perf: add model pinning infrastructure + rate-limit telemetry to E2E runner

Default E2E model changed from Opus to Sonnet (5x faster, 5x cheaper).
Session runner now accepts `model` option with EVALS_MODEL env var override.
Added timing telemetry (first_response_ms, max_inter_turn_ms) and wall_clock_ms
to eval-store for diagnosing rate-limit impact. Added EVALS_FAST test filtering.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

test/helpers/eval-store.ts

@@ -42,6 +42,11 @@ export interface EvalTestEntry {
   timeout_at_turn?: number;   // which turn was active when timeout hit
   last_tool_call?: string;    // e.g. "Write(review-output.md)"
 
+  // Model + timing diagnostics (added for Sonnet/Opus split)
+  model?: string;                // e.g. 'claude-sonnet-4-6' or 'claude-opus-4-6'
+  first_response_ms?: number;    // time from spawn to first NDJSON line
+  max_inter_turn_ms?: number;    // peak latency between consecutive tool calls
+
   // Outcome eval
   detection_rate?: number;
   false_positives?: number;
@@ -65,6 +70,7 @@ export interface EvalResult {
   failed: number;
   total_cost_usd: number;
   total_duration_ms: number;
+  wall_clock_ms?: number;     // wall-clock from collector creation to finalization (shows parallelism)
   tests: EvalTestEntry[];
   _partial?: boolean;  // true for incremental saves, absent in final
 }
@@ -546,6 +552,7 @@ export class EvalCollector {
   private tests: EvalTestEntry[] = [];
   private finalized = false;
   private evalDir: string;
+  private createdAt = Date.now();
 
   constructor(tier: 'e2e' | 'llm-judge', evalDir?: string) {
     this.tier = tier;
@@ -615,6 +622,7 @@ export class EvalCollector {
       failed: this.tests.length - passed,
       total_cost_usd: Math.round(totalCost * 100) / 100,
       total_duration_ms: totalDuration,
+      wall_clock_ms: Date.now() - this.createdAt,
       tests: this.tests,
     };