Add agent restart lifecycle command

Extends the existing CheckIn RPC with a one-shot AgentCommand
(restart only -- stop/uninstall need real per-platform OS
service-manager integration and stay deliberately out of scope),
delivered at-most-once: cleared the instant it's handed to the agent
in a response, since a restarting agent's process is gone before it
could ever confirm receipt. On restart, the agent flushes whatever's
buffered, aborts its source task, and exits cleanly, relying entirely
on the host's own service manager to bring it back up.

Issuing a command is gated at RoleAdmin (stricter than config
editing's RoleEditor) and logged into the same audit_log table Phase
7's AI interactions use, via a new agent_command event type.

A real bug was found and fixed during live verification: the first
implementation tried to atomically read-and-clear pending_command in
a single INSERT...ON CONFLICT statement using a sibling CTE
referenced only from RETURNING, on the assumption that Postgres
evaluates every part of a WITH query against one pre-statement
snapshot. That's wrong specifically for FOR UPDATE, which always
reads the latest row version including one written earlier in the
same statement -- confirmed empirically (a restart command was
always coming back empty even when genuinely pending, so the agent
never received it). Fixed by splitting into two real, ordered
statements inside one explicit transaction.

See /docs/agent-management-design.md's "Lifecycle commands" section.

api/agents/handler.go

@@ -18,28 +18,58 @@ type store interface {
 	Get(ctx context.Context, tenantID, host string) (*Agent, error)
 	SetOverride(ctx context.Context, tenantID, host string, override ConfigOverride, updatedBy string) (*Agent, error)
 	ClearOverride(ctx context.Context, tenantID, host string) error
+	IssueCommand(ctx context.Context, tenantID, host, command, issuedBy string) (*Agent, error)
+}
+
+// CommandLogger records an issued lifecycle command into the Phase 4
+// audit trail -- same nil-by-default, fail-open shape as
+// aiapi.InteractionLogger and queryapi.AuditLogger: a single-tenant
+// deployment with no enterprise/ configured just doesn't log these.
+// enterprise/internal/audit supplies the real implementation
+// (event_type = 'agent_command', see
+// metadata/migrations/0039_add_agent_command_event_type.sql) -- this is
+// the one entry point in this package genuinely worth logging even
+// without enterprise/ wired, given "strict RBAC, full audit trail" was
+// the explicit precondition for building lifecycle commands at all (see
+// /docs/agent-management-design.md); it degrades gracefully rather than
+// being required, matching every other optional audit hook in this
+// codebase, but a real deployment should wire it.
+type CommandLogger interface {
+	LogCommand(ctx context.Context, entry CommandLogEntry) error
+}
+
+type CommandLogEntry struct {
+	Host     string
+	Command  string
+	IssuedBy string
 }
 
 type Handler struct {
 	logger     *slog.Logger
 	store      store
 	authorizer authz.Authorizer
+	commands   CommandLogger
 }
 
-func NewHandler(logger *slog.Logger, store store, authorizer authz.Authorizer) *Handler {
-	return &Handler{logger: logger, store: store, authorizer: authorizer}
+// commands may be nil -- see CommandLogger's doc comment.
+func NewHandler(logger *slog.Logger, store store, authorizer authz.Authorizer, commands CommandLogger) *Handler {
+	return &Handler{logger: logger, store: store, authorizer: authorizer, commands: commands}
 }
 
 // RegisterRoutes: viewing inventory is RoleViewer (same bar as viewing
 // a dashboard); editing an agent's remote config is RoleEditor -- an
 // operational-tuning action, not an admin-only one, matching the RBAC
 // matrix's treatment of alert rules/notification targets rather than
-// user/role management.
+// user/role management. Issuing a lifecycle command is RoleAdmin --
+// stricter than config editing, matching the matrix's treatment of
+// similarly consequential actions (e.g. deleting a notification
+// target) rather than day-to-day tuning.
 func (h *Handler) RegisterRoutes(mux *http.ServeMux) {
 	mux.HandleFunc("GET /agents", authz.RequireRole(h.authorizer, authz.RoleViewer, h.handleList))
 	mux.HandleFunc("GET /agents/{host}", authz.RequireRole(h.authorizer, authz.RoleViewer, h.handleGet))
 	mux.HandleFunc("PUT /agents/{host}/config", authz.RequireRole(h.authorizer, authz.RoleEditor, h.handleSetConfig))
 	mux.HandleFunc("DELETE /agents/{host}/config", authz.RequireRole(h.authorizer, authz.RoleEditor, h.handleClearConfig))
+	mux.HandleFunc("PUT /agents/{host}/command", authz.RequireRole(h.authorizer, authz.RoleAdmin, h.handleIssueCommand))
 }
 
 // tenantID mirrors dashboards.Handler.tenantID exactly -- resolved from
@@ -106,6 +136,43 @@ func (h *Handler) handleSetConfig(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, a)
 }
 
+type issueCommandRequest struct {
+	Command string `json:"command"`
+}
+
+// handleIssueCommand queues a one-shot lifecycle command -- see
+// Store.IssueCommand's doc comment for the delivery/clearing semantics.
+// Logs to CommandLogger fail-open (a write failure is logged server-
+// side and otherwise ignored, same posture as aiapi's interaction
+// logging): audit-trail completeness matters, but it shouldn't be able
+// to turn a legitimate restart request into a 500.
+func (h *Handler) handleIssueCommand(w http.ResponseWriter, r *http.Request) {
+	var req issueCommandRequest
+	if !decodeJSON(w, r, &req) {
+		return
+	}
+	if !validCommand(req.Command) {
+		writeError(w, http.StatusBadRequest, `command must be "restart"`)
+		return
+	}
+
+	host := r.PathValue("host")
+	issuedBy := h.updatedBy(r)
+	a, err := h.store.IssueCommand(r.Context(), h.tenantID(r), host, req.Command, issuedBy)
+	if err != nil {
+		h.writeStoreErr(w, err, "issuing agent command")
+		return
+	}
+
+	if h.commands != nil {
+		if err := h.commands.LogCommand(r.Context(), CommandLogEntry{Host: host, Command: req.Command, IssuedBy: issuedBy}); err != nil {
+			h.logger.Error("logging agent command to audit trail", "host", host, "command", req.Command, "error", err)
+		}
+	}
+
+	writeJSON(w, http.StatusOK, a)
+}
+
 func (h *Handler) handleClearConfig(w http.ResponseWriter, r *http.Request) {
 	if err := h.store.ClearOverride(r.Context(), h.tenantID(r), r.PathValue("host")); err != nil {
 		h.writeStoreErr(w, err, "clearing agent config")