Direct review should suppress findings already raised in PR feedback #76

Problem

The production Direct imreview workflow reviews the frozen diff without reading the PR's existing discussion, review summaries, or review threads. As a result, it can return findings that maintainers or reviewers have already raised.

The full/Strict PR workflow already builds a PR context bundle and tells the reviewer not to repeat existing concerns, but this behavior was not carried into Direct mode. Evaluation runs intentionally support PR_CONTEXT_MODE=no_discussion to prevent ground-truth leakage; that evaluation-only boundary should not define production behavior.

Proposed behavior

Keep Direct independent during source analysis, then perform a final duplicate-classification pass against existing PR feedback:

• new: no equivalent existing concern; include normally.
• duplicate: equivalent concern already exists; suppress a new comment.
• extends-existing: materially stronger evidence or impact; point to or reply to the existing thread rather than opening a duplicate.
• resolved/outdated: do not revive unless the pinned head still contains the defect.

The Direct MCP can remain thin and network-free. The host imreview workflow should fetch the bounded PR discussion/review context and retain it in the shared evidence packet. Evaluation must retain an explicit no-discussion mode.

Acceptance criteria

• Direct PR reviews inspect bounded existing conversation comments, review summaries, and thread-aware inline comments before final output.
• Candidate source findings are generated independently before duplicate classification.
• Final output suppresses duplicate findings and distinguishes extensions from genuinely new findings.
• Resolved and outdated threads are represented in the classification.
• Local/worktree reviews continue to work without GitHub context.
• Evaluation can explicitly disable discussion context.
• Tests cover duplicate, extension, resolved/outdated, unavailable-context, and local-review behavior.

Motivation

PR vllm-project/vllm-omni#5666 exposed the gap: Direct review independently found two valid issues, but both had already been raised in the PR discussion. The findings were technically correct but would have produced redundant review noise if posted.