[TheMeinerLP]

Summary

• WindowConfig defaults: window_seconds 10 -> 20, stride_seconds 5 -> 10 (half-window overlap kept), max_packets 1024 -> 2048.
• Raw build and evaluate/train job parameters take their defaults from WindowConfig instead of repeating literals.
• max_packets = 2048 is added to the tuning search space so the new default is searchable.
• Docs explain the defaults with the pilot measurement 2026-10-16 (single player, 2 h 10 min, 269,566 packets): packets per 20 s window p50 524, p90 1,386, p99 1,969, max 8,180 (serverbound p99 888, clientbound p99 1,084); 10 s windows had p99 1,084, so 1024 already truncated.

Compatibility

Specs are stored and hashed with the model and written with every window field, so existing models keep their 10 s / 5 s / 1024 window; only new builds change. Tests cover loading and re-saving an old spec unchanged.

Not changed

• Aggregate feature window (FeatureSpec.window = 256 events) is event-count based, not time based.
• Explicit windows in eval configs and tests (small and fast on purpose).
• Latency budget lines (handled in a separate PR). The latency estimate already extrapolates the last measured slope to 2048 packets; a test pins it as finite and monotonic.

Tests

ml: new test_raw_window_defaults.py, raw jobs/windows/streaming/eval suites; services: scoring suite. ruff clean for both.