[snokvist]

What

A new self-contained subproject p2plink/ that builds against the sibling devourer tree: a standalone 1:1 radio layer (drone RTL8812EU → ground 8812EU and/or MT7612U ears) carrying an H.265 stream from the waybeam encoder frame ring plus an IP tunnel. Range and link robustness are the objective, not throughput.

Nothing outside p2plink/ is touched. The branch sits on upstream OpenIPC/devourer master f46c28a (the fork was fast-forwarded first).

Mechanism

• addr1 is the ACK policy. Video and reports go to a group DA (never ACKed, never retried); the 100 ms probe and in-range IP go unicast to the ground ACK target (hardware ACK, retry limit 3). On Realtek the radiotap NOACK bit is ignored and BMC derives from addr1, so the DA switch is the only per-frame lever.
• The probe's hardware ACK is the primary uplink sensor; the 10 Hz LinkReport is secondary. ACKs OK + no reports = HOLD the rung (ground hears us, its reports don't reach us); no ACKs + no reports = descend to the floor.
• Ladder of 7 rungs, 20/40 MHz frames only, ground parked at 40 MHz, drone picks width per frame (a 40-parked receiver decodes a 20 MHz frame on its primary with no retune). No 5/10 MHz anywhere (the MT7612U ear cannot decode it).
• Blind descent: report timeout 500 ms → encoder budget + FEC drop two rungs immediately, MCS follows 0.5 s later, floor within ~2.5 s. Promotion needs 5 s dwell + loss/RSSI/probe-rung evidence.
• N-ear merge: first copy wins per (stream, block, symbol); separate emit/retire cursors; silent-ear exclusion; Reed-Solomon erasure FEC (zfec-compatible).

Contents

piece	files	tested
wire format + LinkReport/StatsSnapshot	include/p2p/wire.h, report.h	wire_test
RS erasure FEC	fec.h, src/fec.cpp	fec_test (every C(12,4)/C(11,3) erasure pattern)
N-ear merge	rx_merge.h	merge_test (132 checks)
link controller + DA machine + scheduler	link_ctl.h, budget.h	link_ctl_test, budget_test (186 checks)
encoder frame ring (waybeam-compatible), tun, stats shm/NDJSON	frame_shm.*, tun.*, stats_pub.*	frame_shm_test
devourer glue	radio.h, src/radio.cpp	never run on hardware
apps	src/p2p_tx.cpp, src/p2p_rx.cpp, src/p2p_feed.cpp	--help/error paths only
docs	docs/DESIGN.md (spec as built), docs/HANDOVER.md (bench cells C1–C9), README.md	—

Build: cmake -S p2plink -B build-p2p -DCMAKE_BUILD_TYPE=Release -DP2P_WITH_MT7612U=ON && cmake --build build-p2p -j && ctest --test-dir build-p2p -L p2p — warning-free, 6/6 suites pass.

Not yet verified (why this is a draft)

All of radio.cpp and both apps have never touched an adapter. docs/HANDOVER.md lists the six design assumptions the bench must confirm first: 8812EU as the soliciting transmitter (only ever measured as responder), 8812EU→MT7612U ACK, 20 MHz frames into ears parked at 40 MHz on these dies, the per-unit wide_tx_igi_floor edge and its unmeasured uplink cost, MT7612U 40 MHz receive, frame-length PER at the floor. Review and hardware trials happen in a local session against this branch; fixes land as further commits here.

Known gaps carried as follow-ups

Frame size fixed at 1400 B (not a rung); ground IP broadcast-only; no encryption; no channel migration; the "no probe ACK for 30 s → cap duty" rule; encoder floor (resolution/fps) co-design.

---

Generated by Claude Code

[snokvist]

Milestone: review + first light (hardware, 2026-10-29). Bench has one 8812EU, so every result is 8812EU drone (1-1, USB2) → MT7612U ground (8-1). C1 and C3's 8812EU ear are deferred until the second 8812EU is available.

