feat(auth): add PAT auth, scoped REST/MCP access, CSRF, and tenant isolation

Introduce Personal Access Tokens (nzp_*) as a stateless auth path alongside
JWT, gated per-endpoint by a scope middleware (nezha:{resource}:{verb}) with
fail-closed empty-scope defaults and a server-id whitelist. Self-management
endpoints (profile, api-tokens, oauth2 bind, refresh-token) explicitly reject
PATs to block privilege-escalation chains. A revoke registry tears down active
long-lived connections (terminal, fm, ws, transfer, mcp) the moment a PAT is
deleted, with a tombstone closing the revoke->register race.

Add an MCP endpoint that proxies tool calls (exec, fs read/write/delete,
transfer) to agents over gRPC, guarded by origin/DNS-rebinding checks, a
per-token rate limiter, audit logging, and a kill switch. Serialize all
sends through the IOStream wrapper to honour grpc-go's concurrency contract.

Add CSRF double-submit protection on unsafe cookie-authenticated methods,
exempting authenticated PAT requests by context identity (not a forgeable
Authorization header). Apply visibility/whitelist filtering consistently
across list, get-by-id, and mutate paths to enforce tenant isolation.

Migrate legacy mcp:* scopes: rewrite read/exec to nezha:* equivalents and
drop dangerous write/delete/wildcard grants.

Co-authored-by: cloudcode <cloudcode@users.noreply.github.com>

service/rpc/mcp_kill_switch_registration_race_test.go

@@ -0,0 +1,80 @@
+package rpc
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/nezhahq/nezha/model"
+)
+
+// Registration-after-sweep race (review issue #1): CancelAllMCPInflight only
+// cancels entries already present in the inflight map. A CallAgent that passes
+// the upfront kill-switch check but has not yet Store()d its entry is invisible
+// to the sweep, so without a post-registration re-check it goes on to SendTask
+// a fresh exec/fs task to the agent AFTER EnableMCP=false.
+//
+// This test drives the worst-case interleaving deterministically:
+//   1. CallAgent passes the upfront observer check (observer still false).
+//   2. The operator flips the observer to "disabled" and runs the cancel sweep
+//      while CallAgent is paused between the check and Store.
+//   3. CallAgent resumes; it MUST observe the kill switch on the post-Store
+//      re-check and return ErrMCPDisabled WITHOUT sending the task.
+//
+// With the race present, CallAgent sends the task and blocks until timeout
+// (ErrAgentTimeout) — the agent received a fresh task past the kill switch.
+func TestCallAgent_KillSwitchBeatsRegistrationAfterSweep(t *testing.T) {
+	const target uint64 = 7401
+
+	stream := newFakeStream()
+	cleanup := installFakeServer(t, target, stream)
+	defer cleanup()
+
+	var killed bool
+	var mu sync.Mutex
+	prev := mcpKillSwitchObserver()
+	// Observer returns the operator-controlled flag. CallAgent reads it both
+	// before and (with the fix) after registering the inflight entry.
+	SetMCPKillSwitchObserver(func() bool {
+		mu.Lock()
+		defer mu.Unlock()
+		return killed
+	})
+	t.Cleanup(func() { SetMCPKillSwitchObserver(prev) })
+
+	// Fail loudly if the agent ever receives a task: that means a fresh call
+	// leaked past the kill switch.
+	leaked := make(chan *struct{}, 1)
+	go func() {
+		select {
+		case <-stream.sent:
+			leaked <- nil
+		case <-time.After(2 * time.Second):
+		}
+	}()
+
+	// Arrange the interleaving: hook fires once CallAgent is about to register,
+	// flipping the kill switch and running the cancel sweep so the not-yet-Stored
+	// entry is missed by the sweep.
+	testKillSwitchAfterUpfrontCheck = func() {
+		mu.Lock()
+		killed = true
+		mu.Unlock()
+		CancelAllMCPInflight()
+	}
+	t.Cleanup(func() { testKillSwitchAfterUpfrontCheck = nil })
+
+	_, err := CallAgent(context.Background(), target, model.TaskTypeExec,
+		model.ExecRequest{Cmd: "x"}, 1*time.Second)
+
+	if !errors.Is(err, ErrMCPDisabled) {
+		t.Fatalf("CallAgent must return ErrMCPDisabled when the kill switch fires during registration; got %v", err)
+	}
+	select {
+	case <-leaked:
+		t.Fatal("a fresh MCP task leaked to the agent past the kill switch")
+	default:
+	}
+}