[josephnef]

Part of #260 (epic #254). Follows #558 (sequential uplink visits): the downlink side across channels.

Why

Before the switch, the UE takes DL only from its serving cell. With the source dead, it had no DL from the kill to the switch (1.4–1.9 s), although the target held every PDU since the command (DAPS DL).

What (examples/cell)

• The answer: when the target DU hears a visit copy from a UE it staged for an inter-frequency arrival, before the switch, it answers at once with the queued DL PDUs no answer has offered yet (up to 8).
  • The frame (build_visit_dl_frame, cell.h) is broadcast-addressed: no ACK, no CW escalation.
  • It is tagged with the target cell's key, and the body names the UE's target identity.
• The UE: while still holding the target's channel, it takes the PDUs into its DL stream (parse_visit_dl_frame; cell.ue.visit_dl).
• After the switch: the queue keeps everything (nothing ACKs an answer), and the target sends not-yet-offered PDUs first, then the offered ones. Without that ordering, the newest PDU waited behind ~5 frames the UE already had (0.53–0.63 s).
• Sending: answers go out from a dedicated thread. Sending from inside the RX callback froze the DU on the rig (the RX loop recycles the TX pages).
• Only across channels: the target's neighbour list must put the source on another channel. On one channel the target hears every uplink, and an answer would land in the source cell's schedule.
• Selftests: the visit DL frame (broadcast RA, target identity, key, not a unicast DL frame, PDU sn under the tag). ctest 95/95.

Rig (testrig scripts/m6_daps_xch.sh, 2 runs per case, all pass)

longest DL silence at the UE	runs
source killed at the accept, visits answered	205, 205 ms
source killed, no visits	1395, 1498 ms
no fault, visits answered	308, 205 ms

• Answers: 19–21 per handover; DL DAPS lost no PDU.
• Uplink: unchanged (0 lost before the switch with visits).
• What remains: the switch's own pause (~2 superframes). That is next.
• Regression: same-channel DL DAPS (no answers sent, silence unchanged), the three-cell crossing with a source failure, plain and fault handovers.

🤖 Generated with Claude Code

[qodo-free-for-open-source-projects]

PR Summary by Qodo

Deliver target downlink PDUs during inter-frequency DAPS visits

✨ Enhancement 🧪 Tests 📝 Documentation 🕐 20-40 Minutes

AI Description

• Target DUs answer pre-switch, inter-frequency visits with queued downlink PDUs, reducing silence
 after source failure.
• UEs accept authenticated answers; targets retain offered PDUs and prioritize unoffered ones after
 the switch.
• A dedicated sender avoids RX callback stalls; selftests and rig results document the behavior.

Diagram

sequenceDiagram
    actor UE
    participant DU as Target DU
    participant Q as DL Queue
    participant Sender as Answer Thread
    participant Stream as UE DL Stream
    UE->>DU: Pre-switch visit copy
    DU->>Q: Select unoffered PDUs
    Q-->>DU: Up to eight PDUs
    DU->>Sender: Queue broadcast answer
    Sender-->>UE: Authenticated DL answer
    UE->>Stream: Accept and deduplicate
    Note over DU,Q: Retain offered PDUs
    DU->>Q: Prioritize unoffered after switch
    Q-->>UE: Scheduled DL frames

Loading

High-Level Assessment

The visit-answer approach fits the brief pre-switch target-channel window. Using a broadcast-addressed frame avoids ACK handling, while a separate sender avoids the observed RX callback stall. Direct callback transmission and ACK-dependent unicast are less suitable here.

Files changed (4) +266 / -9

Enhancement (2) +188 / -6

cell.hDefine authenticated broadcast visit DL frames +31/-0

Define authenticated broadcast visit DL frames

• Adds a builder and parser for broadcast-addressed target DL answers carrying the target UE identity and up to eight PDUs. Parsing reuses DL frame authentication while rejecting ordinary unicast DL frames.

examples/cell/cell.h

main.cppAnswer staged visits and deliver their downlink PDUs +157/-6

Answer staged visits and deliver their downlink PDUs

• The target identifies eligible inter-frequency visits, selects unoffered queued PDUs, and transmits answers through a dedicated thread. The UE accepts authenticated answers into its DL stream; after switching, the target sends unoffered PDUs before retained offered ones.

