Bug: project highlights are extracted by the LLM but discarded in transform_projects()

[VanshGupta18]

Description

transform_projects() in transform.py discards the highlights returned by the LLM. Instead of reading item["highlights"], it reads a non-existent "type" key, so the expression always evaluates to an empty list:

# transform.py:339
"highlights": [item.get("type", "")] if item.get("type") else [],

"type" is not a field on the Project model in models.py, and it is never requested in prompts/templates/projects.jinja, so no resume can populate it.

The sibling transform_work() handles the same field correctly at transform.py:218:

"highlights": item.get("highlights", []),

Impact

Every bullet point under every project on a resume is silently dropped before evaluation.

The rest of the plumbing is already complete — Project.highlights exists in models.py, and convert_json_resume_to_text() already renders project.highlights at transform.py:812. This single assignment is the only break in the chain.

As a result the evaluator receives only project names and URLs, losing all project descriptions, metrics, impact statements, and any demo/deployment links written in bullet text. This under-scores the self_projects category (weight 30/100).

Evidence

The raw LLM response already contains the highlights, confirming this is purely a transform-layer loss (model gemini-2.5-flash, structured output):

{"name": "Multimodal Earnings Call Intelligence System",
 "highlights": [
   "Built a multimodal sentiment analysis pipeline for 1,038 earnings calls, fusing FinBERT text embeddings with prosodic audio features using a turn-level gating and attention pooling model.",
   "Improved macro-F1 from 0.699 to 0.741 (+5.1 points) over a text-only baseline through cross-modal fusion, validated using cross-validation and significance testing."
 ],
 "url": "https://github.com/.../earning-calls"}

But the text actually passed to the evaluator contains none of it:

=== PROJECTS ===
1. Multimodal Earnings Call Intelligence System
   URL: https://github.com/.../earning-calls
2. Corporate Compliance Environment
   URL: https://github.com/.../corporate-compliance-env
3. AI-Powered Geopolitical Risk Intelligence System
   URL: https://github.com/.../forsyt

Steps to Reproduce

1. Run python score.py <resume.pdf> --role software_engineering_intern on a resume whose projects have bullet points.
2. Inspect cache/resumecache_<name>.json — every entry under projects has "highlights": [].
3. Equivalently, print convert_json_resume_to_text(resume_data) — the === PROJECTS === section contains only names and URLs.

Environment Info

• OS: macOS (Darwin 25.5.0)
• Python: 3.11
• Model: gemini-2.5-flash
• Commit: 70fd3ea

[VanshGupta18]

Working on this — will open a PR shortly.