Document repeater validation and operating limits

This commit is contained in:
Lars Nolden
2026-09-21 12:31:22 +02:00
parent 8c44c54131
commit 3a0d342314
11 changed files with 890 additions and 267 deletions
+151 -103
View File
@@ -5,18 +5,40 @@ the development bench to a real site. Goal: run the repeater standalone
(no USB host) and verify it over RF alone. This log records what happened,
because the sequence of wrong theories is itself a finding.
## Final correction (same date, after source and IQ audit)
The late-evening conclusion that the board had an intermittent TX-chain
hardware fault is **withdrawn**. Two software/methodology faults explain
the evidence:
1. The `pa_low` diagnostic used SetPaConfig
`(0x01,0x00,0x00,0x01)`. Its PA-duty and hpMax bytes came from the
SX1261 +10 dBm row, but `deviceSel=0x00` still selected SX1262. This
undocumented hybrid is not a low SX1262 row. Same-capture A/B testing
against valid SX1262 `(0x02,0x02,0x00,0x01)` measured about 32 dB more RF.
2. The beacon verdict called RF absent from total energy alone. Spectral
reanalysis of the canonical “dead” IQ capture finds weak but repeated
FSK at all 25 TX times: median +1.5 dB over floor versus +0.2 dB in
nearby controls, with +7.6 dB FSK-selective excess.
After deleting the invalid PA row, a flashed-firmware smoke test produced
9/9 beacons at median +27.81 dB over floor (range +27.02..+27.90), nearby
controls +0.00 dB, and +40.73 dB FSK-selective excess at configured
-9 dBm. This report retains the earlier narrative as a chronology of
misdiagnosis; the final conclusions are in the corrected section below.
## Timeline of evidence
| Setup | LED (TxDone) | RF at SDR | Verdict at the time |
|---|---|---|---|
| Laptop USB, board on table (~16:07) | double-blinks | **7/7 TX clock-aligned, median +37 dB over floor** | baseline: transmits |
| Laptop USB, board on table (~16:07) | double-blinks | reported 7/7 at +37 dB; later invalidated | baseline claim came from the repeat-borrowing clock-fit bug |
| Power bank (USB-C, then USB-A) | went dark after seconds | none | bank auto-off at ~60 mA idle load |
| UGREEN 100 W GaN wall charger, C-to-C | double-blinks "sometimes" | none in 4 captures (~8 min) | *wrongly* blamed supply sag |
| LiPo pack attached | blinks, 2-4 double-pulses/90 s | none | supply theory started cracking |
| LiPo + dangling USB cable (no host) | ~9 double-pulses/90 s | none | RX rate recovered; TX still invisible |
| Mid-capture unplug (laptop -> LiPo, position frozen) | blinks throughout | **none in either phase** | laptop connection alone does NOT explain it: power theories all dead |
| Wall power, board hanging free, no LiPo | blinks | none at prior SDR-antenna position | placement of the DUT alone is not it either |
| **SDR antenna moved to ~20 cm from the hanging board** | blinks | **same-duration burst pairs at 50-54 ms: 10-10, 17-17, 16-16 ms** | repeater IS repeating standalone; radiated/received level is simply far lower than the laptop-connected reference |
| **SDR antenna moved to ~20 cm from the hanging board** | blinks | **same-duration burst pairs at 50-54 ms: 10-10, 17-17, 16-16 ms** | standalone RF exists; comparison with the invalid laptop reference is unsupported |
## What was actually established
@@ -24,23 +46,18 @@ because the sequence of wrong theories is itself a finding.
RX->50 ms->TX cycle runs (LED) and its repeats are observable at RF
(same-duration pairs at the configured delay) once the observer antenna
is close enough.
2. **The laptop-connected "+37 dB at 50 cm" reference was not a pure
radiated-field measurement.** With the board on the host's USB, the
cable shield and shared ground with the SDR dongle add counterpoise
and possibly conducted common-mode coupling. Standalone, at a similar
claimed distance, the same transmitter was invisible; at ~20 cm it shows
up at +14..+24 dB. The gap between those numbers is the combined effect
of geometry, orientation, antenna environment and the lost cable/ground
contribution -- not a firmware or supply fault.
3. **Every power-source theory failed when tested properly.**
- Power bank: real failure, but trivial (auto-off). The LED found it
in one glance.
- Wall-adapter "sag during PA burst": WITHDRAWN. The adapter is a
100 W GaN unit, and the frozen-position unplug test showed the
laptop connection failing at the same spot too. The earlier
conviction (committed to FINDINGS.md at the time) was wrong.
- LiPo (stiffest possible rail): same invisibility -> supply exonerated
entirely.
2. **The laptop-connected “+37 dB at 50 cm” reference was invalid.**
The old clock fit could align each expected repeat to its original
meter burst. It cannot establish a USB-cable counterpoise or conducted
coupling effect.
3. **No tested power source correlated with the weak RF.**
- Power bank: real failure, but trivial (low-load auto-off). The LED
identified it directly.
- Wall adapter and LiPo both showed chip-level TX with weak/invisible RF.
- The frozen-position unplug run did not isolate a supply effect.
These observations reject the original wall-adapter-sag theory, but
“supply exonerated entirely” was stronger than the available evidence.
The later PA-row bug is sufficient to explain the large RF deficit.
4. **TxDone still proves nothing about radiated power.** The task spec's
warning (§36: "TX successful is not evidence") extends beyond FEM
configuration to antenna geometry: the SX1262 completes its packet into
@@ -62,110 +79,141 @@ because the sequence of wrong theories is itself a finding.
- **A conducted RF test port would have ended this in minutes.** Half a day
of over-the-air inference replaced one attenuated SMA measurement.
Already in devboard_signals.csv; priority upgraded.
- **Antenna system design is not optional polish.** The stock whip against
a tiny PCB ground radiates so poorly, and so directionally, that
deployment placement dominates link budget. The gateway needs a real
antenna with a defined counterpoise, and the battery pack must be kept
mechanically away from the radiator.
- **On-device persistent telemetry (NVS counters, VBAT) is still the
missing piece.** The LED collapsed "is it powered/alive/repeating" from
four SDR captures to one glance, but every RF question still required
the laptop. Retained counters would have separated "not receiving" from
"not transmitting" instantly.
- **The antenna conclusion was premature.** Geometry, orientation and
counterpoise always affect over-the-air measurements, but the invalid
SX1261-derived PA configuration was not controlled and accounts
for about 32 dB by itself. A proper antenna A/B remains useful; this
session did not quantify stock-whip efficiency.
- **Persistent counters would improve, but not prove, RF diagnosis.** NVS
RX/TxDone counters and VBAT would separate power, receive, and chip-TX
failures without a host. They still could not distinguish weak RF from a
healthy radiated packet; that requires a conducted detector or external
receiver.
- **LED patterns earned their place** and should carry to the gateway
(with the FUTURE_UI.md caveat: an LED can only report chip state, never
radiated RF).
## Still open after this session
## Questions left by the afternoon run (status after audit)
- Clock-aligned close-antenna run (device on laptop serial + SDR antenna
at 20 cm) to attribute bursts per TX event and put a number on the
standalone radiated level. Protocol ready; not yet executed.
- RESOLVED for the attached-DUT method: meter-anchored event windows plus
spectral controls now attribute RF to each TX. Absolute standalone field
strength is still uncalibrated.
- Quantify laptop-cable contribution: repeat the alignment run twice at
identical geometry, cable in vs out (LiPo), antenna close.
- Proper antenna A/B (e.g. quarter-wave with radials vs stock whip) --
directly relevant to the gateway RF architecture.
## Location 3 (evening session) -- the beacon, and the real fault
## Location 3 (evening session) — corrected beacon postmortem
Coordinating physical actions with an async agent kept failing (missed
timestamps, tuner deaths mid-capture), so we changed instruments: a
firmware **beacon mode** (`set beacon_period_ms N`) transmits a synthetic
OMS frame (manufacturer SLN, ID 00000001, access_nr = sequence) every N ms.
No meter traffic, no serial coordination, dense uniform sampling.
Beacon mode (`set beacon_period_ms N`) was the right instrument: it emits
known frames at known times without relying on meter traffic. The first
verdict applied the wrong detector to that data.
The beacon caught what two days of repeat-chasing could not:
### What the retained IQ establishes
- **TX radiated power is intermittent at minute scale.** One 90 s window:
0/25 beacons produced ANY RF (max +0.5 dB over floor) while 27 ambient
meter frames in the same IQ showed median +26.8 dB (observer healthy) and
the chip reported TX_DONE, mode 6, device_errors 0 for every beacon.
Thirty minutes later, same boot, same config: repeats, test_tx and
beacons all radiated at a constant ~+10 dB.
- RX stayed healthy through dead TX windows. Chip-side diagnostics are
identical in dead and live states. The fault therefore sits between the
SX1262 PA output and the antenna: KCT8103L PA path (CPS/DIO2, CTX, PA
supply) or a mechanical/thermal TX-chain fault. Firmware cannot observe
it -- exactly the RADIO_VDD_SENSE / conducted-test-port gap in
docs/devboard_signals.csv.
- **Methodology bug found and fixed:** verify_repeats_iq's single-burst
clock fit can lock at -repeat_delay and "confirm" every repeat with its
own original's burst. The "7/7 at +37 dB" laptop confirmations were this
artifact. Trustworthy methods: structural same-duration pairs at the
configured delay, content decodes (duplicate payloads), and
hil/meter_anchored.py (clock fitted on strong meter receptions; ambient
meter levels double as per-capture observer health).
