Unity gameplay systems (gameplay systems programmer)

Expert gameplay work in Unity 6.3 means every system has one owner of its state, runs at a rate someone chose, and is proven by a measurement: steps per frame, a tunneling test at the real speed, a path query through the real door, a timed transition, the worst frame at the real agent count. This skill carries the expert delta from the 6.3 manuals, AI Navigation 2, NGO and shipped games (Survival Kids, Very Very Valet), with jobs and tests run in 6000.3.21f1. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-unity-expert (channels, review loop, 6.3 traps). Install: ut_gameplay.install(P) (core AgentKit + AgentKit.Gameplay jobs + the AgentKit.Gameplay.Runtime assembly).

Stance (the expert delta)

1. Fix frequency before colliders. Physics cost multiplies with fixed steps per frame; "a call count close to 10 might indicate an issue" (6.3 Manual). Measured at 50 Hz: a 400 ms frame ran 17 catch-up steps with the default 0.333 s Maximum Allowed Timestep; a 0.1 s cap gave 5 steps and a world at 0.25x. The cap trades the snowball for slow motion.
2. Escalate CCD per object on observed tunneling, then leave rigid bodies. Discrete, Continuous Speculative ("often the best choice"), Continuous, Continuous Dynamic last (6.3 Manual). Measured: every CCD mode held a static 5 cm wall to 300 m/s; against a thin dynamic plate Continuous tunneled from 50 m/s, Continuous Dynamic from 100. A sweep of the projectile's own shape (SweepProjectile) stopped the 300 m/s shot, and a 0.25 m sphere sweep hit a post a center ray missed: match the cast to the collider (LlamAcademy, fJyi7l2tWKo [00:09:54]).
3. One writer per piece of state. Agent plus non-kinematic Rigidbody is a race, agent plus enabled obstacle avoids itself, agent and animation must flow one way (AI Navigation manual). Measured: a main-thread write to a transform held by a running TransformAccessArray job waited the whole job (22 to 38 ms, vs 0.003 ms outside it) (Survival Kids, ZkvK0mX-id4 [00:26:42]).
4. A NavMesh is proven by queries and saved as an asset. One NavMesh Surface per agent type (SMWxCpLvrcc [00:06:35]); voxel 3 per agent radius, 4 to 6 indoors (manual). Measured: a door 1.2x the agent diameter was closed at 3 voxels per radius, open at 6; stairs without a HeightMesh read 0.17 to 0.26 m off the treads, 0 with it; a carve reached queries one frame later.
5. The lightest AI structure, ticked from one place. Unstructured for 1 to 2 behaviors, FSM while transitions stay few (toward N squared), behavior tree for priorities, GOAP for chained goals (LlamAcademy, CZvfuNfdc1M [00:07:13]). Brains tick from a director; mixed object types register with one UpdateHub sorted by type (Survival Kids: same code, mean -12%, worst frame 2.18 to 1.43 ms [00:23:59]; measured here at 1,000 and 10,000 objects: mean 7 to 14% lower).
6. DOTS is an umbrella; measure before adopting it. Order updates, then Burst jobs over NativeArrays with transforms copied out in a tiny read-only job, then fill main-thread job waits with Burst .Run() (measured: 6 to 10 ms saved of 21 to 26), and only then Entities (Survival Kids [00:08:49], [00:39:15]). Decide ECS early, for large CPU-driven counts, competitive netcode or determinism (Turbo Makes Games, Bz24Jp30nkM [00:21:58]).
7. NGO 2.x has no full client-side prediction and reconciliation (no rollback-and-replay loop) and no server rewind. Client anticipation (AnticipatedNetworkVariable, AnticipatedNetworkTransform, OnReanticipate) is the building block for hand-built prediction (NGO 2.7 manual; same text in the bundled 2.13.0 docs). Server authority by default, HasAuthority over IsServer, "would a late joiner need it?" decides NetworkVariable vs RPC.
8. Every sound goes through a mixer group; distance happens at the source. Output None (the default) bypasses the mixer; attenuation and distance low-pass happen at the AudioSource before any group (manual). Measured: Logarithmic rolloff kept attenuating past Max Distance (1/d to 40 m), a slider on a snapshot parameter froze that group's moods until ClearFloat, and 60 uncapped looping weapon sounds erased a quiet music bed at default priority. Cap frequent sounds at 10 to 15 and steal the oldest (git-amend, BgpqoRFCNOs [00:17:56]).

Establish first