LGTM. Two coordinated changes that together unblock the test:

socket.destroy() → socket.end(cb) then destroy — correct fix for a real bug. handle.hasPendingData() returns false once nghttp2's internal want_read/want_write are both 0 (session terminated, GOAWAY drained from nghttp2's side), but the GOAWAY bytes are still in the JS socket's writable-stream buffer waiting for the OS write. socket.destroy() would discard those, so the peer never sees the COMPRESSION_ERROR(9) GOAWAY and never surfaces ERR_HTTP2_SESSION_ERROR. socket.end(cb) flushes first, then FIN, then the cb destroys for good measure (some peers don't surface an error on a half-open FIN alone). The !socket.destroyed re-check inside the cb handles the race where the socket got destroyed by another path while the writable was draining.

The condition guarding this block (!handle.hasPendingData() && !this.destroyed, fired from socket.on('data')) only triggers post-session-termination — for a normal idle HTTP/2 connection want_read stays non-zero, so this won't fire on every quiet receive. Verified has_pending_data in ext/node/ops/http2/session.rs:2124 is want_write != 0 || want_read != 0.

Internal internal/http2/core shim — the upstream test does require('internal/http2/core').ServerHttp2Session for an instanceof assertion. Standard pattern for --expose-internals Node compat tests. The new file just re-exports the four classes (ClientHttp2Session, Http2Session, Http2Stream, ServerHttp2Session) which all already exist in ext/node/polyfills/http2.ts (lines 1363, 3048, 3512, 3638) — verified the names line up.

Test enrollment alphabetically positioned correctly between no-wanttrailers-listener and options-server-request.

CI still building, no failures on completed jobs.