- Positive control: with a healthy observer, 30 structural pairs at
137 ms and 35 duplicate decoded payloads in 90 s -- the repeater
function itself is solid whenever the TX chain is in its live state.
- The historical capture
`captures/2026-09-01T17-35-12-beacon-diag` has a healthy observer:
27/32 strong meter anchors, 0.7 ms median clock error, meter median
+26.8 dB.
- Beacon windows are weak in total energy (median +1.5 dB; none reaches
+6 dB), but their median spectrum contains the expected approximately
+/-45 kHz FSK energy. Relative to nearby non-TX controls, FSK-selective
excess is +7.6 dB across 25 windows.
- A known-live repeat capture
(`2026-09-01T17-41-21-repeat-pair-test`) classifies strong: 29/29 above
+6 dB, median +10.9 dB, FSK-selective excess +18.8 dB.
- Therefore “0/25 produced any RF” and “genuinely zero RF” were false.
TxDone did not prove RF, but the retained spectrum does.
## Location 3, late evening -- soak verdict and physical isolation attempt
The soak's exact historical PA-row state cannot be reconstructed: serial
telemetry logged SetTxParams power but not the `pa_low` selection, and the
old soak harness deleted IQ after reducing each segment to scalar energy.
Its claimed +8 -> +6 -> zero thermal decay is consequently not auditable.
A board-level thermal/mechanical TX fault was never demonstrated.
Antenna swap + IPEX reseat performed (TX disabled during swap). Immediate
post-swap check: 74/74 TX events radiating. Beacon soak started to look
for recurrence (2-min segments, meter-anchored verdicts, observer health
gating, dongle auto-recovery).
### Source-level root cause
Soak result: **the fault recurred post-swap.**
The firmware's alleged low-power SX1262 PA row was copied from the wrong
line of Semtech table 13-21:
- Segments 1-8 (~16 min usable): alive, TX median +7.0..+8.8 dB.
- Segment 9: marginal (median +5.9) -- levels visibly sagging.
- Segment 13: **dead** -- 106 TX attempts, median +0.5 dB, observer
independently verified healthy (56/60 meter anchors at 2.2 ms clock
error, meters at +21.8 dB). Airtight.
- The decay pattern (+8 -> +7 -> +6 -> dead over ~10 min of sustained
~50 TX/min) suggests but does not prove a thermal mechanism.
```
old diagnostic row: 01 00 00 01 undocumented SX1261-values/SX1262-select hybrid
fixed SX1262 row: 02 02 00 01 documented +14 dBm-optimal PA config
SetTxParams: -9 02 configured level -9 dBm, 40 us ramp
```
**Conclusion: the old antenna and pigtail are acquitted. The intermittent
TX-kill is on the board itself**, in the chain between SX1262 PA output
and the antenna connector (KCT8103L FEM prime suspect), invisible to
every chip-level diagnostic.
Same-stream A/B measurement found about 32 dB between the two rows. The
invalid path has been removed rather than retained as a diagnostic.
Configured SX1262 power is now limited to -9..+2 dBm; the default remains
-9 dBm because the external KCT8103L adds gain and absolute e.r.p. has not
been calibrated.
Live provocation session (35 s verdict cadence):
- Baseline: marginal-alive (+5.2), died to +2.0 within ~2 min (pressure
state during death window not recorded).
- ~90 s of breath-cooling the FEM area: no recovery (+3). Weak cooling,
so thermal is not acquitted -- merely unconfirmed.
- Pressure test: not completed; session ended.
### Corrective verification
Instrument attrition tonight: the RTL-SDR dongle required one physical
replug, dropped off the USB bus once more, produced four short/empty
captures and two deaf windows. Roughly a third of tonight's wall time
went to working around the observer rather than observing the DUT.
Flashed fixed firmware, 1.6 Msps SDR, fixed 0.9 dB tuner gain:
```
scheduled beacons 9/9 TxDone
RF envelope over floor +27.81 dB median (+27.02..+27.90)
nearby non-TX controls +0.00 dB median
FSK-selective excess +40.73 dB
SDR clipped samples 0.000 %
computed airtime accounting 24,480 us = 9 x 2,720 us
```
The detector also scored nine fake event times in an off interval at
-0.16 dB FSK excess, versus +40.73 dB for the real aligned events.
`hil/meter_anchored.py` now reports `rf_state=strong|weak|not_observed`
and requires event-correlated FSK spectral excess before claiming RF.
`hil/beacon_soak.sh` retains raw IQ for every weak, absent or unclassified
segment. Strong segments alone may be reduced to a JSON verdict.
The unrelated duty-cycle bug found during the audit is also fixed: the
384-event ring silently forgot airtime at high rates. A 61-minute bucket
ledger now retains every event overlapping the rolling hour, and duty
usage is computed from packet bits rather than SPI/polling latency.
## Where this leaves the project
1. Repeater firmware: fully functional and standards-shaped -- structural
pairs, duplicate decodes, duty accounting (modulo ring saturation,
see OPEN_QUESTIONS), recovery ladder, LED patterns, beacon mode.
2. This particular board: not deployable unattended until the TX-chain
fault is fixed or understood. Next diagnostic steps in order of value:
a) proper cooling/heating provocation (freezer spray / hot-air at low
temp) with the live monitor; b) pressure/flex mapping; c) conducted
measurement through an attenuator to split FEM vs antenna-port; all
need bench equipment or at least controlled temperature, not breath.
3. The observer (NESDR SMArt v5) is end-of-life for serious use: replace
before the next HIL session.
4. Every hard-won methodology upgrade is now in the repo: beacon mode in
firmware, hil/meter_anchored.py (borrow-immune verification + observer
health), hil/beacon_soak.sh (unattended state mapping).
1. The connected board demonstrably transmits correct FSK after the PA-row
fix. No intermittent board-level TX fault is currently established.
2. TxDone and the LED remain chip-level evidence only. RF conclusions need
the spectral detector or an independent receiver/decode.
3. Absolute radiated power and regulatory e.r.p. remain uncalibrated. Keep
the configured SX1262 level at -9 dBm until a conducted measurement or
calibrated field-strength setup exists.
4. The NESDR observer remains unreliable for long unattended work; its
per-capture health gate and anomaly-IQ retention are mandatory.
## Late evening — standalone battery test with fixed firmware: PASS
First deployment after the PA-row fix and the OLED/button UI. DUT on
battery, no USB host; all verification RF-only plus the on-device screen.
Capture: `captures/2026-09-01T22-22-32-standalone-battery` (360 s IQ at
42.1 dB gain, verdicts archived alongside).
- **Repeat function confirmed standalone** by two independent methods on
one stream: 3 decoded duplicate payloads, and 3 structural RF pairs with
matching durations (10-10, 7-7, 10-10 ms) at 53-54 ms gaps against the
50 ms default delay. The +3-4 ms is the radio-preparation offset between
TX-routine entry and RF onset (standby + packet params + buffer write),
consistent with `tx_operation_us` measurements.
- **Observer health control in-capture**: 46 C1 decodes from 38 unique
meters; floor +2.9 dB.
- **Counter funnel closed from the OLED alone** (no laptop): RX 292 =
VALID 100 + CRCFAIL 63 + NOT-C1 130 + LEN 0 (one frame arrived between
page reads). TX 109/109 at 100%. The 34% valid ratio matches the site
benchmark; CRC failures are the characterized long-encrypted-Techem
sensitivity edge, not a fault.
- **UI-driven beacon toggle verified end to end** earlier in the session:
an independent wmbusmeters receiver decoded beacon ID 00000001 with
valid CRC every ~3.3-3.5 s (3000 ms + scheduling drift) after a button
press, on battery. Beacon was off during this capture (0 beacon decodes,
no comb -- matches the TX counter arithmetic).
- The 44 railed ~+40 dB short bursts in the capture are our repeats PLUS
the fingerprinted neighbour meter's singles; only delay-and-duration
matched pairs were attributed to the repeater.
- Duty note for future runs: the 3000 ms button beacon consumes 907 of
1000 ppm; running it for long alongside repeats drives `DUTYBLK` up by
design. Toggle it off when measuring repeat behaviour.
+102 -51
View File
@@ -19,14 +19,15 @@
- TX proof of life: SDR decoded our synthetic C1 format-A frame as valid
(crc_ok=1, mfr 0x4D8E "SLN", ID 00000001) at configured -9 dBm, and the
wM-Bus CRC blocks computed by firmware matched rtl_wmbus's checks.
- First real repeat observed: 100-byte Techem frame received, repeated
after 50.002 ms (configured 50 ms).
- Nooelec NESDR SMArt v5 R820T PLL wedges ("PLL not locked"), producing
gain-independent white noise on all frequencies until USB reset.
`usbreset 0bda:2838` reliably restores operation. Trigger correlated
with `rtl_test -t` direct-sampling probe.
- rtl-sdr-blog and osmocom rtl_sdr builds behave identically w.r.t. the
wedge; not a driver-fork issue.
- First real repeat observed: a 100-byte Techem frame was received and the
firmware entered its TX routine 50.002 ms after the reconstructed frame
end (configured 50 ms). This timestamp precedes radio preparation; it is
not the exact on-air start time.
- The Nooelec NESDR SMArt v5 can enter gain-independent, railed-noise
states until USB reset or physical replug. The tune-time
`[R82XX] PLL not locked!` message is benign and `rtl_test -t` was not the
cause. `hil/sdr_health.py` diagnoses the actual failure from gain response
and spectral structure.
- Real C1 traffic exists at this site: 8 meters in 10 min baseline
(docs/SDR_BASELINE.md); one strong plaintext meter (61471660), several
marginal encrypted Techem meters.
