Skip to content

LR2021 legacy calibration on W12 - 2026-11-08

The final feb7c43dd candidate passed independent legacy calibration, guarded primary restoration, multi-SF RF checks and a real 240-second combined primary plus old-calibration cycle on the Meshnology W12, with no reported failure or unplanned recovery. Supporting 0d880be86 runs also passed native RXPS and single-profile multi-SF. The normal build passed its final 3/3 RF check, and both radios' recorded settings were restored and verified. These are functional bench results, not sensitivity or lossless-traffic proof.

Setup and source

A Meshnology W12 ran diagnostic Full Companion firmware with MESH_AGC_TIMING_DIAGNOSTICS=1. A RAK Tag sent unique direct RAW_CUSTOM probes. These were nearby-radio tests with matched tuples, not controlled weak-signal or power measurements. Dual-profile probes used 909.5/919.5 MHz, SF7 and 128-symbol transmitted preambles; multi-SF probes varied the primary SF.

  • Final hardware candidate: feb7c43dd; regression coverage: e61c08960.
  • Supporting full-mode hardware results: 0d880be86.
  • Supporting independent legacy tests used set agc 0, set agc.old 4.
  • Counts below cover individual test windows, not the whole boot.

Final candidate: both maintenance halves verified

Stage Result Evidence
Independent legacy calibration Three successful native helpers Helper count 0→3
Short primary refresh, agc=4, agc.old=auto Three requests admitted and completed; zero failures req=3, admit=3, ok=3, fail=0; cold starts 1→4; old-helper count stayed 3
Multi-SF RF checks 3/3 received with matching actual tuples Primary SF5 and SF8; secondary SF7, verified by get pramscan.last
Real 240-second combined cycle Primary refresh and separate automatic old half completed Primary successes 3→4; old helpers 3→4; old diagnostics pr=4, pa=4, skip=0; cold starts 4→5
RF after the combined cycle 2/2 received SF7 primary and secondary probes

After the combined cycle, PRAM was on and hooked, with zero uncertainty, recoveries, bad handoffs, lost preambles, device/command/sequence errors and pending repair. All four native old-helper calls took 26.301-26.357 ms; admission to continuous RX took 33.787-33.864 ms.

The combined check waited for the actual 240-second cadence. Its roughly 250-second observation included nine-second polling and RF probes; that is not the duration of a maintenance operation or an RF outage. The cold restore belongs to primary restoration: the separate old-helper increment proves it was not silently counted as satisfying the automatic old half.

Supporting functional stages on 0d880be86

Stage Native calibrations RF result Native helper timing Admission to continuous RX
Single profile, continuous RX 7 3/3 baseline; 12/12 during maintenance 20.586-21.221 ms 31.939-32.599 ms
Quiet dual-profile PRAM 5 12/12 alternating profiles after maintenance was disabled Last 26.264 ms Last 33.755 ms
Quiet dual-profile PRAM, primary multi-SF5-8 4 6/6 after maintenance was disabled Last 26.272 ms Last 33.770 ms
Single profile, native RXPS level 6 9 12/12 incoming; 1/1 outgoing — —
Single-profile multi-SF5-8 5 8/8 during maintenance — —

All these stages ended with zero reported device, command and sequence errors and no pending repair. The PRAM timings are last samples, not isolated ranges: their cumulative counters also included earlier single-profile samples.

The five quiet dual-profile calibrations advanced native helpers 8→13 and warm resumes 1→6 without changing cold starts (4). Intentional multi-SF reconfiguration increased cold starts to 5; four more calibrations advanced native helpers 13→17 and warm resumes 0→4 without another cold start. Both stages kept the scanner hooked, with no errors, bad handoffs or recovery. Subsequent RF checks passed 12/12 and 6/6 with zero lost-preamble increments. The six multi-SF probes included actual primary SF5, SF7 and SF8 and secondary SF7, verified by get pramscan.last. Quiet-then-RF stages do not establish simultaneous-traffic qualification.

The RXPS stage explicitly reported effectively armed level 6, configured receive preamble 16, RX 41704 us and sleep 11882 us; the source transmitted 128 preamble symbols. Its valid 12/12 incoming and 1/1 outgoing results cover nine legacy calibrations, not desired RXPS running in continuous fallback.

Three supporting short primary cycles also passed with actual guarded cold restores and no old-helper increment. Their lost-preamble counter rose by 1 without feeder transmissions; that counter was not treated as a lost test packet.

Implementation principles

Legacy maintenance uses checked RC-standby Calibrate(0x6F) and frontend calibration at the actual current frequency and all enabled profile frequencies. Correction is applied once; encoded 4 MHz points and RF paths are deduplicated into up to three frontend slots. Missing frontend coverage (0x0200) is repairable by that sequence; unrelated busy-trigger and oscillator/PLL-lock faults remain preserved and rejected.

Ordinary single-profile/RXPS legacy calibration includes retained sleep/wake. A known-resident PRAM/LDO session omits only that sleep/wake step, retaining full checked calibration and verified scanner resume.

Primary duty-cycle refresh does sleep. Its owned cold restoration preserves packet/pulse and generation guards, performs an actual checked radio reset, calibrates both enabled profiles, then reloads and verifies PRAM. It never rewrites an active image. Descriptor words alone do not prove working code or mapping; native version, SCN3 identity, hook and generations are checked.

Cold restoration can add full calibration below the 240-second policy threshold. It belongs to primary restoration and does not satisfy the separate automatic old half at 240 seconds or longer. The final combined test verified both stages. Pure queued noise-floor requests do not block the checked matching paused-PRAM maintenance lease; active sampling batches/windows still do.

Limits and validation status

Earlier concurrent-traffic stress encountered a post-stop activity IRQ (001100A0) with FIFO level 0. The safety guard refused the handoff and bounded recovery selected host fallback. That traffic case remains unresolved by the calibration fixes; PRAM/multi-SF can remain off until profile change or reboot.

Timing covers software boundaries, not RF downtime or analog AGC state. Native-helper samples include sleep/wake only on the single-profile path. Admission-to-continuous-RX includes host restoration but excludes the earlier scanner pause/standby and later rehook. No calibrated sensitivity, battery saving or interference-performance claim follows from these tests.

All 165 focused host tests passed (68 LR2021 and 97 shared), as did the final full PRAM suite and normal W12 build. The W12 application uses 2355688 of 6553600 bytes. LILYGO X1 Repeater and Companion builds passed; X1 has build-only evidence.

Normal build and final cleanup

The W12 was left running the normal feb7c43dd Full Companion image with timing diagnostics disabled. With agc=0 and agc.old=4, it received 3/3 additional probes and reported radio.apply code=0 repair=0. The normal image SHA-256 is 057ea2302e5518b09c6eb6a7c988a120181c09cac26349522a1e1bb13e89c977.

Final reads verified both radios back on 910.525 MHz, BW62.5, SF7, CR4/5 with automatic preambles. W12 transmit power is 5, trim is 9,9, primary multi-SF5-8 is on, radio2 is off with automatic operation allowed, and cross-repeat is off. Power saving is on; desired RXPS remains level 6 with preamble 16 and its expected continuous-RX hold for multi-SF. Maintenance is agc=16, agc.old=auto. The RAK Tag has transmit power 22, agc=16, old maintenance off, RXPS off, power saving off and radio2 manually off. Both serial ports were released after verification.