Add LLM-powered PR classification

[pavsoss]

Summary

Replaces heuristic-only PR classification with LLM-powered structured classification using the existing llmCall abstraction.

The classifier now analyzes PR title and body context through the shared LLM provider pipeline while preserving the existing heuristic logic as a graceful fallback.

Changes Made

LLM-Powered Classification

• Converted classifyPrAsAi to an async function.
• Added structured LLM classification using llmCall.
• Validates classification output with a Zod schema containing:
  • isAiGenerated
  • confidence
  • reason
• Uses the validated isAiGenerated result as the classifier decision.

Heuristic Fallback

• Preserved the existing classification logic in classifyPrHeuristic.
• Automatically falls back to the heuristic when:
  • all LLM providers fail
  • the LLM request returns an error
  • structured output cannot be validated

This preserves the existing deterministic behavior when AI classification is unavailable.

Async Pipeline Updates

Updated all functional callers of classifyPrAsAi to await the new async classifier:

• pr-backfill.ts
• process-pr-event.ts

Tests

• Updated existing heuristic tests to explicitly exercise the fallback path.
• Added coverage for AI-generated LLM classifications.
• Added coverage for genuine PR classifications.
• Verified heuristic fallback when llmCall returns an error.

Closes #696

[vercel]

@pavsoss is attempting to deploy a commit to the codersogs-3057's projects Team on Vercel.

A member of the Team first needs to authorize it.