@@ -41,6 +42,18 @@
called) -> no notifications despite correct radio-side IRQ config.
3. `timeout bash -c "rtl_sdr | rtl_wmbus"` orphans the pipeline children
on timeout (device stays claimed). Apply timeout to rtl_sdr directly.
4. The `pa_low` diagnostic selected SetPaConfig
`(0x01,0x00,0x00,0x01)`, documented in code as a low SX1262 row. Its
first two parameters came from the SX1261 +10 dBm row, but
`deviceSel=0x00` selected SX1262, creating an undocumented hybrid.
Same-capture A/B testing showed about 32 dB less RF than the valid
SX1262 row `(0x02,0x02,0x00,0x01)`. Fix: remove the invalid row
and diagnostic toggle; always use a valid SX1262 row.
5. The rolling-hour duty ledger retained only 384 individual events. At
beacon rates it wrapped after about 2.7 s of airtime and silently
under-counted the hour. Fix: 61 keyed minute buckets retain all airtime
in the overlapping hour; host tests cover high event counts, expiry,
boundary overlap, and invalid timestamps.
## LIKELY
@@ -61,20 +74,23 @@
docs/RF_FRONTEND_TESTS.md.
- Repeat payloads are byte-exact: 63/63 logged `tx_payload` values equalled
mode bytes + the received raw frame. No corruption in the RX->TX path.
- Real repeats are emitted as correct RF: aligning firmware TX timestamps
to bursts in the same stream's IQ confirmed 48 of 57 TX events (84 %),
with alignment error median 1.3 ms (max 3.6 ms) and burst duration within
0.5 ms (median) of the computed airtime. Firmware repeat delay measured
200.0 ms min/median/max against a 200 ms setting.
- Real repeats are emitted as correct RF: aligning firmware TX-operation
timestamps to bursts in the same stream's IQ confirmed 48 of 57 TX
events (84 %), with alignment error median 1.3 ms (max 3.6 ms) and
burst duration within 0.5 ms (median) of computed airtime. Firmware
entered the TX routine at 200.0 ms against a 200 ms setting; radio
preparation shifts actual RF onset later.
- Firmware TX reliability: 58/58 TX success over 600 s and 72/72 over
1200 s; tx_timeout 0, tx_failures 0, radio recoveries 0, MCU resets 0.
- The KCT8103L RX path is essential: holding CSD low in RX costs ~41 dB of
chip-referred level and drops the visible meter population from 18 to 1
(600 s per arm, docs/RF_FRONTEND_TESTS.md). No evidence of excessive
front-end gain hurting at this site.
- Repeat-delay accuracy: +1 to +8 us against a 200 ms setting across 72
repeats. RX deaf time per repeat ~225.7 ms median, of which 12.2 ms is
the fixed-255-byte RX tail and 13.4 ms the TX airtime (docs/TIMING.md).
- Repeat-operation scheduling: +1 to +8 us against a 200 ms setting across
72 repeats. This is accuracy at entry to the TX routine, not RF onset.
The historical 13.4 ms value called “TX airtime” was operation latency
through observed TxDone; computed channel occupancy was 8.8 ms for the
representative 102-byte radio payload (docs/TIMING.md).
- The SX1262 GFSK engine cannot express wM-Bus length semantics, so RX uses
a fixed 255-byte capture; RxDone therefore fires 9.4-17.1 ms after the
real end of frame and the firmware reconstructs the true timestamp.
@@ -147,36 +163,38 @@
- Long-run stability at >=100 and >=500 verified repeat events.
- Early-stop RX (poll GetRxBufferStatus) to cut the ~20 ms deaf tail.
## Location 3 follow-up (2026-09-01 evening): intermittent TX RF death
## Location 3 follow-up (2026-09-01 evening): PA configuration postmortem
- CONFIRMED: the DUT's radiated TX output is intermittent on minute scales
-- genuinely zero (not weak) for >=10 min windows, then a stable ~+10 dB
at the SDR. Verified with beacon mode against in-capture ambient-meter
controls (observer healthy: meters +26.8 dB median while 25/25 beacons
produced <= +0.5 dB). RX unaffected; chip reports TX_DONE/mode 6/no
errors in both states => fault is post-chip (KCT8103L PA path or TX
chain), invisible to firmware. Hardware follow-up required; strengthens
the RADIO_VDD_SENSE and conducted-RF-test-port devboard requirements.
- UPDATE (same evening): antenna + IPEX pigtail EXCLUDED -- after a full
antenna swap and pigtail reseat the fault recurred (soak segment 13:
106 TX attempts at median +0.5 dB with observer verified healthy).
The fault is ON-BOARD. Decay signature before death (+8 -> +6 -> dead
over ~10 min under ~50 TX/min load) points at thermal in the FEM PA
path; a 90 s breath-cooling attempt did not revive it (inconclusive --
needs real thermal provocation or conducted measurement).
- WITHDRAWN: every single-burst "TX aligned to IQ" confirmation (e.g.
"7/7 at +37 dB", "33/41 at +19.3 dB"). verify_repeats_iq's clock fit on
TX events can lock at -repeat_delay and match each repeat to its own
original's burst. Borrow-immune evidence only: structural same-duration
pairs at the configured delay, duplicate content decodes, or
hil/meter_anchored.py (meter-anchored clock, exact TX windows).
- CONFIRMED: observer health must be judged per capture from ambient meter
levels inside the same IQ (meter median >= +15 dB over floor). The
Nooelec went deaf mid-session again; runs without this control are void.
- CORRECTION to yesterday's "placement dominates": placement governs
whether PAIRS are detectable (originals must be SDR-visible), but the
standalone invisibility episodes are better explained by the intermittent
TX death now that it has been directly observed under controls.
- **WITHDRAWN:** the conclusion that this board had an intermittent
post-SX1262 hardware TX death. The canonical “0/25 beacons produced any
RF” capture was judged only by total energy in 1 ms windows. Reanalysis
of the retained IQ finds a repeatable event-correlated FSK signature:
TX median +1.5 dB over floor, nearby controls +0.2 dB, and FSK-selective
excess +7.6 dB across 25/25 windows. It was weak RF, not zero RF.
- Root cause in firmware: the `pa_low` diagnostic combined the SX1261
+10 dBm PA-duty/hpMax values with SX1262 `deviceSel=0x00`:
`(0x01,0x00,0x00,0x01)`. In a same-stream A/B test, documented SX1262
`(0x02,0x02,0x00,0x01)` raised the event envelope by about 32 dB. The
invalid toggle and row are now deleted. Configured power
is restricted to -9..+2 dBm; the external KCT8103L still follows.
- The old logs recorded configured SetTxParams power but not which PA row
the diagnostic had selected. Exact row state cannot be reconstructed for
every old soak segment, and the old soak deleted each IQ file after its
scalar verdict. Therefore its claimed +8 -> +6 -> zero thermal decay is
not auditable and is not evidence of a board fault.
- Fixed-firmware RF smoke test at configured -9 dBm: 9/9 scheduled beacons,
median +27.81 dB over floor (range +27.02..+27.90), nearby controls
+0.00 dB, FSK-selective excess +40.73 dB, and no ADC clipping at 0.9 dB
SDR gain. Firmware also reported 9/9 TxDone and exactly 24,480 us of
computed airtime (9 x 2,720 us).
- `hil/meter_anchored.py` now requires event-correlated FSK spectral excess
before calling RF present and reports `strong`, `weak`, or
`not_observed`. The historical weak capture classifies `weak`; a known
live repeat capture classifies `strong`. `hil/beacon_soak.sh` deletes IQ
only for `strong` segments and retains every anomaly.
- Observer health must still be judged per capture from ambient meter
levels. TxDone remains only chip-level evidence. Neither fact supports a
board-fault conclusion without the RF signature and retained raw IQ.
## Field test 2026-09-01 (onsite standalone deployment)
@@ -192,12 +210,14 @@ Full narrative with the evidence table: docs/FIELD_TEST_2026-09-01.md.
~20 cm of the free-hanging, wall-powered board, same-duration burst pairs
appeared at the 50 ms repeat delay (10-10, 17-17, 16-16 ms; received
+14..+24 dB over floor).
- CONFIRMED: the laptop-connected "+37 dB at ~50 cm" reference overstates
standalone radiated field -- the USB cable/host ground contributes
counterpoise (and possibly conducted coupling into the shared-ground
dongle). Radiated performance is dominated by antenna geometry:
orientation, surroundings, and the LiPo pack's position relative to the
antenna. Placement, not power, ruled the day.
- **WITHDRAWN:** the claimed laptop-connected “+37 dB at ~50 cm”
reference and the inferred USB-cable/counterpoise delta. The +37 dB
number came from the repeat-borrowing clock-fit bug, so it cannot support
a cable contribution estimate.
- **NOT ESTABLISHED:** that the stock whip itself explains the weak
standalone results. Antenna geometry unquestionably changes an
over-the-air measurement, but the invalid SX1261-derived PA configuration is sufficient to
explain a roughly 32 dB deficit and was not controlled in those runs.
- CONFIRMED: TxDone (and therefore the LED double-blink) proves chip-level
packet completion only. It is indistinguishable between a well-matched
antenna and a shadowed/mismatched one -- §36's lesson extended to
@@ -206,3 +226,34 @@ Full narrative with the evidence table: docs/FIELD_TEST_2026-09-01.md.
singletons", 23-24 ms trains at ~30 ms spacing) are a neighbouring meter
railing the dongle: rate invariant across firmware versions, reboots and
power sources. Excluded from pair detection by fingerprint.
