[P1] Preserve RF source/link identity and calibrated temporal cohorts for room learning #2419

Parent plan: #2418

Scope

Implement the source-preserving input/cohort boundary for the final room learner. Reuse canonical RfFrameV2/RfTensor concepts and existing adapters, with a bounded application envelope for verified owner/installation/room/source registry and capture role.

Preserve per-source/link identity, exact raw grid/channel, representation/phase validity, source and receiver clock provenance/uncertainty, calibrated geometry and baseline revision. Magnitude-only CSI, RSSI-only Wi-Fi and validated complex RF are different capabilities. Do not fabricate phase or convert missing data into an empty observation.

Acceptance

• Bounded per-source temporal windows and deterministic cohort formation; no global pooling or source overwrite.
• Per-source grid/sequence/clock/age rules, source-set completeness and missing/optional masks are tested at the public boundary.
• Physical/claimed/synthetic provenance and receiver time versus measurement time remain explicit.
• Normalization preserves baseline-relative gain/mean shift and fits only the designated training empty references.
• Frozen trial/session train/validation/held-out identity with no shared frame IDs or overlap across splits.
• Wrong owner/room/revision, replay, drift, source changes and excessive timing uncertainty fail with bounded actionable reasons.
• Composed adapter/cohort tests and causal guard mutants pass; no hardware or raw-data retention claim.