpec suite passes locally.
• RuboCop clean.
• Mutation-checked through existing specs: a table answering :txt for every key fails 7; skipping the scrub on a miss fails 1.
• CI green.

🤖 Generated with Claude Code

[github-actions]

Danger Report

No issues found.

View run

[github-actions]

Danger Report

Errors

• One of the lines below found in CHANGELOG.md doesn't match the expected format. Please make it look like the other lines, pay attention to version numbers, periods, spaces and date formats.

Markdowns

* [#XXXX](https://github.com/ruby-grape/grape/pull/XXXX): Answer the formatter's common Accept headers from a table built once instead of negotiating them per request - [@ericproulx](https://github.com/ericproulx).
does not include a pull request link

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.