## T-mode + long-frame support (2026-09-02)
- CONFIRMED: T1 transmit works end to end. Synthetic T1 frames (3-of-6
coded in software, 50 kHz TX deviation via lazy SetModulationParams
swap, same 0x543D sync) were decoded by rtl_wmbus as valid `T1;...`
telegrams, including a long L=140 frame (12 decode lines over 6
airings, `captures/2026-09-02T-t1-verify`).
- CONFIRMED: the T1 receive path decodes standard on-air chips. With no
ambient T1 available (below), the proof is: our own T1 airings --
independently accepted by rtl_wmbus -- were FSK-demodulated from raw IQ
by a separate host tool, and the post-sync byte stream was fed through
the unmodified firmware decode code (`wmbus_3of6_first_byte` ->
`wmbus_3of6_decode` -> `wmbus_check_crc`): 5/5 CRC-OK with the exact
transmitted content. A C1 burst pushed through the same path is
correctly rejected at the first coded byte.
- CONFIRMED (site): there is NO ambient T-mode traffic at this site --
every ambient rtl_wmbus decode is C1. **Correction to earlier belief:**
the large `rx_not_c1` bucket (44 % of sync matches overnight) is NOT
T1 meters; it is C1 frames whose two NRZ mode bytes were corrupted by
bit errors, plus false syncs from the deliberately short 8-bit preamble
detector. The Techem meters here are C-mode.
- Long frames: format B >128-byte CRC chunking and format A up to L=200
are host-tested (sanitizers) and the L=200 format A frame was decoded
CRC-OK over the air by rtl_wmbus.
- T1 length cap: the fixed 255-byte capture holds ceil(total * 1.5) coded
bytes, so T1 receive/repeat covers format A total <= 170 (L <= ~150).
Long C1 covers total <= 253.
- 3-of-6 tables verified against rtl-wmbus t1_c1_packet_decoder.h; codes
are constant weight 3, so every single-chip corruption is detected
(host test flips all 384 chip positions of a 32-byte payload).
+93 -6
View File
@@ -1,10 +1,9 @@
# Future feature note: OLED + button for immediate local feedback
# OLED + button for immediate local feedback
Status: **design note only. Not implemented.** The task description
explicitly excluded a "production OLED UI" from this experiment, and the
display is untouched by the current firmware. This note records what a
local UI should show and the constraints it must respect, while the
reasoning is still fresh.
Status: **implemented** (firmware/main/ui.c, 2026-09-01). This note holds
the design rationale and constraints; deviations of the implementation
from the original sketch are listed at the end. Pins were verified against
RTNode-HeltecV4 and are now exercised by our firmware (`ui_init ok:true`).
## Why this is worth building
@@ -182,3 +181,91 @@ does **not** share a rail or bus with anything RF-critical, plus at least two
buttons so navigation and confirmation are separable — a single button
forces gesture overloading, which is how dangerous actions end up behind
timing tricks.
## Implementation notes (what was actually built)
- `firmware/main/ui.c`: esp_lcd SSD1306 panel over the new i2c_master
driver, 5x7 font (`font5x7.h`, host-verified by
`firmware/test/test_font_render.c`), UI task at priority 3 pinned to
core 0 (radio task is priority 10 on core 1). Rendering is a 1024-byte
full-frame draw every 500 ms while awake.
- Data flow: radio task publishes `g_ui` (last RX / last TX / heartbeat
probe results) in `repeater.c`; the UI additionally reads `g_stats`,
`g_cfg`, and `repeater_duty_used_us()`. The UI never touches the SX1262.
- Button: short press = next page / wake; hold ~1.5 s = page action.
Actions emit `{"event":"ui_action","cmd":...,"value":...}` so serial
logs remain authoritative. A press while the display is blanked only
wakes it and fires nothing.
- Page actions: Status = toggle `repeat_enabled`; Health = clear stats;
Config = toggle beacon at a fixed 3000 ms (907 ppm at L=19 — inside the
default 1000 ppm duty budget, so a button-started beacon cannot violate
the guard). Last-telegram page has no action.
- Display blanks after 3 min idle (panel off command; Vext is never
toggled). Init failure is non-fatal: `{"event":"ui_init","ok":false}`
and the task exits.
- Deviations from the sketch: no RESETS counter on the health page (needs
NVS boot counting, still unimplemented); the duty bar shows used/budget
from the minute-bucket ledger; the config page omits `PA LOW` (the
invalid PA row was removed from the firmware entirely) and `fem_rx_csd`
(console-only diagnostic).
## Wi-Fi dashboard and QR addition (2026-09-09)
Implemented in `wifi_debug.c`, `debug_history.c`, `wifi_qr.c` and
`firmware/main/web/`. Two pages follow Config: Wi-Fi credentials and Wi-Fi
join QR. A long press on either page toggles the AP; no boot-time gesture
is used. The existing four page actions remain unchanged.
The radio publishes a separate 256-entry immutable final-outcome history,
including silent/verbose-only rejection and skip paths, with synchronized
per-row copies. Web clients do not read the deliberately racy OLED snapshot
or query the SX1262. Wi-Fi, TCP/IP, HTTP and control processing use core 0;
the radio remains on core 1. No network/JSON allocation was added to the
radio history producer. Final row IDs correlate reception and transmission
directly, not by frame-hash matching.
The AP is off at boot and stays on until explicitly disabled or rebooted;
display blanking is independent. The password is randomly generated once,
persisted in NVS and shown only on the physical display. The QR encodes
the same credentials and is generated once per boot when its page is used.
Its encoder is the unmodified Nayuki C library v1.8.0 under MIT, with
copyright/license notices retained in `firmware/main/vendor/qrcodegen.*`.
The 64-pixel display height requires small integer-scaled modules; do not
crop the quiet zone or promise camera readability without a physical scan.
The dashboard is entirely offline. CSV labels, templates,
exports, manufacturer registry lookups and device-type interpretation run
in browser JavaScript. Labels never reach the device or persistent browser
storage. See README and `docs/TESTING.md` for usage, exact success semantics
and the distinction between localhost, USB, phone and RF verification.
Both surfaces now use German UI text (ASCII transliterations on the OLED).
The phone dashboard puts current events before collapsed diagnosis and
optional CSV naming. A shared, nonmodal **?** help panel explains wM-Bus
terms without hover, stays outside the refreshed rows, and supports outside
click, explicit close and Escape with focus restoration. Its scrollable
body and visible close control are bounded to the viewport. Detail-cell
spans follow the 560 px radio-column breakpoint, including resize, so
expanded rows do not introduce a phantom column on phones.
## Writable settings addition (2026-09-09)
`settings.c` is the common coordinator for all ten existing USB settings and
OLED toggles. It serializes writers and two-key NVS persistence, publishes
accepted values under a short spinlock, and lets the radio owner copy a whole
revision at a processing boundary. A pending-only IRQ check drains completed
captures and defers known in-flight captures; required radio reinitialization
can still interrupt reception. The applied revision acknowledges copying,
not RF success. OLED rendering reads the applied snapshot, while settings
responses and serial status report accepted values plus both revision numbers.
`settings_http.c` adds bounded JSON writes guarded by AP-host/origin checks,
a random per-AP CSRF token and boot/revision comparison. WPA2 access remains
the authorization boundary. The mobile panel has separate drafts/submit
actions, dirty-value preservation, conflict review, persistence labels and
confirmation for dangerous duty limits or AP shutdown. Expert controls remain
collapsed. Wi-Fi shutdown is queued to its control task, not performed inside
the server being stopped. Only delay and duty budget persist, as before.
The build embeds deterministic gzip assets and serves them directly, avoiding
runtime compression and keeping the existing app/NVS partition layout.
+15 -11
View File
@@ -9,9 +9,10 @@
interferers? -> It helps decisively. CSD low costs ~41 dB and reduces
visible meters from 18 to 1; no overload degradation observed
(docs/RF_FRONTEND_TESTS.md).
- CLOSED: RX dead time during retransmit -> ~225.7 ms median at a 200 ms
repeat delay, of which 12.2 ms is the fixed-255 RX tail and 13.4 ms the
TX airtime (docs/TIMING.md). Early-stop RX would cut the tail.
- CLOSED: historical RX-unavailable interval during retransmit was
~225.7 ms median at a 200 ms delay: 12.2 ms fixed-255 RX tail + 200 ms
scheduling delay + 13.4 ms TX operation latency. The representative
packet's computed RF occupancy was 8.8 ms (docs/TIMING.md).
- Can the SX1262 re-arm RX fast enough to catch a meter's immediate
follow-up telegram after our TX?
- Exact production repeater framing/hop requirements (EN 13757-5).
@@ -40,12 +41,15 @@
EN 13757-4 text, or a meter that emits one. Until then, a real meter using
those lengths is an interoperability risk worth a targeted test.
## Duty ledger ring capacity (noted 2026-09-01, loc3)
## Duty ledger event-capacity bug (closed 2026-09-01, loc3)
DUTY_RING=384 entries saturates at ~2.7 s of history at ~7 ms mean
airtimes -- BELOW the 3.6 s budget, so the guard silently never triggers
at high TX rates. OBSERVED ACTIVE during the 2026-09-01 beacon soak
(duty_used_us plateaued at ~3.46 s while true hourly airtime was ~20 s;
effective duty ~0.6 %). Fix: replace the ring sum with an hourly-bucketed
accumulator (or ring + saturation flag that fails CLOSED, i.e. suppresses
when the ring wraps inside the window). Scheduled for next flash.
