fix(supervisor): stop terminate() racing watchChild's Cmd.Wait()

[nhuelstng]

What

Fixes a data race between terminate() and watchChild()'s reaper goroutine, both of which called os/exec.Cmd.Wait() concurrently on the same sidecar process.

Why

This is the root cause of the agent did not exit within 5m0s E2E hangs investigated in #60/#61 — every matrix job registers at least one sidecar (echo-rest/self-audit), so every harness (including claude-code) is equally exposed, which is why #61's "shared GLM-backend contention" theory stopped holding up once claude-code hung too in a later run.

watchChild() (added in #38) spawns a goroutine that calls r.Cmd.Wait() on a sidecar as soon as it starts, so process.exit can be audited at real termination time. But terminate() (called by StopSidecar/ShutdownAll at teardown) also spawns a goroutine calling cmd.Wait() on the same *exec.Cmd. os/exec.Cmd.Wait must only ever have one caller in flight — calling it concurrently races on shared internal state, and the loser can block forever reading a channel the other caller already drained, even though the child process has actually exited. Once that happens, ShutdownAll's WaitGroup.Wait() never returns, so the parent omac start process (which the e2e test's cmd.Run() is blocked on) never exits either.

How

Running now carries a waitDone channel that watchChild closes after it reaps the child. terminate() waits on that channel instead of calling Cmd.Wait() itself whenever a reaper is already running (any child reachable via s.children — i.e. StopSidecar and ShutdownAll). The one call site that runs before watchChild starts (a health-check failure during sidecar startup) still spawns its own Wait(), since no reaper exists yet there.

Verification

• Goroutine dump from a real hang: reproduced locally (register echo-rest, omac start opencode --no-sandbox in a loop against a fake fast-exiting harness) and captured a SIGQUIT dump showing the main goroutine wedged in ShutdownAll's WaitGroup.Wait() → terminate() → Cmd.Wait(), forever.
• Deterministic repro, before/after: same loop hung on every iteration after the first, 100% reproducible, before this fix; 10/10 clean runs (~0.5s each) after, both under --no-sandbox and under the real bubblewrap sandbox.
• go test -race: added TestShutdownAllReapsLongRunningChildWithoutHanging — confirmed it fails with WARNING: DATA RACE between terminate.func1 and watchChild.func1 on the pre-fix code (reverted via patch to check), and passes clean post-fix.
• go build ./..., go vet ./..., gofmt -l, full go test ./... (non-e2e) — all clean.
• Full internal/supervisor suite passes under -race.

Relationship to #60 / #61

• fix(e2e): read audit-output.txt in TestE2EProvenance instead of chat text #60 (provenance test file-read fix) is unrelated and still valid.
• fix(ci): serialize GLM-backed E2E harnesses to reduce backend contention #61 (E2E workflow concurrency mitigation) was based on a theory (shared GLM backend contention) that this investigation disproved — recommend closing fix(ci): serialize GLM-backed E2E harnesses to reduce backend contention #61 in favor of this fix, since this addresses the actual cause directly rather than reducing exposure to it.