[Sealdot]

AI assistance: The implementation and regression tests were developed with Codex, with a separate agent reviewing the cancellation path.

Problem

A pending fetch_read_body call retains an Arc to the response while awaiting Response::chunk(). Closing its resource-table entry alone does not interrupt that operation. When a streaming server stays silent, cancellation therefore waits for another chunk; in the local HTTP/1.1 reproduction, the server also keeps seeing an open connection.

Closes #3713.

Solution and key decisions

• Give response resources a close notification, and race the pending body read against it using the existing Tokio sync dependency.
• Store the response in Mutex<Option<Response>>, dropping it on cancellation, EOF or error. This also replaces the raw mutable Arc cast and serializes concurrent reads.
• Keep commands, permissions, JS cancellation hooks, response sentinels, dependencies and lockfiles unchanged. Include matching http / http-js patch entries and streaming cancellation documentation.

Tests

macOS arm64, stable Rust 1.97.1 and MSRV 1.90.0, using the checked-in lockfile:

• Original read algorithm plus the pending-read regression: fails at the expected 2-second cancellation timeout.
• Final crate: 23 tests pass with stable default features, stable all features, and MSRV default features. These are the same 23 tests across configurations, including 4 new tests.
• The new tests cover 3 real silent HTTP/1.1 TCP cases (pending read, cancel before read, and a second read waiting for the mutex), with server-observed EOF/reset assertions; the fourth is an in-memory normal EOF case. Repeated RID closure keeps its existing error behavior.
• Strict Clippy for all targets with default and all features; cargo fmt --all --check; HTTP guest Rollup build, ESLint and TypeScript; documentation format, paired changeset and license checks pass.
• cargo test --doc succeeds with 0 doctests.

Risks and follow-up

No native webview/IPC, Linux, Windows or mobile execution is claimed. HTTP/2 stream cancellation and TLS-specific transport behavior are not independently exercised by these tests; dropping a response does not imply closing a shared HTTP/2 TCP connection. ResourceTable::clear does not call Resource::close, so this change does not claim to fix webview teardown with outstanding reads. Upstream platform CI and maintainer review remain to be completed.