fix(sensing): reduce empty-room adaptive false positives

[davestanyer]

Summary

• Expand adaptive training samples from RuView multi-node JSONL into per-node sequences so runtime per-node classification matches training data shape.
• Preserve temporal window boundaries by grouping training windows by recording/node sequence key instead of concatenating unrelated nodes or captures.
• Add an absence consensus gate: when adaptive motion says present but independent multistatic/person-count fusion remains zero, clamp the fused output and active node API view to absent / person_count: 0.

Why

In the current empty-room acceptance case, all six live ESP32 CSI nodes could report present_moving while the independent person-count path reported zero people. That creates an empty-room false positive. This PR treats that contradiction as absence unless the independent person-count/fusion path agrees there is a person.

Related: #518, #521

Local validation

• cargo fmt --package wifi-densepose-sensing-server
• cargo test -p wifi-densepose-sensing-server --bin sensing-server adaptive_classifier::tests -- --nocapture — 4 passed
• cargo test -p wifi-densepose-sensing-server --bin sensing-server -- --nocapture — 74 passed
• cargo test -p wifi-densepose-sensing-server — passed
• docker compose -f docker/docker-compose.yml up -d --build sensing-server — rebuilt/restarted
• Live empty-room check after rebuild: final /api/v1/sensing/latest reported motion_level: absent, presence: false, estimated_persons: null; final /api/v1/nodes sample showed all six active nodes as absent with person_count: 0.

Notes

This is intentionally focused on sensing-server adaptive/runtime behavior. It does not include local recordings, generated model files, firmware binaries, or unrelated local worktree changes.