Record annotated and warm-cache evidence

Add repeated production annotation timing, real app-server cache telemetry, the context-delta decision, and correct --judge=false handling.\n\nAI-assisted: OpenAI Codex.

site/data/live-codex-worker-result.ts

@@ -44,6 +44,9 @@ export const liveCodexWorkerResult = {
     annotatedQualityRuns: 3,
     coldInputTokens: 41066,
     coldOutputTokens: 2148,
+    warmRepeatInputTokens: 46973,
+    warmRepeatCachedInputTokens: 38656,
+    warmRepeatCacheRate: 0.823,
   },
   productionBrowser: {
     runs: 1,
@@ -55,6 +58,14 @@ export const liveCodexWorkerResult = {
     sourcePreflightMs: 52,
     validationMs: 10,
   },
+  annotatedBrowser: {
+    runs: 2,
+    firstVariantMs: 14316.21,
+    allVariantsMs: 45424.24,
+    comments: 1,
+    strokes: 1,
+    screenshotAttached: true,
+  },
   result: '4/4 full-task quality gate',
   passed: true,
   architecture: 'A dedicated persistent Live app-server connection and worker thread with native Impeccable skill, sub-command, project context, and annotation-image inputs. Accept interrupts unfinished variants; cleanup stays on the foreground control lane while the next generation can start independently.',

site/data/live-harnesses.ts

@@ -97,10 +97,10 @@ export const liveExperiments = [
   {
     rank: 8,
     title: 'Trim the producer contract — selected',
-    evidence: 'Progressive compact passed 5/5 for Claude Sonnet 4.6 and GPT-5.5. It was OpenAI’s fastest and cheapest progressive path; Claude’s other candidates all missed reliability.',
+    evidence: 'Progressive compact passed 5/5 for Claude Sonnet 4.6 and GPT-5.5. A repeated warm Sol turn cached 38,656 of 46,973 input tokens (82.3%) and completed in 21.8 s versus 26.0 s cold.',
     move: 'Keep the stable generation contract cacheable and send only the picked-element context plus the current action and identity lock.',
-    expected: 'Reduces prefill and output work without weakening the design brief.',
-    confidence: 'Five runs per candidate',
+    expected: 'Use provider caching instead of risky context deltas; always send current project and brand evidence.',
+    confidence: 'Candidate matrix + token telemetry',
   },
   {
     rank: 9,