The old `DUTY_RING=384` implementation saturated at about 2.7 s of history
at ~7 ms per event, below the 3.6 s hourly budget. It silently forgot old
events while they were still inside the hour; this was observed during the
beacon soak.
Fixed: 61 keyed minute buckets accumulate every successful transmission.
The rolling-hour query includes every bucket overlapping the window,
including the oldest partial minute in full (safe over-count). Host tests
cover 10,000 events in one bucket, all 61 overlapping minutes, exact expiry,
invalid timestamps, and zero-length records.
+48 -14
View File
@@ -1,6 +1,34 @@
# Reliability
## Soak result (final firmware, `captures/2026-09-01T11-34-08-soak`)
## Overnight standalone battery soak (2026-09-01/02): milestones A-D reached
DUT on LiPo, no USB host, no observer needed -- firmware counters read via
OLED overnight and archived exactly over serial in the morning
(`captures/2026-09-01T22-48-32-standalone-soak/overnight_final_stats.log`).
```
uptime : 42 339 s (11 h 45 m), zero resets
RX candidates : 9 617 = 4 905 valid + 1 423 crc_fail
+ 3 274 not_c1 + 15 len_invalid (exact closure)
TX attempts / success : 2 822 / 2 822 (100 %)
TX timeouts / failures : 0 / 0
radio recoveries/reinits : 0 / 0
duty suppressions : 2 092 (~41 % of valid traffic over budget)
cumulative TX airtime : 23.72 s computed occupancy (8.4 ms avg/frame)
lifetime duty cycle : 560 ppm (guard held the night at 56 % of cap)
```
Milestones A (>=10), B (>=100), C (>=500), D (>=1000 repeats, no restart):
**all REACHED** in one unattended run. Independent RF confirmation the
following morning: meter-anchored capture verified live repeats (2 decoded
duplicates, 2 RF pairs at 53-54 ms vs the 50 ms setting).
The 2 092 suppressions are the first production evidence for the
STANDARDS_GAPS conclusion that repeat-everything exceeds the 0.1 % duty
budget at a real site (~1.7x here); an installation-list filter is
mandatory for a product.
## Historical soak result (`captures/2026-09-01T11-34-08-soak`)
Repeater running unattended, no SDR involved, repeat delay 200 ms,
duty-cycle guard at 0.1 %.
@@ -18,12 +46,12 @@ TX attempts : 380
TX success : 380 (100.0 %)
TX timeouts / failures : 0 / 0
repeat delay : 200.001 ms min, 200.001 median, 200.006 max
TX airtime : 10.2 / 13.4 / 14.9 ms (min/median/max)
TX operation latency : 10.2 / 13.4 / 14.9 ms (min/median/max)
radio recoveries : 0
heartbeats not in RX : 0
heartbeats config lost : 0
duty-cycle suppressions : 8 (guard working as designed)
old-guard suppressions : 8 (based on operation latency, not RF airtime)
```
## Milestones (firmware-reported repeats, no restart)
@@ -35,23 +63,29 @@ duty-cycle suppressions : 8 (guard working as designed)
| C | >= 500 | not reached (run stopped early at 380) |
| D | >= 1000 | not reached |
Milestone C needs roughly 2 h at the observed rate, or longer once the
duty-cycle guard caps output at ~268 repeats/hour. The run was ended early
by operator request; nothing in the data suggests it would not continue --
the last heartbeat was as healthy as the first, with zero recoveries and
zero resets across 1.5 h. To continue:
Milestone C needs roughly 2 h at the observed rate. The run was ended
early by operator request; nothing in the data suggests it would not
continue -- the last heartbeat was as healthy as the first, with zero
recoveries and zero resets across 1.5 h. To continue:
```
nix develop --command ./hil/soak.sh 9000
```
## Duty-cycle guard behaviour
## Duty-cycle guard correction
At the observed traffic rate the repeater wants ~320 repeats/hour, which at
13.4 ms measured airtime each is ~0.12 % -- above the 0.1 % ceiling for
868.7-869.2 MHz. The guard suppressed 8 repeats to stay inside the budget
(`used_us` 3 594 561 against `budget_us` 3 600 000) and reported a final
`tx_duty_cycle_ppm` of 905, i.e. 0.0905 %. See docs/STANDARDS_GAPS.md.
The old report multiplied ~320 repeats/hour by 13.4 ms TX-operation
latency and called the result ~0.12 % RF duty. For the representative
102-byte radio payload, computed occupancy is 8.8 ms, so the same rate is
about 0.078 %. The eight suppressions at `used_us` 3,594,561 were caused by
the old conservative wall-time metric, not proof that actual channel
occupancy reached 0.1 %. Likewise, the reported 905 ppm was cumulative
operation latency divided by uptime.
Current firmware budgets computed packet occupancy and stores it in a
61-minute bucket ledger. Historical stability/TxDone counts above remain
valid; historical airtime, suppression, and duty-rate interpretations do
not. See docs/TIMING.md and docs/STANDARDS_GAPS.md.
## What has NOT been demonstrated
+32 -2
View File
@@ -41,8 +41,15 @@ RX gain: boosted (reg 0x08AC = 0x96)
TX params: default -9 dBm configured (min), ramp 40 us; PA config
(0x02,0x02,0x00,0x01) = +14 dBm-optimal table, deliberately
capped -- the external KCT8103L PA adds gain on top
```
PA-row invariant: `deviceSel` must remain `0x00` for SX1262, and the other
bytes must come from an SX1262 row. The removed diagnostic
`(0x01,0x00,0x00,0x01)` mixed the SX1261 +10 dBm PA-duty/hpMax values with
SX1262 selection. This undocumented hybrid reduced measured RF by about
32 dB while all chip-side TX diagnostics still passed.
## Consequences of the fixed-255-byte RX capture
RxDone fires ~ (255 - frame_len) * 80 us after the actual end of frame;
@@ -54,5 +61,28 @@ time per frame. Acceptable for the experiment; optimization candidate
## TX power
Configured SX1262 power is NOT radiated power: the KCT8103L PA follows.
Development default is the SX1262 minimum (-9 dBm). Log lines carry
`configured_tx_dbm` to keep this distinction explicit.
Development default is the SX1262 minimum (-9 dBm), and the console accepts
only -9..+2 dBm. +2 is the level used by known-good Heltec V4.3 firmware,
not a declaration of legal radiated power. Keep -9 dBm until the complete
FEM/antenna path has a conducted or calibrated e.r.p. measurement.
`wmbus_tx.configured_tx_dbm` records SetTxParams. `airtime_us` is computed
channel occupancy (48-bit preamble + 16-bit sync + payload at 100 kbps);
`tx_operation_us` is the longer firmware-entry-through-TxDone latency.
## T-mode (T1) support
T1 shares the 0x543D access code, so ONE radio configuration receives
both modes: C-mode announces itself with two NRZ mode bytes after the
sync; T-mode continues directly with 3-of-6 coded data (decoded in
software from the same fixed 255-byte capture; wM-Bus C/T share the
100 kchip/s rate). Tables verified against rtl-wmbus.
TX: T1 frames are 3-of-6 encoded in software and sent with 50 kHz FSK
deviation (raw 0x00CCCD) instead of C-mode's 45 kHz; the swap is a lazy
per-TX SetModulationParams (deviation is a TX-only GFSK parameter, RX is
unaffected). Verified: rtl_wmbus decodes our synthetic T1 frames,
including L=140.
Length caps from the fixed 255-byte capture: C1 total <= 253 bytes;
T1 total <= 170 decoded bytes (coded 1.5x).
+22 -10
View File
@@ -127,6 +127,11 @@ negligible. In the *full* state, however, an earlier IQ capture showed our
repeat at +44.8 dB over floor with IQ amplitude at the dongle's full scale
(rms 88.6 of 127) while the meter original sat at +8.4 dB - a ~36 dB gap.
These TX amplitudes predate PA-row telemetry. They demonstrate the
observer's compression behavior in those captures, but must not be used as
the output level of the fixed firmware: the now-removed `pa_low` diagnostic
could select an undocumented SX1261-values/SX1262-select hybrid.
Consequence: with a fully sensitive dongle, meter originals decode well at
high gain but our own transmission is driven into compression and
rtl-wmbus receives it with a few byte errors (CRC fails). Observed
@@ -134,15 +139,20 @@ directly - the decoded hex differed from the transmitted frame in 1-2
bytes, e.g. transmitted `...78534f4c5458...` received `...78514f4c5448...`,
with the rest of a 113-byte frame bit-exact.
Mitigations applied in firmware: SX1262 at its minimum configured power
(-9 dBm) and the lowest documented PA table (paDutyCycle 0x01, hpMax 0x00,
SX1262 datasheet table 13-21). Remaining mitigation is physical (antenna
separation or an inline attenuator) and is not scriptable from this host.
Firmware correction: the former “lowest documented PA table”
`(paDutyCycle=0x01, hpMax=0x00, deviceSel=0x00, paLut=0x01)` mixed the
PA-duty/hpMax values from the SX1261 +10 dBm row with SX1262 selection.
It is not a documented SX1262 row and cost about 32 dB. It has been
deleted. Current firmware uses documented SX1262
`(0x02,0x02,0x00,0x01)` with SetTxParams -9 dBm. Receiver compression must
be handled at the observer with fixed low gain, physical separation, or an
attenuator—not by misconfiguring the DUT PA.
Note: `rtl_sdr -g 0` selects *automatic* gain, not 0 dB. The lowest real
fixed gain on this tuner is 0.9 dB. At 0.9 and 7.7 dB fixed gain the
degraded dongle decoded nothing at all (neither meters nor our TX), so low
gain is not a usable workaround in that state.
fixed gain on this tuner is 0.9 dB. In the historical degraded-observer
test, 0.9 and 7.7 dB decoded neither meters nor the then-configured TX.