Review fixes (one commit each):

commit	defect
fe3b003	open through claim_interface_reset_reopen (a re-enumerating reset left the handle stale)
457af4b	stop() could hang forever: StartRxLoop clears a StopRxLoop that lands before it
f66b2e0	TX-report tag join: stale probe slot could claim a unicast frame's report after the 8-bit wrap; slot now written before send_packet
00192cf	merger rule (3) dated from the first symbol → every IDR block lost; now a quiet dwell from the last symbol (+ test, DESIGN.md)
1c09184	--txpwr-offset-qdb on both apps; DEVOURER_TX_PWR was inert (the library reads no env)
c5ce1d6	ground spun at 100 % CPU with --tun and --uplink none; tx_timeouts count
70814cf	--fixed-rung drone never fed reports → FLOOR, no Ucast
8adb0a1	default --log-level warn

First light, ch 100 HT40+, R2, −10 dB power offset, near field (−37 dBm), run 2:

criterion	result	pass
uniq vs video frames sent	434/s vs 436/s; 760/760 blocks (25 via FEC)	PASS (single ear, saturation not excludable: MT7612U has no EVM)
fec_fail‰	0	PASS
ack8 / retries / report coverage	8 / 0.00 / 250 of 250	PASS (ON arm only)
report_age_ms	19–21, TRACK	PASS
p2p.da	Ucast at +1.5 s	PASS
tx_timeouts / tx_failed	0 / 0	PASS (25 s)
ping via tun	39/40 fwd, 19/20 rev	PASS (1 loss each way unexplained)
diversity	n/a, one ear	NOT RUN

Upstream note (not changed here): UsbTransport logs every successful bulk-OUT at info (bulk_send EP n OK), one stderr line + flush per frame. That line probably belongs at trace.

[snokvist]

Milestone: C1–C4 (+ C5/C6 run, 2026-10-29). Pairing as before: 8812EU drone (1-1) → MT7612U ground (8-1), ch 100 HT40+, near field (−36 dBm, no attenuator). Details + log names: HANDOVER ## Results.

cell	assumption	result	verdict
C1	A1 (8812EU responder)	needs a second 8812EU	NOT RUN
C2	A2 8812EU → MT7612U ACK	ON 223/223 ok, retries 0.01, coverage 1.00; OFF 283/283 pinned at 3; group decode 435/s unchanged	PASS (one unit pair)
C3	A5 MT7612U at 40, 20-in-40	R3 99.8 %, R5 99.7 % of submitted, 0 failed blocks	PASS (MT7612U ear only; 8812EU ear NOT RUN)
C4	A4 floor edge	unit 1-1 clears 3/3 soft starts at floor 0	edge 0 (soft re-init, not VBUS cold; one unit)
C4	A4 floor cost	needs attenuation (~50 dB margin here)	NOT RUN
C5 (b)	blind descent	R5→FLOOR on the 500 ms grid, +2.05 s after kill	PASS
C5 (c)	restore, no snap-back	TRACK +1.06 s after restart, R0 kept	PASS (needed afdb65c, the session-nonce fix)
C5 (a)	HOLD with ACK alive	HOLD at report timeout, no R−2 step first	CONTRADICTS HANDOVER (code/test intend it)
C5	promotion	never fires: no R5→R6 in 20 s clean TRACK; parked at R0 after restore	FINDING
C6	DA machine	Bcast +0.31 s; duty unchanged before the PHY step; 0 IP drops	PASS

Fixes since first light, from an independent review: afdb65c (peer restart = new session; byte 11 nonce), bdc3b56 (40 MHz frame on a 20 MHz tune race + retune readback), 14319ad (spectator IP drops, stoi throw, ClearAckResponder), plus a SIGUSR1 report-mute bench hook for C5(a).

