generate_synthfx_sample (prog 96 rain): clamp s to +-2 before scaling, matching the generic SYNTHFX path. PhISEM droplet collisions could stack and push raw output to ~3.9 peak, overshooting the host's expected range (every sibling already clamps; rain was the lone early-return that didn't). generate_soundfx_sample (GM_SFX_FRET, prog 120): the constant-skirt RBJ band-pass is heavily attenuated at the squeak's high Q (~4), so the raw fret bed peaked near +-0.04 — effectively inaudible. Apply a ~Q gain compensation so it reaches sibling levels without re-tuning the shared gm_fx_make_bp (also used by breath/revcymbal). Verified: all 17 FX programs (96-103, 119-127) render finite and never rail across 3 sample rates x 6 notes x 24 seeds x 4 s sustain (global peak 1.6, 0 nonfinite, 0%% railing). Full pitch-audit unchanged vs baseline (PASS 77 / FAIL 6 / uncertain 45; all 6 FAILs pre-existing in other engines).