feat(advisor): snapshot builder + deterministic detectors

- internal/snapshot: compact cluster state via schedulerclient (nodes,
  PEND top-N most-starved with parsed pending_reason, cluster sums, queue
  garnish, recent exits). Field whitelists + stall-threshold filter keep
  the state a few KiB. Module-local PEND-reason table mirrors
  mcp-server/internal/pendreason.go (lift into schedulerclient is a P1
  cleanup).
- internal/detect: pend_stall / worker_health / util_skew / exit_burst
  triggers — threshold, then per-scenario cooldown, then dedup on a hash
  of the situation (job set + reasons / node set) so an unchanged cluster
  is never re-evaluated.

Co-Authored-By: Claude Code <noreply@anthropic.com>

src/frontend/advisor/internal/snapshot/pendreason.go

@@ -0,0 +1,95 @@
+package snapshot
+
+import (
+	"encoding/json"
+	"fmt"
+)
+
+// pendingReason mirrors the apiserver rawPendReason (convert_pure.go),
+// surfaced by schedulerclient as the JSON-encoded string
+// Extensions["pending_reason"].
+type pendingReason struct {
+	MainReasonName string `json:"main_reason_name"`
+}
+
+// reasonScenario maps a PREASON_* name to the human-readable scenario.
+// This mirrors mcp-server/internal/pendreason.go (module-local copy: the
+// advisor and mcp-server are separate modules; lifting the table into
+// schedulerclient is a deliberate P1 cleanup, not forgotten).
+func reasonScenario(name string) string {
+	switch name {
+	case "bad_resreq":
+		return "the resource request is invalid or cannot be satisfied"
+	case "no_resource":
+		return "insufficient resource capacity"
+	case "match_fail":
+		return "no host matches the job's host-selection condition"
+	case "limit_global":
+		return "a global quota is exhausted"
+	case "limit_host":
+		return "a host-level quota is exhausted"
+	case "queue_closed":
+		return "the queue is closed (not accepting new jobs)"
+	case "queue_draining":
+		return "the queue is draining (no new jobs until it empties)"
+	case "time_window":
+		return "outside the queue's dispatch window"
+	case "host_unavail":
+		return "the host is unavailable"
+	default:
+		return fmt.Sprintf("unknown reason (%s)", name)
+	}
+}
+
+// parsePendingReason extracts the main reason name + scenario from a job's
+// Extensions. Returns nil when absent or unparseable — a missing reason is
+// reported as such, never guessed.
+func parsePendingReason(ext map[string]any) *JobReason {
+	raw, ok := ext["pending_reason"]
+	if !ok {
+		return nil
+	}
+	var pr pendingReason
+	switch v := raw.(type) {
+	case string:
+		if v == "" {
+			return nil
+		}
+		if err := json.Unmarshal([]byte(v), &pr); err != nil {
+			return nil
+		}
+	case map[string]any:
+		b, _ := json.Marshal(v)
+		if err := json.Unmarshal(b, &pr); err != nil {
+			return nil
+		}
+	default:
+		return nil
+	}
+	if pr.MainReasonName == "" {
+		return nil
+	}
+	return &JobReason{
+		MainReasonName: pr.MainReasonName,
+		Scenario:       reasonScenario(pr.MainReasonName),
+	}
+}
+
+// round2 keeps snapshot numbers at two decimals so the JSON stays small.
+func round2(f float64) float64 {
+	return float64(int(f*100+0.5*sign(f))) / 100
+}
+
+func sign(f float64) float64 {
+	if f < 0 {
+		return -1
+	}
+	return 1
+}
+
+func frac(a, b int) float64 {
+	if b == 0 {
+		return 0
+	}
+	return float64(a) / float64(b)
+}