[milos-lk]

One more fix.

Follow-up to #66, which keeps a dead HTTP/2 connection from sitting until the idle deadline. This PR is about what happens to the requests when a connection ends anyway: the GCS client now retries the two errors with which Go's HTTP/2 client fails every open request on a connection it is closing.

How this differs from #66

• Health-check HTTP/2 connections with pings #66 changes detection: a silent peer is pinged and the connection is dropped after 45 s with client connection lost, which the SDK already retries. It prevents the stall from ever reaching the idle-deadline race, but it only helps when the peer has gone quiet.
• This PR changes classification: whatever closes a shared connection, the requests still open on it fail with one of two errors the SDK does not retry. After this PR it does. It covers two cases Health-check HTTP/2 connections with pings #66 cannot:
  • the server's own GOAWAY followed by a close before the in-flight streams finish (Google frontends do this at the end of a session's lifetime, debug session_timed_out); the peer is alive and answering pings the whole time;
  • a hard reset racing the same way as the stall, where there is no silence for a ping to notice.
• Without this PR, one such connection event loses every upload in flight on it; with it, each is re-sent on a fresh connection with the existing backoff.

What changes

• gcsShouldRetry is storage.ShouldRetry plus http2: client connection force closed via ClientConn.Close and http2: server sent GOAWAY and closed the connection, set client-wide with SetRetry(WithErrorFunc). Per-object retryers (gcsRetryAlways, the conditional write) inherit it and keep their own policy and backoff; the policy still decides which operations retry at all, this only widens which errors count.
• The README's Retries bullet states it.

Worth knowing

• Go re-sends only the streams a GOAWAY says the server never accepted (above LastStreamID); the ones already in flight get GoAwayError, which has no Temporary() and is not in the SDK's string list. Both error types are unexported in net/http, so a string match is the only handle, as the SDK uses for its own transport errors.
• Retrying either is as safe as the SDK's existing retries of "connection reset" or "broken pipe": the request may have been received, which the overwrite policy and the DoesNotExist precondition both tolerate.
• GCS is the only backend here whose endpoint negotiates HTTP/2, so no other store needs this.

Tests

• TestGCSRetriesConnectionClose: a TLS HTTP/2 server that takes the request's HEADERS, answers GOAWAY with that stream still open and closes, which yields the real GoAwayError from Go's client; asserts the SDK's predicate rejects it and ours accepts it. The force-closed case uses the error as net/http wraps it. A nil error, a 404 and a connection reset are classified as before.

🤖 Generated with Claude Code