Correct Morse human spacing and harden speed qualification

docs/test-results/morse_rx_hil_2026-10-12.md

@@ -166,3 +166,195 @@ remain in progress. Default 25 WPM, public ceiling 60 WPM, radio bandwidth,
 and shipping RF filtering have not been changed. Packet control stays at
 909.5 MHz / 500 kHz / SF5 / CR5; CW remains separate at 902.100 MHz by default.
 Raw results remain in `out/morse-rx/` and the Pi run directory above.
+
+## Controlled speed and human-timing follow-up
+
+These are **conditional laboratory results**, not a promise of error-free
+reception on every board or with automatic noise detection. CW stayed at
+902.100 MHz, packet control at 909.5 MHz / 500 kHz / SF5 / CR5, and requested
+chip TX power at -9 dBm. The Pocket / RAK pair used a volatile **HIL-only
+-80 dBm receive floor** against strong received tones. Shipping images do not
+contain this receive-floor setting. Earlier automatic-threshold quiet controls
+produced false characters on the Pocket and T096; that remains a separate
+qualification limitation.
+
+The original 2.25 / 5.5-dot letter / word boundaries failed ordinary uneven
+timing even at 25 WPM. With seeded +/-20% phase jitter, gradual -10% duration
+drift, and **no extra letter or word spacing**, the first sweep passed only
+2/6 messages at each of 25, 60 and 120 WPM. A short, fully retained RF trace
+decoded `VW MN DOG` as `VW MNDOG` in all six directions/rate combinations.
+All 20 expected marks remained present; the second word gap measured
+246 / 246 ms at 25 WPM, 100 / 102 ms at 60, and 54 / 52 ms at 120. The old
+word thresholds were 264, 110 and 55 ms respectively.
+
+The receiver now uses the ordinary **2 / 5-dot gap midpoints**, without changing
+mark classification, RX allocation, RF filtering or the public speed limit.
+The same tight human-timing profile (seed 6789), with two repetitions of all
+three mixed alphabet/digit, word and punctuation messages in both directions,
+then passed strictly:
+
+| Requested WPM | Pocket to RAK | RAK to Pocket | Total exact messages |
+| --- | --- | --- | --- |
+| 25 | 6/6 | 6/6 | 12/12 |
+| 60 | 6/6 | 6/6 | 12/12 |
+| 120 | 6/6 | 6/6 | 12/12 |
+
+Only leading/trailing capture whitespace is removed in the physical harness;
+extra internal spaces, noise characters, missing letters and `?` fail the test.
+There were no cleanup errors. This establishes zero observed errors in those
+36 messages under the stated laboratory conditions, not universal 100%
+reliability or a 120-WPM release qualification.
+
+The production-equivalent RF comparison uses cold-XOSC TX, 3400-us PA ramp,
+RSSI window 0, two-ms sampling and ordinary OFF debounce. Compile guards and
+actual PlatformIO/SCons flag tests prevent faster experimental settings or a
+different receiver clock from being mislabeled as shipping RF behavior.
+The lab maximum override and diagnostics still make these HIL applications,
+not release images.
+
+Requested WPM is not measured RF pace. Four strict 13-word `PARIS` trials per
+rate transmitted 643 dot units, omitting the unsent final seven-unit word gap:
+
+| Requested WPM | Host-observed duration | PARIS-equivalent pace |
+| --- | --- | --- |
+| 60 | 14.354-14.365 seconds | 53.7 WPM |
+| 120 | 7.937-7.940 seconds | 97.2 WPM |
+
+The pace includes CLI response and polling overhead; it is not an oscilloscope
+measurement of RF edges. It also does not substitute for varied-text testing.
+
+Latest focused verification passed 129/129 test methods. The actual HIL
+scheduler / decoder matrix has 175 exact results plus one explicitly
+unsupported 240-WPM / two-ms-sampling / 50%-extended-spacing case. The separate
+human envelope matrix has 84 exact results plus 14 explicitly unsupported
+doubled-character-spacing cases. Unsupported transcripts remain checked and
+visible; they are not normalized or counted as exact-decoding passes.
+Former-threshold negative controls reproduce the tight-gap failure. The
+production receiver remains 4168 bytes, and compact STM32 still links at
+228868/229376 flash bytes with its unchanged feature boundary.
+
+Evidence filenames in the directories above:
+
+- `squelch80-production-rf-human-tight-20261013.json` (before gap correction)
+- `production-tight-gap-trace-20261013.json` (complete short RF traces)
+- `production-tight-gap-fixed-repeat-20261013.json` (36/36 corrected messages)
+- `production-pace-baseline-20261013.json` (host-inclusive PARIS pace)
+
+The tested application-only gap-correction packages were:
+
+| Package | SHA-256 |
+| --- | --- |
+| `RAK_4631_companion_radio_usb-prod-rf-gap-20261013.zip` | `26cc93db4f6a602e24a35828cca79efaef002b13b18192fe763eb5426515ec6b` |
+| `Mesh_pocket_companion_radio_usb-prod-rf-gap-20261013.zip` | `9d1f76c1cc42e1d30ca5daa3d30fe30232d426d72b80278cac67d2d98c21a178` |
+
+The T096's earlier image and measurements do not qualify its speed under this
+new gap correction. No LR1110 or SX127x physical speed limit is established by
+the SX1262 pair. Default 25 WPM and the public 60-WPM ceiling remain unchanged.
