Labeling transcript text rows: neighbor/turn inheritance is ~50%, read them

When classifying agent-session rows (reasoning, assistant text) by purpose, inheriting the label from the nearest tool call or from the turn's dominant activity was only ~50-60% accurate on spot checks, because one user turn often mixes code edits, tests and cluster verification, and turn-ending summaries sit next to trivial read/git calls. Reading each text row's first ~260 chars and labeling manually took 5 batches for ~540 rows and resolved it. Store manual labels keyed by row number plus event_id (event_id alone is not unique in agentrecap CSVs) and assert the event_id still matches, so regenerating the source CSV fails loudly instead of mislabeling.