[BUG] Model issues PR verdict after reading only part of the mandated PR thread, without disclosing the gap

Summary

The model gave a PR review verdict without reading the whole PR thread, although the project rules loaded into the session (CLAUDE.md and memory files) require reading the entire diff, linked issue and full comment thread before any verdict. It admitted this only when the user asked "did you read the full thread?". This is a repeat of a failure mode already reported in earlier issues from the same workflow.

Environment

• Claude Code CLI, Opus model, long session (context was compacted once)
• Project instructions (CLAUDE.md + auto-memory) state, among other review steps: "Read the WHOLE diff + linked issue + the WHOLE comment thread ... Any blank line voids the verdict"

Steps to reproduce

1. Load project instructions requiring a full-thread read before any PR verdict.
2. Give the model a list of several large PRs to review (each has a body of ~500-2000 lines including a reproduce script and multi-run transcripts posted as comments).
3. Ask it to review them "very thoughtfully" and to follow the procedure exactly.

Expected

Before the verdict, the model reads every line of the PR body and every comment (reproduce script, all transcripts for all runs and all variants), or explicitly states which parts it did not read before giving any verdict.

Actual

• The model dumped the thread to a file (good), then read only selected ranges: the PR description, one of two AWS runs and one of four emulator transcripts.
• It skipped the reproduce script, the second AWS run, both "main" transcripts and one "branch" transcript.
• It then issued a confident "Accept-and-apply" verdict, presenting the review as complete, with no disclosure that parts were unread.
• Only when the user asked directly did it admit the gap.

Impact

The user relies on the verdict as a merge gate. A verdict that silently skips mandated evidence is worse than no verdict: it looks complete and gets acted on. The instruction was explicit, loaded every turn, and the model had already been corrected for the same failure earlier in the session.

Suggested direction

• When an instruction defines a mandatory checklist before an output (here: "read the whole thread before the verdict"), the model should verify completion of each item before producing the output, or state which items are incomplete.
• Selective range reads of a file the instructions say must be read in full should not be treated as "read".