feat: Add Git Commit Forensics for granular template and AI-clone detection

[Shreesh-Sree]

Problem Statement

The current evaluation pipeline relies purely on repository descriptions, languages, and star counts to evaluate a candidate's projects. This creates a significant blind spot where the agent cannot accurately differentiate between:

1. A candidate who cloned a large tutorial or template repository and made a single, trivial commit.
2. A candidate who authentically contributed to an existing open-source project over an extended period.
3. A candidate who iteratively built a complex system from scratch.

Without commit forensics, the LLM frequently misinterprets cloned repositories as high-impact original work.

Proposed Solution

This PR introduces Git Commit Forensics into the retrieval pipeline. By querying the GitHub contributors endpoint, the system extracts the candidate's exact commit volume and compares it against the repository's total commit volume. This data is injected into the prompt context, allowing the LLM to accurately deduce authentic development effort.

Key Changes

1. GitHub API Enhancements (github.py):
   • Implemented fetch_repo_contributors to retrieve contribution statistics.
   • Implemented fetch_contributions_count to calculate author_commit_count versus total_commit_count.
   • Enhanced fetch_all_github_repos to dynamically classify projects as open_source (multiple contributors) or self_project (single contributor).
2. Schema & Model Updates (models.py):
   • Updated the Project model schema to ingest the new forensic metrics.
3. LLM Context Injection:
   • The evaluation context now receives exact commit ratios, explicitly instructing the LLM to penalize projects that have massive codebases but minimal author commits.

Proof of Enhancement & Validation

Scenario A: The Cloned Tutorial Repository

Before this PR:
The agent evaluates a "React E-Commerce Dashboard" repository with 50 stars and assigns a 25/30 Self-Project score, praising the complex architecture.

After this PR:
The LLM context receives the following structural data:

{
  "name": "react-ecommerce-dashboard",
  "project_type": "self_project",
  "author_commit_count": 1,
  "total_commit_count": 142
}

Result: The LLM correctly identifies this as a cloned template. The score is reduced to 5/30 with the corresponding evidence: "Candidate claims ownership but only authored 1 out of 142 commits; highly likely to be a cloned tutorial or boilerplate."

Scenario B: Authentic Open Source Contribution

Before this PR:
The agent identifies a fork of facebook/react on the candidate's profile and incorrectly assumes the candidate authored the entire framework, heavily skewing the score.

After this PR:
The LLM context receives the following structural data:

{
  "name": "react",
  "project_type": "open_source",
  "author_commit_count": 12,
  "total_commit_count": 15830
}

Result: The LLM accurately identifies this as a legitimate open-source contribution, scoring it under the open_source criteria rather than self_projects, and rewards the candidate proportionally for their 12 commits to a major production codebase.

Testing Performed

• Verified github.py gracefully handles rate limits when fetching contributors for large profiles.
• Verified score.py accurately passes the author_commit_count to the Jinja templates.
• Evaluated locally against sample resumes; successfully detected and penalized low-effort template clones while accurately scoring legitimate open-source contributions.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.