[Model behavior] PR thread content read through filters that drop parts of it, after an explicit instruction to read it whole

What happened
When reading PR bodies and comments, the model used sed to delete collapsible <details> sections, read large comments in byte-range chunks (head -c / tail -c), and filtered lines with grep -v. The user had explicitly told it to read threads whole and unfiltered.

Expected
Read every comment in full, unfiltered.

Impact
Content that matters for the review (reproduce scripts, transcript lines) is silently skipped, and verdicts rest on partial reads.