Fix AI assignment workload imbalance: enforce caps and rebalance

Root cause: batches of 15 projects were processed independently -
GPT didn't see assignments from previous batches, so expert jurors
got assigned 18-22 projects while others got 4-5.

Fixes:
- Track cumulative assignments across batches (feed to each batch)
- Calculate ideal target per juror and communicate to GPT
- Add post-processing rebalancer that enforces hard caps and
  redistributes excess assignments to least-loaded jurors
- Calculate sensible default max cap when not configured
- Reweight prompt: workload balance 50%, expertise 35%, diversity 15%

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/server/routers/roundAssignment.ts

@@ -134,7 +134,13 @@ export const roundAssignmentRouter = router({
 
       // Build constraints
       const configJson = round.configJson as Record<string, unknown> | null
-      const maxPerJuror = (configJson?.maxAssignmentsPerJuror as number) ?? undefined
+      const configuredMax = (configJson?.maxAssignmentsPerJuror as number) ?? undefined
+
+      // If no explicit cap, calculate a balanced one: ceil(total_needed / juror_count) + 2 buffer
+      const totalNeeded = projectStates.length * input.requiredReviews
+      const jurorCount = round.juryGroup.members.length
+      const calculatedMax = Math.ceil(totalNeeded / jurorCount) + 2
+      const maxPerJuror = configuredMax ?? calculatedMax
 
       const constraints = {
         requiredReviewsPerProject: input.requiredReviews,