Go node: furrowd drops child output when Wait races the stdout copy; supervisor test reads empty PID file #161

Two pre-existing races in the Go node turn up as intermittent go test -race failures in CI. Both reproduce on main and are independent of any open PR.

1. furrowd loses child output (cmd/furrowd, runChild). runChild starts io.Copy(conn, stdout) and calls cmd.Wait() concurrently, then waits for the copy. os/exec documents that Wait closes StdoutPipe, so it must not run before all reads finish. When the network write is slow, the tail of the child's output is dropped:

• CI symptom (Treat Claude usage caps and expired logins as fatal #156's first run): TestHappyPathRoundTripThroughDialer got "MARKER:", want "MARKER:ciphertext" (main_test.go:103).
• Also seen locally: TestChildServesRecordedStoreDirNotRawRunID returned an empty data dir (main_test.go:148), 4 failures in 20 runs on main.
• A deterministic probe that delays the first write by 200ms reproduces it on main every time, with copy child output: read |0: file already closed.
• Fix shape: wait for the copy goroutine to drain stdout before calling cmd.Wait().

2. internal/furrow supervisor test reads an empty PID file. In TestSupervisorCancellationKillsChildAndReturns (supervisor_test.go:167), waitForPID parses the file as soon as it exists, but the shell creates or truncates it before echo $$ writes the PID, which fails with strconv.Atoi: parsing "": invalid syntax. It reproduces in -count=10 runs on main. Fix shape: keep polling until the file contains a full line, or have the helper write to a temp file and rename it.

Item 1 is a real product bug: a slow client can lose the end of a child's output. Item 2 is only a test.