Agent-drafted interpretation, pending human review

The repo-side quantitative analysis is refreshed at path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-continuous-v1/, and the collection is now complete (36/36 analyzable, 0/36 running). Interpretive writeup is still deferred here pending human review.

Candidate interpretation topics to revisit after review:

1. whether the beta0_const_pv0 regime should be framed as the highest-decoding corner of the sweep rather than as a straightforward win, given its near-zero reconstruction and very large end-of-training KL;
2. whether the positive constant-beta TWR averages are mostly driven by list_summation and grid_pathfinding, with weaker or negative router-probe deltas on sorting_algorithms, base_conversion, and multidigit_addition;
3. whether the persistently low sorting_algorithms unique-strategy rates should be treated as a task-specific bottleneck that remains after both beta-schedule and TWR changes.