v1 strips the incoming accept-encoding header before proxying (#914). The v2 ignoredHeaders set in src/utils/internal/proxy.ts added accept but dropped accept-encoding, so getProxyRequestHeaders forwards the client's accept-encoding to the target.

proxy() removes content-encoding from the upstream response, relying on the runtime's fetch having already decoded the body. Forwarding the client's accept-encoding lets the target pick an encoding the runtime's fetch may not decode (the zstd-on-undici case from nitrojs/nitro#2880 that #914 addressed); the still-encoded body would then be passed on without a content-encoding header. Since h3 v2 runs across runtimes whose fetch clients support different encoding sets, this restores the v1 behavior by adding accept-encoding back to the set so the runtime fetch negotiates and decodes the encoding itself.

Added a regression test: proxying a request that sends accept-encoding: zstd, br, gzip now asserts the upstream does not receive it. Without the change the header is forwarded (web and node test targets); with it the header is dropped. Full suite passes.

Ref: #914

Summary by CodeRabbit

• Bug Fixes
  • Updated proxy handling so incoming accept-encoding headers are no longer forwarded upstream.
  • This helps ensure proxied requests use the expected header behavior.
• Tests
  • Added coverage verifying that accept-encoding is stripped before reaching the upstream request.