$ the-wire · showcase
gopls stops dropping cloned views, ssh fixes a permanent OpenChannel block
By RepoJournal · Filed · About Go · Composed from the cited sources · methodology
Two correctness fixes headline the day: gopls was leaking views out of its diagnosis queue after a snapshot clone, and crypto/ssh could hang OpenChannel forever when the connection died mid-open.
gopls/internal/server: don't forget views whose snapshot was cloned golang/tools
The server removed a view from viewsToDiagnose as soon as a diagnosis finished, checking only the modification context, which DidModifyFiles has already cancelled by cloning the snapshot, so a diagnosis of the stale snapshot could complete inside that window and the view was never diagnosed again. The window widens when DidModifyFiles waits on initialization, which is presumably why it surfaced...
ssh: fail OpenChannel when the connection is lost golang/crypto
mux.loop closed known channels once, at failure time, so a channel registered afterwards was never closed and OpenChannel blocked forever if its open message was written before the transport recorded the write error; a channel registered before that got a nil message and a misleading "unexpected packet" error. Channels are no longer registered after mux shutdown, and OpenChannel returns io.EOF ...
cmd/compile: don't apply make+copy optimization to side-effecting operands golang/go
The make+copy rewrite (m = make([]T, n); copy(m, s) becomes m = makeslicecopy(T, n, s)) evaluates s as an allocation operand, before make checks n, so a source with side effects or a possible panic reorders relative to the length check. The compiler now skips the optimization when the operand is side-effecting.
acme: bump pebble v2.10.1 -> v2.10.3 golang/crypto
Daniel McCarney's update to the acme test dependency notes it "fixes a Pebble order completion deadlock we've hit in CI" and brings ML-DSA support for testing future acme package changes. Only test infrastructure moves; the acme API does not.
go/types, types2: don't unpack prematurely in Named.TypeParams golang/go
In the long tail, go/types and types2 stop calling t.unpack() in Named.TypeParams, which could cache t.fromRHS before validType sets Typ[Invalid] during a cycle; gopls's LookupModOrZeroGrouped doc lost its pseudocode, and arm64's OrNot became a peephole folding x.Or(y.Not()) into VORN.