Use paired controls to evaluate LoRA warm-start additions

To test whether LoRA adapter-weight warm-start preserves an incumbent better than fresh initialization, do not compare a single warm-start candidate only with the incumbent. Start from two hash-identical incumbent adapter copies with fresh optimizers: continue one on the original corpus as a rehearsal-only control, and continue the other under identical settings with the proposed additions. Their paired difference isolates the additions from generic continued-training drift; the untouched incumbent remains the zero-update anchor. If initialization is the question, add a fresh-init arm with the same corpus, learning rate, update budget, scheduler, seed, and runtime.

Call this adapter-weight warm-start, not checkpoint resume, unless optimizer and scheduler state are restored. Predeclare whether fixed updates or fixed epochs define the estimand, because corpus-size changes make them different. One seeded pair remains exposed to numerical nondeterminism and needs a clear paired margin. Repeated synthetic gains with real-development regressions should trigger better representative training coverage before more tuning.