+
+The next faster **production-RF** timing buckets did not meet the zero-error
+criterion, even with the same HIL receive floor:
+
+| Requested WPM | Standard timing | Tight human timing, two repeats |
+| --- | --- | --- |
+| 130 | 5/6 | 11/12 |
+| 150 | 2/6 | 6/12 |
+| 170 | 0/6 | Not repeated in this corrected sweep |
+| 200 | 0/6 | Not repeated in this corrected sweep |
+
+At 130 and 150 some failures affected the initial short mark; 150 also had
+later body errors. The first RF mark currently begins before the command
+reply and the remaining foreground-loop work. That is a possible delay window,
+not a measured explanation of all these errors. No unproved transmitter or
+mark-classifier adjustment was made to improve the reported score.
+
+Additional boundary evidence:
+
+- `production-gap-fixed-standard-boundary-20261013.json`
+- `production-gap-fixed-human-boundary-20261013.json`
+
+A second human-style pattern, using +/-20% jitter, gradual **+10% duration
+drift** (slowing down), no extra gaps and seed 12345, passed another 6/6 mixed
+messages at each of 25, 60 and 120 WPM in both directions. Across the corrected
+speeding-up and slowing-down profiles, that is 18/18 messages at each rate,
+54/54 overall, under the same controlled receive threshold.
+
+The subsequent 30-second **automatic-threshold**, no-intentional-TX control
+was not clean overall: the RAK captured `E E `, while the Pocket was empty.
+This is evidence against an unconditional production zero-error claim; it is
+not a diagnosis of the source of the quiet RSSI excursions. The current
+shipping speed ceiling remains unchanged.
+
+- `production-gap-fixed-human-slowing-20261013.json`
+- `production-gap-fixed-auto-quiet-20261013.json`
+
+Short, fully retained two-second initial-window traces sharpen the initial-mark
+finding. At 130 WPM, the failed first `E` measured 18 ms, exactly twice the
+nominal nine-ms unit; a successful first `E` measured 16 ms. At 150, the
+corresponding failed/successful widths were 16/14 ms against an eight-ms unit.
+Those widths were also common among later dots, so these recordings do **not**
+show an unusually long first mark. The later learned units were 11/10 ms,
+making comparable later dots classify correctly. Two-ms sample phase and the
+initial untrained threshold can therefore matter near this boundary. The
+trace has no transmitter deadline timestamps and does not prove that CLI work
+caused the errors. These partial diagnostic captures are not complete-message
+qualification failures or passes.
+
+- `production-initial-mark-trace-20261013.json`
+
+## Faster experimental RF comparison
+
+The corrected fast SX1262 HIL applications use warm-XOSC TX, an 800-us PA
+ramp, RSSI window 2, one-ms sampling and the experimental short OFF-confirmation
+minimum. These are **not shipping RF settings**. The same -80 dBm HIL receive
+floor and exact-message checks were used on the same Pocket / RAK pair.
+
+| Requested WPM | Standard timing, two repeats | Tight human timing |
+| --- | --- | --- |
+| 120 | Previously all-pass; not repeated in this boundary sweep | 6/6 |
+| 150 | Not tested in this boundary sweep | 5/6 |
+| 200 | 12/12 | 3/6 |
+| 255 | 10/12 | Not tested in this boundary sweep |
+
+The tight human profile is again +/-20% jitter, -10% duration drift, no
+added letter/word spacing and seed 6789. At 150, `DOG` became `DKG` in one
+direction. At 200, three messages had real body or punctuation errors.
+The shortest timing bucket therefore does not qualify as zero-error, and
+neither does the higher uneven-keying speed. Integer rounding maps requested
+219-255 WPM to the same five-ms dot, so 255 is not 255 measured WPM.
+
+The fast steady-PARIS pace check gives:
+
+| Requested WPM | Host-observed duration | PARIS-equivalent pace | Exact PARIS messages |
+| --- | --- | --- | --- |
+| 120 | 6.868 seconds | 112.3 WPM | 2/2 |
+| 200 | 4.273-4.277 seconds | 180.5 WPM | 2/2 |
+| 255 | 3.669 seconds | 210.3 WPM | 1/2 |
+
+The failed 255 PARIS capture contained `PANIS` in place of one `PARIS`.
+These rates include the same host/CLI overhead as the production-equivalent
+pace measurement; pace is calculated only for the 643-unit PARIS payload,
+not by applying that reference count to the other test messages.
+
+The strongest observed all-pass points are therefore **200 requested WPM
+for machine timing on the fast HIL pair**, and **120 requested WPM for the
+tested uneven timing**. Their steady-PARIS references are approximately
+181 and 112 WPM respectively. With production-equivalent RF timing, the
+tested uneven-timing reference is approximately 97 WPM. None of these
+conditional limits resolves the automatic-threshold quiet-control failure
+or qualifies another board/radio family. Shipping defaults remain unchanged.
+
+- `fast-gap-fixed-standard-max-20261013.json`
+- `fast-gap-fixed-human-boundary-20261013.json`
+- `fast-gap-fixed-pace-20261013.json`
+
+| Fast HIL package | SHA-256 |
+| --- | --- |
+| `RAK_4631_companion_radio_usb-fast-gap-20261013.zip` | `7722ca7a0008c660217535c19e9dd17d5923fc90510f9850b2ce188a6d753951` |
+| `Mesh_pocket_companion_radio_usb-fast-gap-20261013.zip` | `8d153c544f19e01d45e178105b704840f855a53cb8fda2044ba102cb19b85521` |
+
+Final cleanup confirmed both selected radios had Morse stopped, configured
+25 WPM / 902.100 MHz, default HIL timing and automatic squelch restored.
+Their saved primary settings still held an earlier 919.5-MHz lab tuple; those
+two test devices were explicitly changed to **909.5 / 500 / SF5 / CR5**, and
+that tuple was also confirmed live. No service radio was changed. The pair
+remains on the identified fast HIL applications, not shipping images.