[Model behavior] PR re-review reduced to a head-commit check; new test file and current-main behavior never read before the verdict

Summary

Asked to review a pull request again, the model checked only that the head commit had not changed and restated its earlier verdict. It had never read the PR's new 322-line test file, and it did not re-pull the diff, the discussion thread, or mergeability against the current main branch. When the user asked "did you actually review it?", the model did that work for the first time.

In the same session the model:

• called a change "breaking" without checking that the main branch already failed in the same case;
• accepted a PR touching background threads and wake events without reading the project's concurrency module or its reapers.

The loaded procedure requires both checks before any verdict.

Environment

• Claude Code CLI, Opus-class model, 1M context
• Code-review workflow with a written procedure loaded every session: read the whole diff, grep callers, check current main, then give one verdict

Expected

A re-review repeats the full procedure. Any claim of "breaking" or "regression" is checked against the main branch's current behaviour before it is made.

Impact

Verdicts that look complete are not, so the user has to re-derive every one.