docs: add ADR-051 sensing server decomposition plan (Sprint 2)

14-module extraction plan for 3,765-line main.rs god object.
6 phases, each independently committable and testable.
Target: no file over 500 lines, AppStateInner split into domain sub-states.

Refs #174

Co-Authored-By: claude-flow <ruv@ruv.net>

docs/adr/ADR-051-sensing-server-decomposition.md

@@ -0,0 +1,109 @@
+# ADR-051: Sensing Server Decomposition — main.rs God Object Breakup
+
+| Field | Value |
+|-------|-------|
+| Status | Proposed |
+| Date | 2026-03-06 |
+| Deciders | ruv |
+| Depends on | ADR-050 (Quality Engineering — Sprint 2) |
+| Issue | [#174](https://github.com/ruvnet/RuView/issues/174) |
+
+## Context
+
+`sensing-server/src/main.rs` is 3,765 lines with cyclomatic complexity ~65. It contains 12 structs, 60+ functions, 10 constants, and a 37-field `AppStateInner` god object. This violates the project's 500-line file limit (CLAUDE.md) and makes unit testing individual components impossible.
+
+The file mixes concerns:
+- CLI argument parsing and server bootstrap
+- HTTP route handlers (health, models, recordings, training, pose, vitals)
+- WebSocket upgrade and client management
+- UDP CSI frame receiver and parser
+- Signal processing pipeline (feature extraction, classification, smoothing)
+- Simulated data generator
+- Windows WiFi scanning integration
+- Pose estimation from WiFi signals
+- Vital sign smoothing and filtering
+- Model/recording file management
+
+## Decision
+
+Decompose `main.rs` into 14 focused modules. Each module owns its types, constants, and functions. `main.rs` retains only CLI parsing, state initialization, router construction, and server startup (~250 lines).
+
+### Module Extraction Plan
+
+| Module | Source Lines | Contents | Target Size |
+|--------|-------------|----------|-------------|
+| `cli.rs` | 59-152 | `Args` struct, CLI parsing | ~100 |
+| `state.rs` | 154-370 | `AppStateInner`, all DTOs (`Esp32Frame`, `SensingUpdate`, `NodeInfo`, etc.), `SharedState` type alias | ~220 |
+| `signal.rs` | 542-890 | `generate_signal_field()`, `estimate_breathing_rate_hz()`, `compute_subcarrier_variances()`, `extract_features_from_frame()`, `raw_classify()` | ~350 |
+| `smoothing.rs` | 886-1060 | Classification smoothing, vital sign smoothing, `trimmed_mean()`, constants | ~180 |
+| `routes_health.rs` | 1660-2005 | `/health/*`, `/api/v1/info` endpoints | ~350 |
+| `routes_model.rs` | 2058-2230 | `/api/v1/models/*`, LoRA profiles, `scan_model_files()` | ~180 |
+| `routes_recording.rs` | 2233-2440 | `/api/v1/recording/*`, `scan_recording_files()` | ~210 |
+| `routes_training.rs` | 2443-2560 | `/api/v1/train/*`, `/api/v1/adaptive/*` | ~120 |
+| `routes_sensing.rs` | 2562-2710 | Vital signs, edge vitals, WASM events, model info, SONA endpoints | ~150 |
+| `routes_pose.rs` | 1701-1930, 2007-2055 | Pose estimation, `derive_single_person_pose()`, pose/stats/zones endpoints | ~280 |
+| `websocket.rs` | 1492-1660 | WS upgrade handlers, `handle_ws_client()`, `handle_ws_pose_client()` | ~170 |
+| `udp_receiver.rs` | 2725-2890 | UDP CSI frame receiver task, frame parsing | ~170 |
+| `data_sources.rs` | 1063-1465, 2888-3020 | Windows WiFi task, simulated data task, `probe_windows_wifi()`, `parse_netsh_interfaces_output()` | ~400 |
+| `router.rs` | (new) | `build_router()` function assembling all routes | ~80 |
+
+### Extraction Order (6 Phases)
+
+1. **Phase 1**: `cli.rs` + `state.rs` — Zero behavioral change, just move types
+2. **Phase 2**: `signal.rs` + `smoothing.rs` — Pure functions, easy to test
+3. **Phase 3**: `routes_health.rs` + `routes_model.rs` + `routes_recording.rs` — Stateless-ish handlers
+4. **Phase 4**: `routes_training.rs` + `routes_sensing.rs` + `routes_pose.rs` — Remaining HTTP handlers
+5. **Phase 5**: `websocket.rs` + `udp_receiver.rs` + `data_sources.rs` — Async tasks
+6. **Phase 6**: `router.rs` — Assemble all routes, slim `main.rs` to ~250 lines
+
+### State Refactoring
+
+`AppStateInner` (37 fields) will be split into domain-specific sub-states:
+
+```rust
+pub struct AppStateInner {
+    pub config: ServerConfig,        // CLI args, ports, paths
+    pub sensing: SensingState,       // CSI frames, features, classification
+    pub vitals: VitalsState,         // Vital sign buffers, smoothing state
+    pub models: ModelState,          // Active model, discovered models, LoRA
+    pub recording: RecordingState,   // Active recording, file handles
+    pub training: TrainingState,     // Training status, adaptive model
+    pub pose: PoseState,             // Person detections, pose history
+    pub broadcast_tx: broadcast::Sender<SensingUpdate>,
+}
+```
+
+## Consequences
+
+### Positive
+
+- Each module is independently unit-testable
+- No file exceeds 500 lines
+- Domain boundaries are explicit (state sub-structs)
+- New developers can find code by domain
+- Merge conflict surface reduced (parallel module edits)
+
+### Negative
+
+- Large refactor with ~3,700 lines touched — high merge conflict risk
+- `pub(crate)` visibility needed for cross-module state access
+- Some functions share mutable state, requiring careful `&mut` threading
+
+### Neutral
+
+- No behavioral change — all endpoints, WebSocket, UDP behavior stays identical
+- Existing integration tests (if any) continue to pass unchanged
+
+## Implementation Notes
+
+1. Each phase is a separate commit for easy revert
+2. Run `cargo test` and `cargo check` after each phase
+3. Use `pub(crate)` for internal types, keep public API surface minimal
+4. Add `#[cfg(test)] mod tests` to each new module with at least smoke tests
+5. Consider adding `tower` middleware for auth (Sprint 1 remaining item) during Phase 3
+
+## References
+
+- ADR-050: Quality Engineering Response (Sprint 2 plan)
+- Issue #170: Quality Engineering Analysis
+- CLAUDE.md: 500-line file limit rule