[P1 Epic] Integrate the final multi-source room learner and guided calibration #2418

End-user outcome

Make room calibration produce a genuinely trained, account-owned multi-source learner that uses all selected RF sources, can learn from aligned optional references, and keeps RF inference working when the phone leaves. First delivered task: room-selective binary presence, including occupied-static detection and adjacent-area rejection. Pose/count/Vitals need their own qualification.

Owner: Dragan; implementation agent: Codex. Highest current RuView integration priority. This tracks source, CI, local/device testing and physical acceptance separately; no merge/release or hardware-write authority is implied.

Existing building blocks to integrate

• Canonical RF records/adapters: ruview-unified RfFrameV2 / RfTensor.
• Genuine CSI encoder and contrastive objective/update: AETHER; guided SpecialistBank/MultiNodeMixture as a comparative statistical baseline.
• Geometry/domain conditioning: MERIDIAN primitives where supported by measured geometry.
• Reference alignment/validation: ruview-groundtruth; bounded derived phone camera/LiDAR labels and geometry.
• Qualified claim fusion: ruview-fusion plus source-aware coherence/contradiction mechanisms.
• Keep selected-source detector and draft feat(calibration): opt-in per-source FieldModel shadow bank #2417 per-source bank as explicit controls, not the final learner.

Design gaps that must be closed

• Preserve source/link identity, exact grid/channel, missing I/Q/phase masks, measured timing uncertainty and geometry; do not collapse three sources into ten pooled features.
• Do not normalize away baseline-relative amplitude/mean-shift information while seeking hardware invariance.
• Same-event views must not become false negatives. Use real optimized loss, not logged-only loss or heuristic pseudo-targets labelled pretraining.
• Freeze trial/session-disjoint train/validation/held-out roles before fitting; prevent time-of-day, PHY changes, overlapping-window and repeated-holdout leakage. Incorporate reported confounds in Controlled reproduction: raw-CSI presence detection sits at chance (AUC ≈ 0.58) once time-of-day is controlled — the two apparent successes were confounds #1440.
• Correlated same-AP receivers are not independent confidence multipliers. Unknown phase/clock quality cannot authorize coherent fusion; incorporate Subject: Multistatic fusion can't hold a 60ms guard on 4x ESP32-S3 — mesh timestamps carry 0.28–1.5s residual error #1710.
• Unvalidated radar/receive-time-only receipts and camera out-of-view observations are not unquestioned ground truth.
• Preserve verified owner/installation/room/model identity and local records across restart, logout, refresh and candidate rejection.
• Keep optional AI/OAuth, optical capture grants and raw retention separate from local learning authority.

Delivery sequence

1. Source-preserving observation/cohort and calibration contracts.
2. Real multi-view representation learning plus room-presence head and frozen inference.
3. Managed capture/train/evaluate/store/adopt lifecycle and protected API.
4. App guided capture, source diagnostics and paired derived-teacher integration.
5. Independent physical acceptance, linked to cognitum-one/ruview-mobile#105.

Implementation children will be linked below. Source should be reviewable in bounded stacked PRs; never mark this epic done from green CI alone.

Acceptance

• All required selected RF sources have distinct trained/scored provenance; missing/stale/changed sources are explicit.
• Behavioral tests prove the actual objective changes intended weights and improves a disjoint controlled fixture; wrong labels/random controls do not manufacture a pass.
• Real managed capture → training → persisted candidate → cold-loaded inference works through composed production code.
• Optional references have verified clock/spatial/data-role context; phone absence preserves the qualified RF-only path.
• Account A/B isolation, expiry, cancellation, refresh, disk corruption and rollback tests pass without data loss.
• Guided UX exposes progress, rejection reasons and actual model/coverage rather than receipt-only readiness.
• Independent empty/occupied-static/movement/adjacent/recovery/missing-modality trials compare against controls, with frozen thresholds, adequate coverage and uncertainty reported.
• Actual delivered Mac/iPhone artifacts are qualified; no unsupported presence/pose/Vitals claim.

Proposed physical targets to lock before collection: occupied sensitivity >=90%, empty/adjacent false-positive rate <=5%, qualifying coverage >=90%, p95 decision latency <=2s. These are targets, not measured results; trial counts, correlation and confidence intervals require review before validation.

Linked implementation trackers

• [P1] Preserve RF source/link identity and calibrated temporal cohorts for room learning #2419 — source/link observation, normalization and cohort contract (in progress)
• [P1] Integrate genuine multi-view contrastive training and a room-presence head #2420 — genuine multi-view representation + room-presence learning (in progress)
• [P1] Wire the room learner capture, model lifecycle and protected managed-server API #2421 — managed capture/train/evaluate/persist/load/adopt lifecycle
• mobile#121 — guided UX and paired derived teachers
• mobile#105 — independent physical/spatial benefit acceptance

Current controls/source evidence: draft#2417, implementation head6a9aa03b289bda26204b45f010dbad73204135d7, and the selected-source shipped detector. Source audit completed; final-learner implementation, installation and physical qualification are not complete.

Implementation branch: feat/room-learner-integration-20261101, isolated from published PRs. Public issue checklists will be updated with exact PR heads, executed tests and device/trial receipts as work lands.