The extended-KS pluck voices could blow up to inf/NaN at high sustained notes (clamped for safety, but pinned). Two in-loop energy sources fixed: - pluck-position comb was applied per-sample INSIDE the feedback loop, edging loop gain >1. Moved it to the EXCITATION (baked into the string buffer at init) on both buffer paths, matching gm_ks_big_init — the loop is now pure delay + loss filter, unconditionally stable. (Jaffe-Smith.) - jawari buzzing-bridge nonlinearity (sitar/shamisen) was fed back into the string, pumping the loop. Now applied to the OUTPUT only; clean signal feeds back. Sitar stable. Local stress test (60k-sample high-note renders): all plucks finite, clampHits→0 (was inf/NaN). Pluck-position notch + jawari character preserved.