PR-aware reviews

### Problem Statement

OCR reviews do not take into consideration previous discussions or feedback in a PR. They only look at code differences. So if there is

* another AI or other people leaving feedback
* discussions and responses with important context about intent and trade-offs decided
* previous posts and comments from OCR

Then all of that is ignored by OCR. The resulting review feel disconnected, particularly compared with, say, what CodeRabbit does.

### Proposed Solution

When I invoke OCR reviews from a GitHub Action, I want to specify a file full of "prior context" that will be first summarized using a first inference request to the configured LLM. Then I want the summary provided as context in the second, actual review request to the configured LLM. The response should not include duplicative findings, only new ones.

### Alternatives Considered

_No response_

### Affected Area

Review Agent / LLM interaction

### Additional Context

_No response_

Start with the Review Agent / LLM interaction and the GitHub Action review invocation. Trace how the configured LLM receives review input, then identify where prior-context summarization and the actual review request are assembled. Done means prior context is summarized, supplied to the review, and duplicate findings are excluded.

Written by the indexing model from the issue text.