fix(worker): 🐛 Fix ConverterWorker not picking up pages after split

[jorben]

Summary

• Fix a bug where ConverterWorker fails to process pages after PDF splitting because orphaned TaskDetail records from terminal tasks (FAILED, CANCELLED, etc.) exhaust claimPage() retry attempts
• Refactor claimPage() to pre-filter active tasks before searching for PENDING pages
• Add cleanup logic in WorkerOrchestrator to mark orphaned PENDING pages as FAILED on startup
• Improve test connection endpoint with realistic image and token limit

Changes

Core Fix — ConverterWorker.claimPage()

• Before: Query all PENDING TaskDetail records first, then verify parent Task status one-by-one. If 5+ orphaned records exist from terminal tasks, all maxAttempts are exhausted and valid pages are never reached.
• After: Pre-query Task table for active tasks (PROCESSING/COMPLETED), then only search for PENDING pages within those task IDs. Orphaned records are completely bypassed.

Startup Cleanup — WorkerOrchestrator.cleanupOrphanedWork()

• Add a new cleanup step that finds tasks in terminal states (CREATED, FAILED, CANCELLED, COMPLETED, PARTIAL_FAILED) and marks their remaining PENDING TaskDetail records as FAILED
• This prevents orphaned pages from accumulating over time
• Update CleanupResult interface with new orphanedPendingPages field

Minor — completion.handler.ts

• Use a meaningful 32×32 image instead of 1×1 pixel for connection testing
• Add maxTokens: 8 cap to limit cost during test connection
• Update prompt for better vision capability testing

Test Plan

• All 759 tests pass (35 test files)
• TypeScript compilation clean (tsc --noEmit)
• New tests for claimPage(): no active tasks, search within active task IDs
• New test for orphaned PENDING page cleanup in cleanupOrphanedWork()
• Updated existing tests to match refactored query strategy

Refs #19

🤖 Generated with Claude Code