With fixed firmware, a 0.9 dB-gain smoke capture measured all 9 beacons at
median +27.81 dB over floor without clipping.
## Repeat-delay dependence of observer decoding
@@ -164,6 +174,8 @@ the firmware's own repeat delay measures 200.0 ms min/median/max.
1. Meter originals and Heltec RX match rate: rtl-wmbus decodes at high gain.
2. Heltec TX framing is decodable as C1: proven with synthetic bursts that
are not preceded by a weak frame (`hil/verify_tx.py`).
3. Real repeats: proven at RF level from the same stream's IQ by aligning
firmware TX timestamps to observed bursts (`hil/verify_repeats_iq.py`),
which is immune to both compression and observer sensitivity drift.
3. Real repeats: proven at RF level from the same stream's IQ by
`hil/meter_anchored.py` -- clock fitted on strong meter receptions
(never on TX events, which can borrow their original's burst), then
event-correlated FSK spectral excess against non-TX controls. Immune to
compression and observer drift; classifies strong/weak/not_observed.
+66 -14
View File
@@ -5,36 +5,88 @@ repeater. Known deviations:
| # | Experimental behavior | Standard requirement |
|---|---|---|
| 1 | Repeats the raw telegram bit-exact, including hop-related fields | EN 13757-5 repeaters must manage the hop counter / C-field semantics so telegrams are not repeated endlessly |
| 2 | Fixed development repeat delay 20-200 ms (default 50 ms) | OMS repeaters use randomized delays in the seconds range to avoid collisions |
| 3 | Repeats every CRC-valid C1 telegram (optional dev whitelist) | Real repeaters filter by installation list |
| 1 | Marks supported OMS uplinks with H=1 and repairs the affected CRC; unsupported layouts are skipped | Full repeater behavior still requires installation handling and timing compliance; see implementation notes below |
| 2 | Fixed repeat delay 20-1000 ms (persisted; 500 ms in current field testing) | OMS repeaters use randomized delays in the seconds range to avoid collisions |
| 3 | Repeats CRC-valid C1/T1 uplinks passing C-field, hop/access-bit and supported-header checks (optional dev whitelist) | Real repeaters filter by installation list |
| 4 | Rolling-hour duty-cycle guard, default 0.1 % (1000 ppm), suppresses repeats over budget and counts them (`tx_suppressed_duty`) | A production repeater needs proper duty-cycle management, or LBT+AFA, plus per-sub-band awareness |
| 5 | Duplicate suppression is a 2.5 s hash ring | Standard requires proper duplicate/loop protection semantics |
| 5 | H=1/R=1 frames are suppressed; a 2.5 s hash ring also suppresses duplicate originals | Full installation and duplicate-management policy still needs review |
| 6 | No AES, no decryption, payload treated as opaque | OMS security profiles apply to end devices; a transparent repeater may pass ciphertext, but production behavior needs review |
| 7 | T1 telegrams are detected (shared 0x543D sync) but dropped | A combined C/T repeater would need 3-of-6 decode/encode |
| 8 | Format B frames >128 bytes CRC-checked but rare path untested | -- |
| 7 | C1 and T1 both received and repeated in their native mode (3-of-6 codec in software, per-TX deviation swap). S-mode out of scope | A production repeater still needs reviewed mode policy (e.g. never re-air T1 as C1) |
| 8 | Format B >128 bytes and format A up to L=200 host-tested and verified on air (2026-09-02) | -- |
None of these shortcuts may be carried into the production Solenos/Gebos
gateway without explicit review.
Scope decision (2026-09-02): S-mode support is explicitly OUT OF SCOPE --
no S-mode traffic observed at either test site. T-mode receive/repeat and
long-frame support were implemented and RF-verified 2026-09-02. Site
finding from that verification: this site has NO ambient T-mode traffic
either (all ambient decodes are C1); the firmware's historical `rx_not_c1`
bucket is mode-byte-corrupted C1 plus false syncs, not T1 meters.
Persistence exception (2026-09-02): `duty_budget_ppm` and
`repeat_delay_ms` survive power cycles (NVS), because battery-standalone
testing needs settings that outlive reboots. **A persisted
`duty_budget_ppm` of 0 disables the duty guard indefinitely** -- the
device transmits over the ERC/REC 70-03 budget whenever traffic exceeds
it, until `set duty_budget_ppm 1000` is sent again. The OLED config page
shows the effective values (`DUTYCAP`, `DELAY`); boot logs `cfg_restore`.
Re-enable the guard after every guard-off test session.
## OMS hop marking (2026-09-09)
`wmbus_mark_repeated()` in `firmware/main/wmbus.c` operates on a CRC-validated
raw format A/B frame, before C1 framing or T1 3-of-6 encoding. It accepts
only C-fields `44` (SND_NR) and `46` (SND_IR), with H=0 and R=0:
- ELL CI `8C` or `8E`: set bit 4 of the Communication Control byte.
ELL takes precedence over any inner Transport Layer configuration.
- Without ELL: an explicitly recognized short/long uplink TPL header in
security mode 0 or 5 uses bit 0 of the Configuration field.
- Already-repeated/access-marked frames, truncated headers, and unsupported
layouts are left unchanged and not transmitted. Unsupported includes
headerless CI `78`, ELL `8D`/`8F`, and AFL without a supported outer ELL.
There is no bit-exact fallback.
Only H and its enclosing block's CRC change. The helper maps logical header
offsets across raw CRC boundaries; it neither decrypts nor copies the payload.
Frame length, meter identity, access number and ciphertext remain unchanged.
The original RX hash remains the RX/TX log correlation key; TX `raw` contains
the actual modified physical payload. Verbose mode reports policy rejection
as `tx_suppressed` with reason `hop_policy`.
H identifies a repeated copy, not a repeater ID, and does not enable multihop.
Synthetic `test_tx`, `test_tx_t1`, beacons and `test_burst` remain unchanged
PHY probes, not examples of received-frame repeater policy. Existing delay,
whitelist and duty-cycle settings are unchanged.
Reference: [OMS Volume 2, issue 5.0.1](https://oms-group.org/wp-content/uploads/2024/10/OMS-Spec_Vol2_Primary_v501_01.pdf),
sections 5.2.2.2, 5.2.5.2 and 7.2.4. Setting H alone does not establish full
OMS compliance. The new marking path has not yet been verified over RF.
## Duty cycle — measured, and why it matters
868.95 MHz falls in the 868.7-869.2 MHz sub-band, which ERC/REC 70-03
(annex h1.6) limits to 25 mW e.r.p. with **<= 0.1 % duty cycle**, unless the
device implements LBT+AFA.
Measured on this bench:
Absolute e.r.p. from the SX1262 plus KCT8103L and antenna has not been
calibrated. The default -9 dBm SetTxParams value is the minimum chip setting,
not proof of compliance with the 25 mW e.r.p. cap.
Current computed occupancy for the historical bench traffic:
```
TX airtime per repeat : 10.4-14.6 ms (median 13.4 ms), frame-length dependent
observed repeat rate : ~5-6 repeats/min when repeating every valid telegram
resulting duty cycle : ~0.11-0.13 %
RF occupancy per repeat : ~5.8-9.8 ms (8.8 ms for 102 radio bytes)
observed repeat rate : ~5-6 repeats/min when repeating every valid telegram
resulting duty cycle : roughly 0.07-0.09 % at the representative length
```
That is at or slightly above the 0.1 % ceiling — while repeating only the
CRC-valid C1 telegrams audible at a single indoor location. A production
repeater in a denser environment, or one that also handles T-mode, would
exceed the budget substantially.
An older report used 10.4-14.6 ms TX-operation latency (median 13.4 ms)
as if it were RF airtime and consequently reported 0.11-0.13 %. That was a
safe overestimate but not channel occupancy. Even the corrected site rate
is close to 0.1%; traffic variation, longer frames, a denser environment,
or adding T-mode can exceed the budget.
Consequences for the production gateway:
+231 -13
View File
@@ -13,8 +13,8 @@ capture -> correlation -> report (text + report.json). Exit 0 = PASS.
Useful flags:
- `--no-flash` reuse the already-flashed firmware
- `--duration N` capture window seconds
- `--tx-power N` configured SX1262 dBm (-9..14; FEM adds gain!)
- `--delay MS` repeat delay (20..200)
- `--tx-power N` configured SX1262 dBm (-9..+2; FEM adds gain!)
- `--delay MS` repeat delay (20..1000; note: run.sh sets it on the DUT, which persists)
- `--gain G` SDR tuner gain
- `--min-repeats N` PASS threshold
@@ -33,15 +33,25 @@ setup.log DUT configuration acks
```
./tools/discover_devices.sh -v # device inventory
./tools/flash.sh # build + flash
./tools/flash.sh # build + flash; no SDR required
./hil/capture_sdr.sh 60 /tmp/sdr.log # SDR-only capture
./result-py/bin/python3 hil/dut_cmd.py PORT --listen 10 status stats
./result-py/bin/python3 hil/iq_energy.py iq.raw # burst energy probe
./result-py/bin/python3 hil/iq_spectrum.py iq.raw # spectrum sanity
./result-py/bin/python3 hil/sdr_health.py # observer sensitivity
./firmware/test/run_tests.sh # frame-layer tests (ASan)
./firmware/test/run_tests.sh # frame + duty-ledger tests (sanitizers)
```
Hop-marker regression coverage in the host frame suite checks ELL precedence,
short/long TPL headers, H/R and security-mode rejection, truncated frames,
format A/B block boundaries, unchanged payload bytes and T1 encode/decode.
