Move expensive paths to native Gemini SDK and add context caching

[PraneelBhatia]

Problem

This repo is Gemini-first in product direction but still uses the OpenAI-compatible Gemini endpoint for all calls.

That is convenient, but it leaves cost and ergonomics on the table.

Proposal

Move the expensive Gemini paths to the native Gemini SDK and add caching where it matters.

Priority targets:

• final report generation
• screenshot annotation
• repeated prompt blocks / large schemas

Potential wins:

• context caching
• cleaner multimodal handling
• more direct Gemini-native optimization paths

Suggested file targets

• backend/llm_client.py
• backend/report.py
• backend/annotator.py

Acceptance criteria

• expensive Gemini paths use the native Gemini SDK
• repeated large prompt blocks are cached where possible
• report and annotation calls are easier to reason about cost-wise