[KunalSiyag]

What

Every Skeleton3D defaults to modifier_callback_mode_process = PHYSICS (Godot 4.7), so the game's GDScript SkeletonModifier3Ds — DogRig/CatRig on every animal body, the hero's strike pose and head look, the T0 conversation head look — ran once per physics frame at 60 tps. The fp layer animates entirely on the idle frame (all mixers are ANIMATION_CALLBACK_MODE_PROCESS_MANUAL, advanced by hand), and the physics server is parked on the web (#89) — but parked physics frames still tick, so the modifiers kept paying per-tick, and were paid again for every catch-up step a slow frame absorbs (up to 2 on web after #84). The day-10 tick-frame attribution puts ~11 ms of the ~45 ms street_tick p50 in exactly this line (round4/web-perf-budget v3.1).

The fix is one line per attach site: modifier_callback_mode_process = IDLE — modifiers run on the idle frame with the hand-advanced mixers, exactly once per rendered frame, immune to catch-up multiplication. Modifiers use real deltas (dog wag pace, LookAt duration), so behavior is unchanged beyond sub-frame timing.

Attach sites

• traffic/people_pool.gd — every people body at build (the T0 LookAtModifier3D head look)
• animals/dog_bodies.gd — DogRig
• animals/cat_bodies.gd — CatRig
• player/hero_strike_pose.gd — StrikePose
• player/hero_look.gd — hero head/neck LookAtModifier3Ds

Tests

• New fp_kit.check_modifier_modes(fails, root): sweeps the scene for any SkeletonModifier3D and asserts its skeleton is on the IDLE mode — a future physics-callback modifier fails the suite. Wired into fp_animals (12 modifiers swept), fp_cats (16), fp_player (10).
• Targeted pins: DogRig (fp_animals), CatRig (fp_cats), a seated people body (fp_people_prewarm — its skeleton carries the mode even though seated drivers hold no look modifier), the hero skeleton (fp_player).

FP_ANIMALS_TEST, FP_CATS_TEST, FP_PEOPLE_PREWARM_TEST, FP_PLAYER_TEST all PASS; tools/gate.sh quick and pre-commit green.

Perf

Native bracket measurement (fixed-fps 60, parked server): 48 modifier-bearing skeletons cost 0.275 ms/physics-frame vs 0.172 ms bare — the modifier share is real and scales with characters on screen; on wasm (GDScript callbacks ~5-10× slower) at day-10 counts this is the ~5-6 ms/physics-step line the attribution measured, double-paid on catch-up frames.

[github-actions]

Automated review

No new findings on the 10 changed files (gdlint, complexity, TODO policy, dead code).

Run python3 tools/checks/run_all.py locally for the full set of checks.