The 2026-09-09 host smoke also marked three captured meter frames and passed
them through `wmbusmeters`: ELL was reported as `relayed`, mode 5 as `hc=1`,
and the unencrypted water telegram still decoded its readings. Encrypted
captures were not decrypted (no meter keys); ciphertext preservation was
checked byte-for-byte. Target firmware build and HIL CLI smoke passed;
the new marking behavior has not been flashed or RF-verified.
## DUT serial commands
All output is single-line JSON; every line the HIL parses starts with `{`.
@@ -54,23 +64,199 @@ radio_status chip mode, command status, device errors, RSSI
radio_reset queue a full radio re-init
probe low-level scripted SPI/IRQ/TX diagnostic with raw dumps
test_tx transmit one synthetic C1 frame
test_tx_t1 transmit one synthetic T1 frame (3-of-6, 50 kHz dev)
test_burst transmit the same frame twice, repeat_delay_ms apart
inject standby fault injection: leave the radio out of RX
inject sleep fault injection: destroy the radio configuration
reboot
set repeat_delay_ms N 20..200
set tx_power N -9..14 (configured SX1262 level; FEM adds gain)
set repeat_delay_ms N 20..1000. PERSISTED across reboots (NVS)
set tx_power N -9..+2 (configured SX1262 level; FEM adds gain)
set repeat_enabled 0|1
set verbose 0|1
set fem_rx_csd 0|1 KCT8103L CSD level while in RX (LNA experiment)
set pa_low 0|1 1 = lowest documented PA table
set preamble_bits N 0|8|16|24|32 preamble detector length
set duty_budget_ppm N rolling-hour TX airtime cap; 1000 = 0.1 %, 0 = off
set duty_budget_ppm N rolling-hour TX airtime cap; 1000 = 0.1 %, 0 = off.
0..100000. PERSISTED across reboots (NVS), like delay.
Re-enable after guard-off tests!
set test_lfield N 19..200, length of synthetic test frames
set whitelist ID[,ID...] | set whitelist clear
set beacon_period_ms N 0 (off) or 250..600000; periodic synthetic frames
set whitelist ID[,ID...] | set whitelist clear (up to 8 unique eight-hex IDs)
```
The ten `set` commands, OLED toggles and web writes share validation and
revisioned configuration. Integers/booleans must be complete, not valid-looking
prefixes; malformed or duplicate IDs are rejected rather than truncated.
`status` retains its previous fields and adds all missing settings,
`revision`, `applied_revision` and the complete whitelist. Accepted values can
briefly precede application by the radio task. A TX-power or preamble change
requires radio reinitialization; a copied revision is not hardware/RF proof.
## On-device UI (OLED + PRG button)
Standalone debug without a host (docs/FUTURE_UI.md for design + pins):
- Pages (short press cycles, wraps): live status w/ duty bar -> health
counters -> last telegram -> effective config -> Wi-Fi credentials ->
Wi-Fi join QR.
- Hold ~1.5 s for the page action, each logged as
`{"event":"ui_action",...}`: status = toggle repeat, health = clear
stats, config = toggle beacon at 3000 ms (duty-safe at L=19), either
Wi-Fi page = enable/disable the debug access point.
- Display blanks after 3 min; any press wakes it without acting.
- `{"event":"ui_init","ok":...}` at boot reports display bring-up; a
missing/broken display never blocks the repeater.
- Config shown is the applied RAM snapshot. Delay and duty budget are restored
from NVS on boot; other settings reset. Unplugging USB with LiPo attached
does not reboot the device.
## Wi-Fi dashboard checks
Enable **WLAN-DIAGNOSE** using the OLED, join the displayed WPA2 network, and
open `http://192.168.4.1`. The following page provides a standard `WIFI:`
join QR (SSID and password, not the dashboard URL). Wi-Fi starts off on
reboot; its password persists in the separate `wifi_debug` NVS namespace.
On enable/disable, serial emits `wifi_debug` states: 0 off, 1 starting,
2 on, 3 stopping, 4 error, plus the ESP-IDF error code. Credentials are
not logged. Never change the test workstation's Wi-Fi without permission;
phone-based testing avoids disturbing its network.
The history API remains read-only: `GET /api/events?after=N`. Start at zero;
responses carry `boot_id`, cursor, retention/eviction metadata, cached
status and immutable final rows. A stale/ahead cursor causes `reset`.
Clients must fetch from zero on a new boot ID. The browser polls once per
second and keeps at most 256 rows. Pause freezes the event display, not
recording/polling. RX/repeat counters exclude synthetic events and survive
OLED/serial `clear_stats`, but not reboot.
### Writable settings API
Open **Einstellungen am Heltec-Gerät ändern**. Each field has its own draft,
accepted device value, persistence indication and explicit submit action.
The ten fields match the serial commands above; T1 synthetic sends still cap
the effective test L-field at 140. The browser warns before duty 0 or >1000
and before disabling the AP. CSV names remain browser-only.
- `GET /api/settings` returns `boot_id`, `revision`, `applied_revision`,
`csrf_token` and `settings` containing all ten keys.
- `POST /api/settings` accepts exactly `{boot_id,revision,key,value}`, one
setting at a time. Values are JSON integers, booleans or an array of up to
eight unique eight-hex-digit whitelist strings; an empty list allows all.
The response has the same snapshot shape.
- `POST /api/wifi/disable` accepts exactly `{boot_id}` and returns `{ok:true}`
before requesting an idempotent stop from the separate Wi-Fi control task.
A lost response leaves the client uncertain; it never retries automatically.
- Writes require `Content-Type: application/json` and `X-CSRF-Token` from GET.
Host must be `192.168.4.1` (optional port 80); foreign Origin/Fetch-Site values
are rejected. The random CSRF token rotates when the AP starts. WPA2 AP
access is the authorization boundary; there is no separate per-user login.
- Bodies are bounded to 512 bytes and shallow JSON, with duplicate/unknown
fields and embedded NULs rejected. Invalid inputs return 400, stale boot or
revision 409, origin/token failures 403, and busy/persistence failures 503.
No failed write is reported as saved.
- Pending settings are copied coherently by the radio owner, never written
into a half-processed reception from HTTP/OLED/USB. NVS runs under the writer
mutex outside the radio task; short spinlocks protect only memory copies.
Settings GETs while pending are capped at 1 Hz and 30 attempts, with no
overlap or blind POST retry. Conflicts preserve drafts and require review.
Static assets are gzip-compressed at build time by `tools/compress_web_asset.py`,
then served directly with `Content-Encoding: gzip`; no runtime compression or
partition-layout change is required.
Acceptance scenarios:
- See each reception's CRC/decode result and repeat completion, failure or
explicit skip. A TX timestamp is software-operation entry (including
driver setup), not an independently measured RF-start timestamp. A TX
outcome timestamp can represent an error, not successful completion.
- Load the downloaded CSV template or observed-ID export, preserving
leading zeros. Quoted commas/newlines are supported. A duplicate ID or
malformed CSV must leave all previous mappings unchanged. Refresh clears
mappings. Labels apply only to CRC-valid received identities.
- Check manufacturer-scoped type meanings (e.g. TCH), unknown raw codes,
version bytes and ambiguity warnings. Identity refers to the link-layer
sender, not necessarily a separately addressed transport-layer meter.
- Pause while new events arrive, reconnect after history eviction, and
reboot: no cross-boot or cross-reception joins are allowed.
- Compare radio timing/health with Wi-Fi off, on and actively serving
clients. Verify disable/re-enable and password persistence. Independently
observe RF if claiming unchanged sensitivity or downstream reception.
- Scan the OLED QR with a phone. Its integer-scaled, positive-polarity
modules retain at least a four-module quiet zone. On this small OLED,
camera focus can still be limiting; use text credentials if needed.
- On 320/390-pixel phone viewports, check that the current signals are easy
to find, controls are tappable and there is no horizontal page overflow.
Open **?** explanations by touch, including near the bottom of the page.
Close with the same button, outside tap, explicit close and Escape.
Explanations must remain usable while live rows update.
All authored visible UI/accessibility text is German; technical raw codes,
exact CSV/API headers and registered company names are intentionally kept.
2026-09-09 integration checks: test/debug builds and existing host suites
passed. A sanitized history smoke exercised 200,000 writes with concurrent
snapshots, overflow, cursor reset and synthetic/counter separation.
The actual browser assets were exercised on a localhost fixture server:
mobile layout, labels/filtering, atomic CSV rejection, downloaded template,
pause, ambiguity, eviction, disconnect/reconnect, reboot and refresh.
An independent ZBar decoder recovered the expected join payload from the
exact 128x64 framebuffer generated by the production QR renderer.
The German UI was checked at 320/390 px, including popup scrolling,
44 px help/button targets, local German CSV labels, outside/same-button/
explicit-close/Escape dismissal and focus restoration. A detail popup
remained open while the live history grew from five to six rows. Expanded
rows initially created a phantom fourth column on phones; the corrected
column spans were exercised across 320 -> 760 -> 390 px without reload.
Page/header widths stayed aligned and there was no horizontal overflow.
Browser checks used coordinate mouse input on narrow/mobile-emulated
viewports; native touchscreen automation timed out, so a physical phone
touch/QR check is still required. Evidence is retained locally under
`captures/2026-09-09-wifi-dashboard/`.
Pre-settings image sizes were test 927,360 bytes and debug 1,041,264 bytes.
