clanker talk doesn't restart Hermes bridge on crash — REPL error-spams until exit

[rafeegnash]

Problem

When the Hermes bridge process crashes mid-session, Prompt returns an error and the clanker talk REPL loop just prints it. Every subsequent prompt hits a dead bridge and the user sees error spam until they type exit. There is no health check or restart.

Where

• cmd/talk.go:111-114 — REPL loop's error handling
• internal/hermes/runner.go — no IsHealthy() or restart mechanism

Fix

1. In the REPL loop, detect "bridge process exited" errors and call runner.Stop() then start a fresh runner before accepting the next input
2. Cap restarts to 3 per minute before giving up
3. On exit, log the underlying cause (look at stderr or exit code captured in feat(k8s/cluster): add errorHint method to EKS provider #10's dispatch loop)

if isBridgeExitError(err) {
    if restartCount.IncrementWithinWindow() > 3 {
        fmt.Println("bridge keeps crashing, giving up")
        return err
    }
    runner.Stop()
    if err := runner.Start(ctx); err != nil { return err }
    continue
}

Acceptance criteria

• Manual test: kill the bridge process mid-session, next prompt restarts the bridge transparently
• Counter prevents infinite restart loops
• Error message on permanent failure includes the bridge stderr tail

[rafeegnash]

Fixed in upstream bgdnvk#187 (merged). cmd/talk now detects hermes.IsBridgeExitError, restarts the bridge transparently, and caps restarts at 3 per minute (sliding window). The error message includes the bridge stderr tail so the user sees the actual Python traceback. Will land here on the next fork sync.