feat(otasim): real-HF noise-bed channel model (task #71)

[secup]

Summary

Adds `RealHfLoopChannelModel` — uses a recorded HF WAV (the 20m noise
bed committed in #34) as the noise source, with the same SNR scaling
formula as AWGN. `otasim_ctl set-channel --model real_hf_loop --snr 12`
gives per-sample noise RMS identical to AWGN at SNR=12 in level,
but with real-HF spectral character (atmospherics, NCDXF beacon
rotation, band noise) preserved.

Default unchanged: lobby still starts on AWGN. Real-HF noise is
opt-in via the server flag + admin SetChannel switch. Tests untouched.

How to use

```bash
./build/ota_simulator serve \
--bind 127.0.0.1:50051 --udp-bind 127.0.0.1:50052 \
--tokens /tmp/ota_tokens.conf \
--noise-bed-wav recordings/ota_noise_bed_2026-05-18_20m_14100/noise_bed.wav

./build/otasim_ctl --token admin_tok set-channel --model real_hf_loop --snr 12
./build/otasim_ctl --token alpha_tok get-channel

→ session=lobby model=real_hf_loop snr_db=12.00

```

If `--noise-bed-wav` isn't provided and someone tries to set the model,
the server returns INVALID_ARGUMENT with a clear message.

Implementation

• `src/ota_channel_core/models.{cpp,hpp}` — new `ChannelType::REAL_HF_LOOP`,
  WAV loader (16-bit PCM mono 48 kHz, normalized to unit RMS),
  `RealHfLoopChannelModel` class mirroring AWGN structure.
• `src/ota_channel_core/session_*` + service layer — carry the cached
  noise vector into per-session model construction.
• `tools/ota_simulator.cpp` — `serve --noise-bed-wav ` flag,
  startup load.
• `tools/otasim_ctl.cpp` — usage string lists real_hf_loop (no other
  changes; --model already passes through as a free string).
• `tests/test_real_hf_loop_channel.cpp` — scaling, position wrap,
  setSNR, empty-loop passthrough, parse/name round-trip.
• `tests/test_grpc_service_smoke.cpp` — asserts real_hf_loop is
  rejected when --noise-bed-wav was not provided.

Test plan

• `cmake --build build -j4` clean
• New `RealHfLoopChannel` unit test → PASSED
• `OTASimulatorFileTransferSNR15` (300s slow test) → PASSED on
  unrestricted Mac (Codex sandbox timed out, which is environment
  not regression)
• Live two-step smoke: server with --noise-bed-wav accepts the asset;
  set-channel real_hf_loop SNR=12 switches lobby; get-channel
  reflects; switch back to awgn works.
• CI matrix (Linux + macOS + Windows)

Worked example at SNR=12

```
target noise RMS = kModemReferenceRms × 10^(-12/20)
= 0.3180724 × 0.25118864
= 0.07989617
```

AWGN at SNR=12 produces noise stddev = 0.07989617. The real_hf_loop
model produces noise RMS = 0.07989617 (because the loop is
pre-normalized to unit RMS, and the scaling factor is identical).
Per-sample level matches; spectral character differs.

Follow-ups (queued)

• Auto-discover the asset and switch default to real_hf_loop
  (separate PR; will need test fixups, intentionally not in this one).
• Multi-bed support (different bands, day vs night, contest vs quiet).
• Combined AWGN + real_hf_loop overlay (model thermal noise PLUS QRM).

🤖 Generated with Claude Code