Build a modular Unity enemy AI with a behavior tree or state machine

You are a senior Unity C# AI programmer who builds composable, debuggable enemy behavior.

Enemy brief: [DESCRIBE THE ENEMY AND WHAT IT MUST DO — e.g. 'a melee guard that patrols waypoints, chases the player on sight, attacks in range, and flees when below 30% HP'].
Architecture preference: [BEHAVIOR TREE / FINITE STATE MACHINE — pick one and say why it fits the brief].

Constraints:
- Unity version: [2022 LTS / Unity 6].
- Use the pattern end to end: if a behavior tree, real Selector and Sequence nodes with Success/Failure/Running returns; if an FSM, explicit state classes and transitions, not an enum switch inside Update().
- Decouple sensing from action: a Sensor component (sight range, FOV, hearing) feeds data to the brain; the brain never reaches into the player transform directly.
- Expose every tunable (sight range, FOV, attack range, cooldowns, flee threshold, decision tick rate) as [SerializeField] private fields with [Header] groups and sensible defaults.
- Use a decision tick (e.g. re-evaluate every 0.1 to 0.2 seconds), not every frame, and cache Transform references.
- Include a clear extension point so I can drop in a new behavior node or state without touching the core.

Output:
1. A 3-bullet design rationale: why this pattern, where the seams are for extension.
2. Each class as its own fenced code block, file name as a comment on line 1.
3. A wiring note: which GameObjects or components hold the Sensor, the Brain, and the Motor, and how they communicate.
4. A tuning-first checklist: the first values to tweak to change how the enemy feels.

State every architecture assumption up front. If the brief is ambiguous, choose a sensible default and flag it.

Success signal: the output is good only if sensing is decoupled from the brain, the chosen pattern is used end to end (not a switch-statement shortcut), and every tunable is serialized with a default.