The writable-settings build with compressed assets is now 862,672 bytes
(test, 18% app space free) and 978,656 bytes (debug, 7% free). The partition
layout and NVS locations were not changed.
After flashing the earlier, pre-settings German test build, USB telemetry at 59.74 s showed
six reception candidates, one CRC-valid reception, one CRC rejection, no
TX errors and no recoveries. Heartbeats reported RX mode and intact radio
configuration. This confirms live firmware operation, not phone access.
The test-profile firmware was flashed to the attached ESP32-S3. In the
pre-QR hardware smoke, the OLED action started the AP successfully; seven
CRC-valid receptions produced five completed repeats with no recoveries or
TX errors. Four repeats while the AP was on had software delays of
50.001–50.005 ms for a 50 ms setting. This small sample is not a controlled
RX-sensitivity comparison, RF delivery proof, or proof of active HTTP load.
Physical phone scanning and end-to-end HTTP checks require a phone report
or an explicitly authorized network connection; do not infer them from the
localhost fixture checks.
Writable-settings checks: 49 sanitized host scenarios ran the production
`settings.c`, `settings_http.c` and ESP-IDF cJSON parser through small host
HTTP/FreeRTOS/NVS adapters. They exercised all ten accepted settings,
non-mutating validation/security failures, two concurrent revision-CAS writers,
simulated storage failure/reboot and radio-copy acknowledgement. This is not
physical NVS or radio-handoff verification.
Chromium exercised all ten forms against that coordinator, preserved unrelated
drafts and edits made while a POST was delayed, rejected malformed input,
handled 409/503 without automatic writes, and gated dangerous duty changes and
Wi-Fi-off with native confirm/cancel dialogs. Confirmed Wi-Fi-off stopped
automatic polling while retaining CSV names. Layout/help were inspected at
320/390 px; setting inputs are 16 px and buttons at least 44 px high. All five
built gzip assets decompressed byte-for-byte to their sources and loaded in
Chromium. Existing frame, duty-ledger, long-T1 and font suites passed.
The settings firmware is built but **not yet flashed**: the Heltec USB device
disappeared before the flash command could open it; both serial discovery and
USB enumeration found it absent. Reconnect USB for the final on-device setting,
radio-handoff and reboot/persistence smoke. Workstation Wi-Fi was not changed.
Physical phone HTTP/QR checks remain separate. Local proof artifacts are in
`captures/2026-09-09-web-settings/`.
## Known pitfalls
- SDR PLL wedge -> gain-independent white noise; `usbreset 0bda:2838`.
@@ -96,6 +282,35 @@ Every run records `SDR sensitivity: full | degraded` from
runs report the same state -- this dongle loses ~17 dB within minutes of
streaming (docs/HARDWARE.md).
## Meter-anchored TX verification
When serial telemetry and raw IQ were captured together:
```
./result-py/bin/python3 hil/meter_anchored.py \
--iq iq.raw --serial serial.log --json
```
The verifier first fits the DUT/IQ clocks only from strong meter receptions,
which prevents a repeat from borrowing its own original burst. It then
compares the median spectrum at exact TX times with nearby non-TX control
windows. `rf_state` is:
- `strong`: event-correlated FSK is present and median event energy is at
least 6 dB over the capture floor;
- `weak`: event-correlated FSK is present below that energy threshold;
- `not_observed`: no event-correlated FSK signature;
- `no_events`: the serial log contained no in-range TX events.
`tx_above_6db` is only an energy count; it is not an RF verdict by itself.
`fsk_excess_db` is the in-band (20-80 kHz offset) event/control excess after
subtracting wideband (250-700 kHz offset) event/control change. At least
five analyzable TX/control pairs are required.
For unattended beacon runs, `hil/beacon_soak.sh` discards raw IQ only when
the result is `strong`. Weak, absent, or unclassified segments retain IQ
for reanalysis.
## Long reliability soak (DUT only, no SDR)
```
@@ -161,10 +376,13 @@ cat iq.raw | rtl_wmbus -f > decodes.log
./result-py/bin/python3 hil/verify_standalone.py --iq iq.raw --decodes decodes.log
```
Reports decoded repeats (same payload twice) and RF-level repeat pairs
(burst pairs of matching duration inside the delay window). The latter
matters because our transmission usually overloads the dongle and fails CRC
even though the RF is correct.
Reports decoded C1/T1 repeats (a supported H=0 original followed by its H=1
counterpart, with every other cooked byte identical) and RF-level repeat
pairs (burst pairs of matching duration inside the delay window). Decoded
pairs use occurrence order because SDR timestamps are quantized; each
original is matched once. Same-H duplicate decodes are not repeat proof.
The RF-level check matters because our transmission usually overloads the
dongle and fails CRC even though the RF is correct.
Note: only one process can hold the RTL-SDR. If another tool (e.g. an
oms-tui / wmbusmeters pipeline) is running, `rtl_sdr` fails with
+38 -28
View File
@@ -9,24 +9,24 @@ repeats, 214 RX candidates, repeat delay set to 200 ms).
| Quantity | min | median | max | n |
|---|---|---|---|---|
| RxDone -> reconstructed end of frame ("deaf tail") | 9 360 us | 12 240 us | 17 120 us | 214 |
| TX airtime (SetTx to TxDone) | 10 421 us | 13 422 us | 14 560 us | 72 |
| Programmed repeat delay (frame end -> TX start) | 200 001 us | 200 001 us | 200 008 us | 72 |
| TX operation latency (routine entry -> observed TxDone) | 10 421 us | 13 422 us | 14 560 us | 72 |
| Scheduled frame end -> TX-routine entry | 200 001 us | 200 001 us | 200 008 us | 72 |
| Repeated frame length (over-the-air bytes) | 63 | 100 | 113 | 72 |
Independent confirmation from the SDR side (raw IQ, same run): observed
burst duration differs from the computed airtime by 0.2-1.5 ms
(median 1.2 ms), and firmware TX events align to observed RF within
0-3.1 ms (median 2.4 ms).
(median 1.2 ms), and firmware TX-operation events align to observed RF
within 0-3.1 ms (median 2.4 ms).
## Derived
```
repeat delay accuracy : +1 us to +8 us against a 200 ms setting
(jitter well under 10 us)
RX deaf time per repeat : ~225.7 ms median
repeat scheduling accuracy : +1 us to +8 us against a 200 ms setting
(at TX-routine entry, not RF onset)
RX deaf time per repeat : ~225.7 ms historical median
= 12.2 ms deaf tail
+ 200.0 ms programmed delay
+ 13.4 ms TX airtime
+ 13.4 ms TX operation latency
```
The programmed delay dominates by design; it exists to make SDR
@@ -59,31 +59,41 @@ a production repeater.
artefact, not a design target (see docs/STANDARDS_GAPS.md).
- The fixed-255 tail is an SX1262 packet-engine limitation, so it applies
to any SX126x-based design. Budget for it, or implement early-stop RX.
- Repeat-delay accuracy of a few microseconds shows the ESP32-S3 side has
ample timing margin; timing precision will not be the constraint.
- TX-routine scheduling jitter is only a few microseconds. Actual RF starts
after packet setup and buffer transfer, so it must be measured at RF when
exact over-the-air delay matters.
## Airtime measurement vs true channel occupancy
## Computed airtime vs TX operation latency
`tx_airtime_us` and the duty-cycle ledger both use wall time from `SetTx`
to the moment `TxDone` is observed. That interval includes SPI round-trips
and up to ~1 ms of polling granularity, so it over-states true channel
occupancy:
Current `airtime_us`, cumulative `stats.tx_airtime_us`, and the duty ledger
all use computed channel occupancy:
```
computed airtime, 102-byte payload : (102 * 80 us) + 640 us = 8.8 ms
measured SetTx -> TxDone observed : 13.4 ms median
airtime = 48-bit preamble + 16-bit sync + payload at 100 kbps
= (radio_payload_bytes * 80 us) + 640 us
102-byte radio payload : 8.8 ms
26-byte beacon payload : 2.72 ms
```
Consequences, both deliberate:
`wmbus_tx.tx_operation_us` separately records the longer interval from
entry into the TX routine through observed TxDone. It includes standby,
packet-parameter and buffer SPI transactions, the actual transmission, and
polling/notification latency. Historical firmware mislabeled this interval
as `airtime_us`; its 102-byte median was 13.4 ms.
- The reported duty cycle is conservative (over-counted by roughly 50 %),
so the 0.1 % guard suppresses slightly earlier than strictly required.
For a compliance guard, erring high is the safe direction.
- Independent confirmation that the real occupancy matches the computed
airtime comes from the SDR side, not the firmware: observed burst
durations in raw IQ differ from the computed airtime by 0.2-1.5 ms
(median 1.2 ms), i.e. the transmission itself is ~9-10 ms for a 100-byte
frame as expected.
The distinction matters:
A production implementation should budget on computed airtime and use the
measured value only for latency diagnostics.
- Computed occupancy is independent of task scheduling and SPI speed, so
the duty guard no longer consumes roughly 50% extra budget.
- Operation latency remains useful for diagnosing stalls and TxDone timing,
but is not a regulatory airtime measurement.
- Raw-IQ burst duration remains the independent check. In the historical
102-byte run it differed from computed airtime by 0.2-1.5 ms (median
1.2 ms).
The rolling-hour ledger uses 61 keyed one-minute buckets rather than a
fixed event ring. It therefore cannot forget high-rate events merely
because an event count was exceeded. The oldest overlapping minute is
counted in full, making the result conservatively high by at most one
minute's included traffic at the window boundary.