Open design questions (no code changed for these):

1. Promotion is dead. probe_evidence_ok vetoes unless the ground saw frames at r+1, and nothing is ever sent at r+1, because the probe is at the current rung. Proposal: while in TRACK with the good-report streak and dwell met, air each video block's repair symbols at r+1. A lost repair costs nothing while the sources arrive, and the ground's per-rung counters then measure r+1 PER in-band with no extra airtime. R3→R4 crosses 20→40 MHz on a 20 MHz tune; R4 has R3's PHY rate, so that step could take rssi + loss evidence alone. Alternative: drop the probe-rung condition and rely on loss < 20 ‰ + rssi margin + 5 s dwell + 30 s lock.
2. HOLD entry. Proposal: keep the code (HOLD at the report timeout, no step) and fix HANDOVER/DESIGN. ACKs prove the forward link is fine, and descending the drone's rate does nothing for the lost reverse path.

[snokvist]

Milestone: C5–C9 (2026-10-29), same pairing: 8812EU drone (1-1) → MT7612U ground (8-1), ch 100 HT40+, near field (−36 dBm, 50 cm, no attenuator). Detail, logs and counterparts: HANDOVER ## Results.

cell	result	verdict
C5 re-run: promotion	three climbs R0→R6 in one session, one rung per 5 s dwell, 0 failed sends	PASS (fixed by 13db3fc + 435c414)
C5 re-run: 20↔40 retunes	frames continued through every R3→R4 retune, no empty 100 ms in 240 s	PASS (the mid-session IGI-floor path, never exercised on air before)
C5 (a) HOLD	HOLD at the report timeout, no step: now the documented design	PASS (docs b386b0f)
C6	Bcast +0.31 s, no duty storm, 0 IP drops	PASS (earlier)
C7 R5 / R6, 30 min each	longest flat tx_submitted 0 ms; 0 failed / 0 timeouts; q_video ≤ 7 / 12 ms; 53 / 167 failed blocks of ~215k / ~431k	PASS. Probe report coverage 0.88 / 0.76 (CCX path saturates with frame rate)
C8	static R0 1400 B baseline: fec_fail 0, 0 failed blocks, loss p99 41 ‰	NOT RUN as designed (no size knob, no motion); baseline recorded
C9	actuator ordering on 22 re-targets: budget + command 500 ms before each demote PHY step, after each promote	PARTIAL (no encoder on the bench)

Fixed along the way:

• c8d8408: the video budget now comes from the scheduler's own airtime model. PHY × duty ignored the 86 µs that every frame pays, so C7 ran at 707‰ against a cap of 500 and 822‰ against 450. With the new budget, R5 runs at 495‰ against its cap of 500.
• 435c414: HT40 rates corrected to 27/54/108 Mb/s. Confirmed on air: R4/R3 frames/s = 1.032 at saturation.
• 86594f5: a repair floor of 2 per block, as in waybeam-link's min_r. Verified on air at 5.66 symbols per block, against 5.85 predicted.
• MT7612U restart stall (3d7c081 → f63d895, devourer core):
  • Symptom: a ground restart under drone traffic took 27–34 s in 22 of 32 tries.
  • Cause: frames reach EP 4 during RXDCOC with MAC RX off, and that holds back the MCU's EP 5 replies.
  • Fix: drain EP 4 on a late MCU reply. Restarts are now 0.9–1.3 s in 31 of 31.
  • Upstream: mt7612u: drain EP 4 when an MCU reply is late, so RXDCOC cannot stall bring-up OpenIPC/devourer#510, with a reproducer and an A/B on upstream code (master 6/6 + 6/6 stalled, fix 10/10 + 10/10 at 1.18 s).
  • It is outside p2plink/ because there is no seam on the p2plink side. The drone-side "no ACK 30 s → duty 0.3" rule is not needed for this.

8812EU HT40 TX (the libc0607/rtl88x2eu#7 stall):

• Not reproduced in devourer's TX path. 0 sagging bring-ups in 28 runs without a ground: txdemo across base, CCA off, 40-in-80, TX reports, and RX thread, plus p2p_tx alone.
• One unexplained sag. p2p_tx at saturation, −20 qdB, with the MT7612U ground 50 cm away sagged in 3 of 8 bring-ups, never under budget load. It is recorded for a re-check at real separation.
• Model calibration. The model's 50 µs inter-frame space is an upper bound on this unit: measured 30–53 µs at saturation.

