Turn-LoRA class balance must account for target-token concentration

Observed 2026-09-04 on e3-solutions/ender-inference, COR-3791, branch vansh/cor-3791-fine-tune-a-freight-turn-detector-for-identifiers-pauses-and, commit bc768b0.

A 1:1 paired synthetic augmentation is not necessarily a balanced next-token objective for turn detection. In this experiment, finished rows all reinforced the same EOU token while unfinished rows distributed supervision across many continuation tokens. The resulting corpus still exerted concentrated finish pressure: a gentle head-only candidate scored 30/42 in production-service replay; exact 3× weighting of 208 partial-number rows improved only to 31/42, below the selected adapter’s 34/42 and 3-early-cutoff safety comparator. A 25-epoch extension also failed. Reusable method: select in the independently loaded service runtime, gate first on false-finished/early-cutoff non-regression, inspect target-token concentration rather than row counts alone, and move to an isolated capacity/objective test instead of adding generic acknowledgement rows.