chore: support MoonBit 0.10.10

[brickfrog]

Why

The repo does not compile under moonc 0.10.10. Two dependencies stopped parsing, for two separate undocumented language changes:

• Loop-else was renamed to nobreak. moonbitlang/protobuf@0.1.1 uses for ... else in reader.mbt and async_reader.mbt → four E3002 parse errors.
• A function-local struct declaration no longer parses. moonbitlang/async@0.19.0 declares one in js_async.mbt:116 → seven errors, js only, so a native-only check would have hidden it. js is a CI target.

Changes

Dependencies and toolchain

moonbitlang/protobuf 0.1.1 → 0.1.2 and moonbitlang/async 0.19.0 → 0.21.1. Upstream protobuf is already at 0.1.3, but mooncakes only publishes through 0.1.2, which compiles clean and needs no call-site changes.

The pinned CI toolchain in .github/scripts/setup-moonbit.sh moves 0.10.6 → 0.10.10+f8a486b6f, with digests computed from the archives. The script's layout assumptions still hold on the new tarball (no moonx binary shipped, so the ln -s does not collide; internal/tcc present) and both moon bundle steps were run against it locally.

HTTP headers

moonbitlang/async/http 0.21 keys headers by CaseInsensitiveString, so Map[String, String] no longer type-checks at @http.post_stream or on Response::headers. A key converts implicitly from String; a whole Map does not. Added to_http_headers at the @http boundary so the public OtlpHttpClient trait keeps plain Map[String, String].

Retry-After is now read straight off the case-insensitive map instead of going through parse_retry_after_ms, which probed the two literal casings retry-after and Retry-After. Casings such as Retry-after are now honored. parse_retry_after_ms(Map[String, String]) stays public and unchanged for callers holding a plain map.

New 0.10.10 diagnostics

E0092 fragile_catch_all flagged three catch-all handlers that cleaned up and re-raised. All three now release with defer, so state restore and span exit also run on the async cancellation paths a catch-all will stop capturing in a future compiler:

// before
try f() catch { err => { restore_trace_state(snapshot); raise err } }
  noraise { result => { restore_trace_state(snapshot); result } }

// after
defer restore_trace_state(snapshot)
f()

span_async cannot go all the way: errdefer runs on error and on cancellation but binds no error value, so recording the message needs a catch. The cleanup moved to defer, the handler now only inspects err, and the lint is suppressed locally with #warnings("-fragile_catch_all") and a comment.

E0020 on StringBuilder::new → StringBuilder(), two sites.

Audit

Nothing else in the 0.10.7 → 0.10.10 range reaches this repo. No @immut/*, BigInt, from_octets, Deque, @diff, Iter2, Char::length, is_prefix/is_suffix, nocancel identifier, @debug.to_repr, or :embed prebuild. @json.parse appears only in tests, over payloads we generate. The trim(chars=) SIMD bug needs an empty character set; env_filter's is the non-empty ASCII constant " \t\r\n".

The default backend changing to wasm is a non-event here: moon.mod sets preferred_target = "native", and CI passes --target explicitly. The @fs create_mode default flip does not reach us either — every write_file/open call passes create_mode explicitly, and the removed create/truncate parameters were unused.

W0079 implicit_impl_as_method has 98 hits behind --warn-list +implicit_impl_as_method, all on derive(...) declarations. It is disabled by default and nothing breaks yet; migrating to explicit extend blocks is left out of this change.

Verification

• moon check --deny-warn on native, js, wasm-gc and wasm: 0 warnings, 0 errors.
• moon test: native 293/293, js 280/280, wasm-gc 265/265.
• moon build --target native --release green.
• moon info produces no .mbti diff — the public API is unchanged.
• moon fmt idempotent.

client_test.mbt now has its test server send Retry-after, a casing the old two-literal probe would have missed, so the existing retry_after_ms == Some(2000) assertion exercises the case-insensitive lookup against a live @http.Server.