Open:

• C1, C3's 8812EU ear, and the C4 edge need the second 8812EU.
• C4's cost, C9's demote half, and C8's sweep need attenuation, an encoder, or a size knob.
• R3→R4 lowers the budget by 4.9 %, under the actuator's 5 % threshold. The ladder or the threshold is your call.

[snokvist]

Addendum — second 8812EU (the ssc338q camera as drone), 2026-10-29 evening. This closes cells the two milestone comments left NOT RUN or PARTIAL.

• Drone: the camera's own RTL8812EU, running p2p_tx cross-built for armv7 from /tmp, with waybeam-hub stopped during the cells and restarted afterwards.
• Ground: the bench host.
• Conditions: ch 100 HT40+, near field (−23 to −45 dBm). Details are in HANDOVER ## Results.

cell	result	verdict
C1: 8812EU drone → 8812EU ACK target	ON: ack8 8, retries 0.01, report coverage 1.00. OFF: ack8 0, 203/203 reports pinned at 3	PASS
C3: two ears, R3 (20 MHz) and R5 (40 MHz)	8812EU ear 99.4 / 99.5 %, MT7612U ear 99.9 / 99.9 % of frames sent; adapters = 2 throughout; 1 / 0 failed blocks	PASS
C4 edge: second unit	all 21 starts aired at once at every --wide-igi-floor from 0x35 down to 0	edge 0 on both units (soft starts)
C9: encoder in the loop	the camera's venc (720p100) through its frame ring, re-targeted live by --bitrate-cmd. The encoder settles within ~0.5 s of each command, so the 500 ms lead holds. q_video max 9 ms, 0 drops, 0 failed sends, 0 mid-run FEC failures	PASS on ordering and the bound. The loss-driven half still needs attenuation

Also:

• R4's duty cap is raised to 0.60 (6399431), so R3→R4 no longer lowers the budget.
• The Jaguar2 read-fail side issue was handled in a separate session. Fix branch: jaguar2-read-fail-survive.

Still open:

• C4's cost, and C9's loss-driven demote. Both need an attenuator.
• C8's frame-size sweep. Needs a size knob and motion.

[snokvist]

p2plink development has moved to its own repository: snokvist/waybeam-p2p. It is private, and registered in waybeam-coordination (snokvist/waybeam-coordination#217). This PR stays open as the record of the hardware bring-up. New p2plink work goes to the new repo, not to this branch.

What moved

• History: all 39 commits under p2plink/ (split with git subtree split, now at the repo root), up to 925e878 on this branch.
• Devourer: vendored under third_party/devourer at OpenIPC d0a9e10, with one pending-upstream delta, mt7612u: drain EP 4 when an MCU reply is late, so RXDCOC cannot stall bring-up OpenIPC/devourer#510 (the MT7612U EP 4 drain from f63d895 here), recorded in third_party/README.md.

What did not move: the devourer-side commits on this branch (3d7c081, f63d895). They go upstream through OpenIPC#510 instead.

Last milestone on this branch (also in the new repo):

• 28e8790: p2p_tx --frame-bytes, the frame-size knob C8 needs.
• 99888e4: the drone no longer charges airtime for a video frame dropped as oversize. On the camera at R6 the duty meter read 1068‰ before and 551‰ after, against 554‰ for the frames actually aired.
• 925e878: the ground's block merger now rebases on the sender's session nonce.
  • Before, a drone restart under a running ground discarded video for up to 1024 blocks. The first untethered C8 walk hit this on all 18 legs: two ears received ~120 000 frames each, and only leg 1 decoded.
  • A new merge test fails on the old code. On hardware, three restarts each decoded from their first symbol.
  • The walk is void and needs repeating.

🤖 Generated with Claude Code

https://claude.ai/code/session_011dReg1L6xtBhf95NpAeuge