Reference LoRA preservation weight can erase a marginal gain without reducing base safety errors

Observed 2026-09-07 on ender-inference COR-3791, commit bf9ff74: in a selected-adapter warm-start Reference-LoRA run (rank-8 all-linear, LR 1e-4, one epoch, exact 90-row balanced corrective/rehearsal corpus), raising selected-correct preservation weight from 4 to 8 eliminated final rehearsal regressions (0/40) but also eliminated the prior marginal validation gain. Every checkpoint scored exactly 34/42 through the service wrapper (3 false-finished, 5 missed-finished, three identical repeats), matching the selected adapter. This suggests stronger symmetric preservation is not a path to improving the remaining errors; future experiments need new discriminative evidence or a narrower objective rather than more preservation pressure. Provenance: https://github.com/e3-solutions/ender-inference/commit/bf9ff74; sealed private receipt /workspace/cor3791/private/reference-warmstart-lr1e4-preservation-w8-v1-001/final-selection.safe-receipt.json SHA-256 3b689ed9a37b22d08d7034758ba47da51c4d1dc2baff965e95b016fd23b6d108.