Rssi and noise floor improvements

[yg-ht]

This PR improves RSSI and noise-floor instrumentation. This should then re-enable the use of the interference threshold configuration, but I think this PR needs wider testing before that is done. So at the moment, it doesn't change the int.thresh / _prefs.interference_threshold from 0.

It enables stats-radio over Mesh as I have been diagnosing the RSSI behaviour from a number of sites concurrently. This might not be desirable, for example owing to concerns over bandwidth usage, but I am not sure I see any real harm in this. This particular stats call now has last_rssi and last_snr come from cached last-packet metrics rather than later instantaneous RSSI sampling. It also exposes stats-noise over the Mesh which goes into more detail about the noise floor sampling results. This reports the current approved noise floor, along with details about the current sampling for the next noise floor calibration. Details include: accepted sample count, sample min/median/max values, and rejected sample counts as the sample was erroneously low/high.

Noise floor calibration is performed after a small number of event triggers, as well as on a periodic basis. The rough execution flow is:

flowchart TD
    A[Periodic trigger<br/>or scheduled refresh] -- Start batch --> B0{batch loop}
    B0 -- Start sample --> B1{Batch maximum window size elapsed?}
    B1 -- No --> C{Radio idle in RX mode?}
    B1 -- Yes --> A

    C -- No --> B0
    C -- Yes --> D{Sample rate limit interval elapsed?}

    D -- No --> B0
    D -- Yes --> E[Read RSSI sample]

    E --> F{Sample acceptable?}
    F -- No --> G[Reject sample<br/>Update reject counters]
    G --> B0

    F -- Yes --> H[Accept sample<br/>Update min/median/max]
    H --> I{64 accepted samples?}

    I -- No --> B0
    I -- Yes --> J{Completed batch sane?}

    J -- No --> A
    J -- Yes --> K[Publish new noise floor]
    K --> A

Loading

So that people can tweak and understand the noise floor sampling process, there are now a few additional configuration options: sampling interval, calibration window, and clamp settings:

• noise.sample.ms - in a calibration batch, how long do we sleep between samples (default: 50ms)
• noise.window.secs - maximum time of a calibration batch (default: 60s)
• noise.clamp.low - the "ludicrous" low RSSI value that can't be true (default: -125dBm)
• noise.clamp.high - the "ludicrous" high RSSI value that can't be true (default: -80dBm)

I have been testing this on a few different repeaters in different locations and using different boards. I haven't seen any anomalous behaviour anywhere and wherever possible I have checked the noisefloor with a calibrated VNA. The values that are being generated by the VNA agree with those established by this codebase.

[usrflo]

I checked your implementation in comparison to my #2933 that goes into the same direction.
In #2933, I accept every idle measurement and reduce the block to the median - outliers on both sides are eliminated without the need for additional filtering. This step depends on some more parameters in your implementation: is there a reason to prefer this logic (trusted_published_floor) over the simplified median approach?

[ACETyr]

@yg-ht Bench comparison of this PR against #2933 and their common base, requested on #2933 — full data
and method there: #2933 (comment). Summary for this PR:

• RAK4631 / SX1262, 869.618 MHz / 62.5 kHz / SF8 / CR5, int.thresh = 0, agc.reset.interval = 0,
  built from the PR head bafa673a with MESH_DEBUG=1 — no rebase, no merge. Stimulus: 250 zero-hop
  adverts at 0.4 s from a node ~1 m away, ~100 s at roughly 75 % channel occupancy.
• Held −105/−106 throughout, idle and loaded. No descent, no clamp, no wedge. It stops publishing
  for the duration (one 126.5 s gap) and resumes ~26 s after the channel clears — suppressed, not
  stuck. The unpatched base under the same stimulus pins at −120 within 17 s.
• stats-noise is what made that readable: accepted froze at 47/64, rejected_high climbed
  31 → 61 → 70, then the 60 s window expired and the batch restarted. Worth keeping.
• Idle block cadence is 4.1 s vs 2.0 s on the base, which is just the 50 ms rate limit × 64 samples.
• Unit tests: all 26 cases in test_rssi_packet_metrics pass, 39 across the native env, with
  g++ 15.2. As a check that the suite is actually wired to the code under test, forcing
  HIGH_BOUND_REJECT_JUMP_DB to 200 fails LowFloorStillRejectsLargeUpwardJumps and
  StrongRssiSamplesDoNotContaminateHealthyBatch; reverting turns them green again.

Two merge notes, neither about behaviour:

• The PR no longer merges cleanly onto current dev (f6c25e6a). RadioLibWrappers.cpp/.h still
  auto-merge; the conflicts are confined to CommonCLI.cpp/.h.
• The cause is that the four new prefs are appended to the positional blob at offsets 295–302, and dev
  has since replaced that blob with the /prefs.json ConfigSerializer (a0b52be4).