examples/cell/main.cpp

Tests (1) +16 / -0

cell_node_selftest.cppTest visit-answer framing and authentication +16/-0

Test visit-answer framing and authentication

• Checks broadcast addressing, target identity, epoch, and PDU sequence numbers. Verifies rejection with the wrong key or cell, rejection as an ordinary unicast frame, and detection of PDU tampering.

tests/cell_node_selftest.cpp

Documentation (1) +62 / -3

cell-m6.mdDocument inter-frequency downlink visit answers +62/-3

Document inter-frequency downlink visit answers

• Explains when the target answers visits, why it retains and reorders offered PDUs, and why answers use a separate sender. Records rig measurements, regressions, and the remaining switch pause.

docs/cell-m6.md

[qodo-free-for-open-source-projects]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0) 🎨 UX issues (0) 🔗 Cross-repo conflicts (0) 📜 Skill insights (0)

1. Rig KPIs rest on two-run maxima ✓ Resolved

Description

The new rig table reports only the longest DL silence per case across two runs (for example `205,
205 and 1395, 1498`), with no distribution, percentiles or spread for handover interruption or DL
delivery. Two samples cannot show a distribution, so a reader cannot judge how much of the gap
between cases is run-to-run variation.

---

2. Rig results hide forced handovers ✓ Resolved

Description

The rig-result paragraph says the A3 hands UE1 over 1 → 2 and the source is killed at the accept,
but never states whether the handover trigger was forced on a stationary UE. The document's own
emulated-crossing section distinguishes runs with nothing forced, so these silence numbers could be
read as a moving-UE crossing when they are not.

---

3. Full answer queues skip downlink PDUs ✓ Resolved

Description

The visit callback advances q.offered before checking whether the bounded answer queue accepts the
frame. When that queue is full, later visits skip those PDUs even though no answer was queued; a
failed send_packet has the same effect because the marker is never restored.

---

View review recommended (2)

4. Lost answers can become downlink gaps ✓ Resolved

Description

The new two-pass scheduler drains every unoffered PDU before returning to offered ones, despite
visit answers having no acknowledgement. If an answer is lost and more than 32 newer PDUs accumulate
before the target serves the UE, the UE confirms the skipped sequence numbers missing before their
retained copies arrive.

---

5. Late answers miss the UE's visit ✓ Resolved

Description

The answer worker sends queued frames without a deadline or a check that the UE remains on the
target channel. If the worker is delayed behind other answers or radio sends, it can transmit after
the UE's four-millisecond hold and return retune, while the PDUs have already been marked offered.

---

Tip of the day

💡 Did you know, you can move or hide resolved findings to focus on what still needs action

More tips ↗ | Customize Qodo ↗ | Qodo docs ↗

[josephnef]

Review addressed:

1. Two-run maxima: 4 runs per case. The doc table gives min / median / max of the DL gap across the switch, the longest DL silence, and uplinks lost before the switch:
   • source killed, answered: DL gap 225 / 225 / 226 ms, 0 uplinks lost;
   • source killed, no visits: 1497 / 1549 / 1805 ms, 10 / 10.5 / 14 uplinks lost;
   • no fault, answered: 225 / 225 / 226 ms, 0 lost.
2. Forced or not: the doc now states the handover is triggered by the CU's A3 at its defaults on the UEs' own measurements, with static UEs, not the emulated crossing, and points to that section.
3. Offered before queued: the offered mark moves only for an answer that gets queued. An answer that is not queued, expires or fails rolls it back, unless a later answer moved it.
4. Lost answers past the reorder window: pick_dl_pdus (cell.h, pure) sends not-yet-offered PDUs first, but an offered PDU goes before it would fall 24 sns behind them, inside the UE's 32-sn window. Selftest: a lost answer with a 60-PDU backlog never puts an offered PDU more than 32 behind the newest sent.
5. Late answers: each queued answer carries a 3 ms deadline (the UE's hold). A stale one is dropped and its PDUs given back. 0 expired in the 8 visit runs.

Rig on this code: m6_daps_xch.sh --runs 4, all 12 runs pass; the same-channel m6_daps_dl.sh passes (no answers sent).

[josephnef]

Applying skip-qodo-gate: Qodo posted its review as a PR comment (5 findings, two of them requirement gaps), not a review with threads, so the gate sees no review. All 5 are addressed and re-verified on the rig (m6_daps_xch.sh 4 runs per case, same-channel m6_daps_dl). All 28 builds pass (macOS gcc was cancelled by the runner once; it passed on re-run).