quest: plan moq fetch close, terminal dial errors, and the interop close log

quest/m1/README.md

@@ -201,3 +201,6 @@ ladders.
 - [WebGPU on Safari](/quest/m1/webgpu-safari.md) - the WebGPU renderer is verified on Safari 26 for macOS and iOS
 - [A/V sync across a break](/quest/m1/watch-break-av-sync.md) - `@moq/watch` never plays pre-break audio out of sync with pre-break video
 - [moqsrc reconnect](/quest/m1/moqsrc-reconnect.md) - `moqsrc` redials after losing its relay and resumes on the same pads
+- [moq fetch closes cleanly](/quest/m1/moq-fetch-close.md) - `moq fetch` closes its session before exiting instead of leaving the relay to time it out
+- [Terminal dial errors](/quest/m1/moq-tokio-terminal-errors.md) - moq-tokio's reconnect loop stops on errors that can never succeed
+- [Interop close log](/quest/m1/interop-close-log.md) - the interop idle-out check reads the relay close log at a pinned level

quest/m1/interop-close-log.md

@@ -0,0 +1,22 @@
+# [XS] The interop idle-out check reads the relay's close log at a fixed level
+
+## Goal
+
+The interop harness's idle-out check never misreads a clean close as an open
+connection because of the relay's log level.
+
+## Plan
+
+Found while merging #5174 (2026-10-10): the relay logs a clean
+`connection closed` at info and an error close at warn, and the harness
+counts those lines to decide every connection closed. If the interop relay
+ever runs at warn, every clean close would read as still open and fail the
+round.
+
+Decided 2026-10-10: pin the harness's relay log filter for the close-log
+target to info in `test/interop/interop.sh`, so the product keeps warn for
+error closes. Rejected: logging both at one level in moq-relay. Also fail
+the round loudly when the relay log shows no connection for it, so a filter
+that hides the `conn{id=...}` and close lines can't pass as nothing open.
+
+Public API: none. Wire: none.

quest/m1/moq-fetch-close.md

@@ -0,0 +1,26 @@
+# [XS] moq fetch closes its connection cleanly
+
+## Goal
+
+`moq fetch` (`rs/moq-cli/src/fetch.rs`) closes its session before exiting,
+so the relay sees a clean close instead of timing the connection out.
+
+## Plan
+
+Found while merging #5174 (2026-10-10), which taught every interop client to
+close cleanly and made the harness fail a round on an idled-out connection.
+`moq fetch` still drops its connection without `Connection::close` or
+`Client::close`; the js-native FETCH_DATA wire-compat cell reproduced the
+relay timing it out. Only the wire-compat lanes reach it, and they skip the
+idle-out check, so it doesn't fail CI today.
+
+Close the session the way #5174 did for the other clients, on every exit
+after connecting, including an error or the `timeout_at` deadline. For the
+regression, run a FETCH round with the current relay and the current
+`moq fetch` under the idle-out check. No such round exists today (FETCH only
+runs in the wire-compat lanes), so add one and wire it into CI. Don't enable
+the check for the wire-compat cells: they mix in released relays and clients,
+which it excludes because they may not close cleanly or log which connection
+closed.
+
+Public API: none. Wire: none.

quest/m1/moq-tokio-terminal-errors.md

@@ -0,0 +1,27 @@
+# [S] moq-tokio's reconnect loop gives up on errors that can never succeed
+
+## Goal
+
+With unlimited backoff, moq-tokio's reconnect loop stops and reports an error
+that can never succeed on retry (a local dial or configuration error such as
+`UnsupportedScheme`, an invalid URL, or a bad TLS configuration) instead of
+retrying it forever. Transient errors keep retrying.
+
+## Plan
+
+Found while planning [moqsrc reconnect](/quest/m1/moqsrc-reconnect.md)
+(2026-10-10): `moqsink` already reconnects with `backoff.timeout = 0`, and a
+mistyped scheme makes it retry forever without a bus error. `moqsrc` will do
+the same once it reconnects.
+
+Classify dial errors at the source: the loop already treats non-retryable
+CONNECT statuses (`status_retryable`) as terminal; extend that to local
+errors, listed by type rather than by message. The dial races QUIC and
+WebSocket, so a local error is terminal only when every raced transport fails
+terminally, as the CONNECT status check already requires; one transport
+rejecting a scheme the other accepts must keep dialing. Fail loud with the
+error.
+Test with mocked time: an unsupported scheme fails at once under unlimited
+backoff, and a refused TCP connect still retries.
+
+Public API: behavior only (the loop ends with the error). Wire: none.