[AR-BULK-01] AskRosetta bulk classification endpoint

Create a bulk classification endpoint for AskRosetta that accepts 100+ products and processes them asynchronously. ## Context AskRosetta's current `/api/classify` endpoint handles one product at a time synchronously. For enterprise customers (eBay, Estafeta, Viettel), we need batch processing. The webhook handler (AR-HOOK-01) will deliver results when done. Existing classification chain: `c8_platform/corridor_engine/classification_chain.py` - `classify(request)` → ClassificationResult (single product) - 2-tier: Rosetta Core (confidence 0.85) → Gemini fallback (0.70) → human_review - ClassificationRequest: product_description, image_data, origin_country, target_corridors, language ## What to generate File: `c8_platform/corridor_engine/bulk_classify.py` A Flask blueprint that: ### 1. POST /api/v1/classify/bulk Accepts batch of products: ```json { "products": [ { "product_id": "client-sku-123", "description": "Men's cotton t-shirt, crew neck", "origin_country": "CN", "image_url": null } ], "callback_url": "https://client.example.com/webhooks/classification", "priority": "normal" | "express", "corridor": "MX_US" } ``` ### 2. Behavior: - Validate API key from `X-Api-Key` header against Supabase `api_keys` table - Validate payload: max 500 products per batch, required fields - Create job record in `classification_jobs` table with status='queued' - Return immediately with job_id and estimated completion time: ```json {"job_id": "uuid", "status": "queued", "estimated_seconds": 120, "poll_url": "/api/v1/jobs/{job_id}"} ``` - Process products in background thread (or queue): - Batch into groups of 10 for parallel processing - Use `asyncio.gather()` or `concurrent.futures.ThreadPoolExecutor` - Call `classify()` from classification_chain for each product - Update job progress in Supabase after each batch - On completion, fire webhook to callback_url (AR-HOOK-01 format) - Handle partial failures gracefully (continue processing remaining) ### 3. Rate limiting: - Max 3 concurrent jobs per API key - Max 500 products per job - Express priority: process immediately, Normal: queued ### 4. Cost tracking: - Track per-product classification cost (Rosetta vs Gemini tier used) - Store in job metadata for billing ### 5. Error handling: - Individual product failures don't kill the batch - Failed products get `error` field in results - Job status = 'partial' if some products failed ### Patterns from existing app.py: ```python from classification_chain import classify, ClassificationRequest, ClassificationResult from flask import Blueprint, request, jsonify import os from supabase import create_client ``` ### Output: Complete Python file with Flask blueprint for bulk classification.