DATA rejection at host speed levels 1–4

The September 19–20 stock VARA HF 4.9.0 campaign found that the eight-symbol continue shape also carries a DATA NACK. Its common first tone pair is not positive acknowledgment. In the preserved 20260919T233737Z-2300-crc-observe run, the old detector retired 47 rejected bytes and sent the next frame; stock received only the preceding 89 bytes. An independent cable decode confirms that the rejected frame had the intended body and deliberately incorrect CRC.

Measured replies and actions

Measurements use W9SSJ calling KC9GHZ, plus an independent N0XYZ peer, on fixed BlackHole virtual audio. The two stock installations are speed limited, so nothing above host level 4 is qualified by this campaign. ARQ record numbers differ from host levels:

A complete queued delivery whose full pending frame was CRC-rejected retries one supported record lower at the wide bandwidths, retaining every unconfirmed byte and holding the lower record for the remaining delivery. BW500 uses host levels 4 → 3 → 1 for CRC-rejection recovery: stock rejected repeated level-2 DATA in the consecutive-failure experiment, whereas the level-1 candidate delivered all bytes. Bench-selected announced clean descents can still use all four levels. At the lowest record a NACK repeats that record. A positive answer to a newly transmitted frame resets the NACK retry budget; repeated NACKs for the same pending data do not.

Before any DATA has been accepted, stock can send the called-keyed 16-symbol 289/31 rejection, including at lower records. After earlier DATA was accepted, it can instead send an eight-symbol two-tone rejection. N0XYZ also sends the long called-keyed 288/63 rejection directly after DATA. These are observed link/state differences, not a rule assigning one NACK shape to each speed. The long form is accepted in a fresh direct-DATA window or a solicited query window; an abandoned query cannot reuse the direct window. When a direct CRC-rejection reply is erased, the called-keyed query answer is also 288/63.

The eight-symbol negative is frame 1 of the called station's answer stream (04 §4.2C, answer tails), which reproduces all three measured tails. Complete clear matches authorize a retry only in the pending full-DATA window. Near matches veto an ACK but do not authorize a retry. Native fixtures retain source hashes and sample boundaries in tests/kestrel/fixtures/data-nak-pairs-0919/manifest.json.

Positive hold and next-lower-record replies are frames 0 and 5 of the same stream, covered by the native data-ack-pairs-0919 fixtures. Any other eight-symbol reply must obtain called-keyed positive query confirmation before retiring bytes. A short final over, a session's only over among them, is its own delivery boundary and takes the same one-step descent (28 geometry; not yet qualified for CRC rejection). Outside the reply window, or after a query, the burst authorizes no DATA and recovery stays with the query.

Evidence and scope

The full development recordings do not ship. The mapping experiments deliberately select entry records and suppress automatic NACK action to measure replies. Its clean descents deliver exact bytes through all four records at every bandwidth. CRC failures and lost frames are separate arms; an arm being a valid measurement does not imply transport succeeded. transport/ exercises production ARQ, with faults applied at the waveform or actual receiver callback, and captures an independent cable in each direction. The early BW500 fault renderer preserved exact encoded body/CRC bytes but used different leading padding from production. From 20260920T004942Z onward, it uses production onset and column settings too; prior runs are retained as negative/candidate evidence rather than padding-identical CRC-only controls.

The missing-DATA answer 288/1117 is distinct from CRC rejection. A preliminary 289/541 reply sometimes precedes it in the mapping arms; its meaning is not qualified as permission to retire or retransmit DATA. BW500 missing-DATA fallback above its lowest record and short-final CRC rejection remain unqualified by this change. Explicit TX ladder configurations are outside the automatic NACK repacketization path. The previously qualified wideband missing-DATA and BW2300 short-final actions remain documented in 23, 24, and 28.

Regression checks

The final focused ACK/NACK, native reply, repacketization and final-answer suite passes 259 tests. Additional BW500 codec tests pass 55 tests and session/continue/window tests pass 41. Source hashes are frozen within each stock arm; cleanup and independent wire audits are recorded beside each run.

K0SI at 01:14 UTC September 20: ACK query scan timing

The capture logs/onair/20260920T011453Z-W9SSJ-K0SI.wav contains exact called-keyed 288/1 answers to both queries after DATA over 2: all 31 payload tones match. The receive cursors are samples 4198048 and 4511904 at 48 kHz. Native excerpts and their hashes are in fixtures/intermediate-query-answer/k0si-20260920-provenance.json.

Uneven receive callbacks at 1.37, 1.46 and 1.89 seconds reproduce the lost answer: the 1.37-second scan finds an unfinished ACK, schedules the usual half-second step and ignores the 1.46-second callback. By 1.89 seconds the complete answer is beyond the 0.5-second freshness limit. An exact partial payload now schedules the next scan at its fitted tail. The full final symbol, called-key match, pending-message identity and freshness checks still apply. Both native excerpts advance the pending frame exactly once at 1.46 seconds; wrong calls and missing tails cannot advance it.

The original short DATA response is also audible in the recording. The rig's receive mute extends into its opening, and several interior two-tone symbols have insufficient separation from competing peaks. Several peaks fall at third-order intermodulation frequencies (for example, bin 63 beside 77/91); the capture does not establish whether transmitter or receiver caused this. At the time of this QSO, no W9SSJ/K0SI ACK tail was known. This evidence does not justify accepting the short shape as an ACK: the query remains necessary for this capture, and its first complete answer should resolve the pending frame.

K0SI at 01:36 UTC September 20: a qualified short ACK

In logs/onair/20260920T013613Z-W9SSJ-K0SI.wav, DATA over 2 draws a complete short reply at receive cursor 3301888. The first query response, at cursor 3467904, matches all 31 payload tones of K0SI's 288/1 ACK. There is no DATA retransmission between the short reply and this confirmation. The pending 89 bytes are then retired, the remaining 55 sent, and the mail exchange completes.

The short reply's eight pairs are (64,67) (58,92) (77,91) (48,64) (65,71) (38,92) (81,85) (58,98). They are K0SI's generated continue tail. Native fixtures, cursors and hashes are in fixtures/k0si-data-ack-0920/provenance.json. The intact reply advances exactly once without a query with 512, 4096, 4800 and 8192-sample callbacks. The earlier distorted reply still cannot meet the existing tone checks, even with the known tail, and remains a query case.