bc62ad02

build: the toolchain moves to go 1.27

a73x   2026-08-20 18:14

build: the toolchain moves to go 1.27

go.mod's language version is this repo's only Go pin — the server image
copies prebuilt binaries rather than building in a golang container — so
one line moves every build, and with it encoding/json's v2-backed
unmarshal and size-specialized small-object allocation.

Three things had to move with it. golangci-lint v2.12.2 bundles a
staticcheck whose IR builder panics on 1.27's field-selector struct
literal keys, and it dies analysing the 1.27 stdlib itself; v2.13.0
carries the staticcheck that understands them. `go install` always
writes the same unversioned path, so the old file target could not tell
a stale linter from a current one — lint now checks what is actually
installed, because a linter built by an older Go refuses a newer
language version outright and the failure looks nothing like a bug in
the code being linted.

That newer staticcheck can see that syncsvc's Serve and the ssh gate's
Serve leave their accept loops only on a non-nil error, so the guards
at both call sites were always taken. The sites now say so directly.

`go fix`'s 1.27 modernizers ran over the whole tree, kept where the
result reads better than what it replaced. hyperlog's tail scan walks a
slice backwards by index; slices.Backward says that directly. And a
HostState carries its Report embedded, so a test that filled one in had
to name the embedded type just to reach a field on it — 1.27 lets a
struct literal take any valid field selector as a key, so the field is
set where it is read.

The atomictypes and unsafefuncs modernizers found nothing here. embedlit
is not applied wholesale: on the `ssh.Permissions{Extensions: ...}`
shape it drops the `Extensions` key and leaves the map entries dangling
in the outer literal, which does not compile (4 sites). Its other
rewrites are valid but fold a call into a literal, or park a closing
brace at the end of the last field, and neither reads better than what
is there.

The coverage table is re-baselined because 1.27 counts statements more
finely: a closure body is its own block instead of folding into the
statement that declares it. No test changed. 25 floors rise, where the
finer blocks credit covered code the old accounting merged away, and
two fall — server/boot 33 to 22 and gateclient 60 to 56, both mostly
wiring closures that nothing exercises. Both cover more statements than
they did under 1.26; only the denominator got honest.