Burst threshold tuning (data-driven) #12

Goal

Validate or tune the v0.1 placeholder burst-detection threshold (5 min / 50k nutrition / 2 min cooldown) using real usage data.

Why

The threshold is a guess. The whole proactive-speech feature feels good or annoying based on this number. We should look at real distribution before locking it.

Acceptance

• Add a debug logging mode (env var or hidden menu) that logs every burst evaluation: sum, threshold result.
• Collect ≥1 week of real-use data from at least 2 users.
• Adjust WINDOW_MS / THRESHOLD / COOLDOWN_MS in core/feeding/burstDetector.ts based on the data.
• Document the chosen values + rationale in docs/02-pet.md or docs/05-pet-design.md.

Hints

• Don't make this configurable; make it good. Users shouldn't have to tune a knob.
• The threshold is in nutrition units, not raw tokens — remember nutrition = input*1 + output*3 + cacheCreate*1.5 + cacheRead*0.1.