Lost answers after rate changes

The pending DATA record determines recovery geometry. The configured target rate can differ from the record actually on air while descending through the ladder, and a NAK can replace a base frame with a smaller pending frame. Checking either the base record alone or the target's payload capacity excluded valid recovery states.

Full frames now use the existing phased intermediate state query when the actual pending body is full at its recorded level and more DATA is queued. Short final frames use the query/request/drained exchange when no further DATA is queued. Both retain the existing bounded attempts, complete-reply checks, same-pending identity and wall-clock/sample freshness requirements. Neither silence nor a query by itself retires DATA. Unsupported states still close unconfirmed.

The seven-symbol clipped-continuation exception remains scoped to the measured base-record links; broader query recovery does not extend that exception.

Qualification

Full BW2300 record 2 remains outside the query scope. The ordinary sender uses that record for a short close; no full-record-2 query exchange was qualified. Each test also receives the same exact 225-byte greeting from stock.

BW2750 runs: 20260919T170553Z (record102), 20260919T170930Z (record101), 20260919T171254Z (record100). BW2300 closing-record run: 20260919T171639Z. Their captures remain in the private station archive.

The virtual-audio harness runs the production mail_session against VARA HF 4.9.0. Each rate has a clean baseline, an erased intermediate ACK, and an erased short-final ACK. Independent cable recordings retain the original stock answer; Kestrel's receive recording contains zeros across the selected callback range. The subsequent query and complete reply are decoded independently, as are the CRC and payload of every transmitted DATA frame. Success requires exact bytes in both directions, no pending/queued DATA, no audio loss, unchanged source hashes throughout the run, and completed process cleanup.

Native control-only excerpts, pending-frame metadata and exact source boundaries are shipped under tests/kestrel/fixtures/lower-answer-recovery/. Tests replay those excerpts through multiple callback sizes and reject unsolicited, stale and wrong-pending answers. Full captures and audits remain in the private station archive.

These are virtual-link measurements. They do not establish RF fading performance, near-capacity short-trailer qualification, or general SNR-based rate adaptation.

Retry phase and the double-loss boundary

A full retransmission advances query phase too. The sender's rate-ladder index still counts new full frames, so a separate retry parity tracks this protocol phase without accidentally skipping a rate. A refused retry advances neither. The next successfully sent short frame and a new session reset the phase.

In the native two-stock BW2750 double-loss experiment, the first base DATA was erased, then the ACK of stock's 22-byte record101 retry was erased. Stock keyed 60/683 before its first NAK and 60/745 after the full retry. The latter drew 288/249 (31/31 payload tones), then stock sent the remaining72 bytes as a base record103 short frame. Both hosts received exactly94/230 bytes; buffers emptied. Stock's lower retry carried field0x91 in this experiment.

Kestrel's short-remainder policy holds record1/101 with field0x89. With the correct 745 query phase its missing retry ACK draws the ordinary 288/1 answer and it can continue at that lower record. The measured288/249 answer also has a scoped handler: it confirms the lower retry and regroups the unsent short remainder at base, preserving later host deliveries and bytes across a refused next TX.

Reusing683 after the retry instead drew a different16-symbol289/541 frame. Treating that response as an ACK failed to complete delivery. It is explicitly rejected; its waveform is retained only as a negative regression fixture. Failed experimental interpretations remain in the development captures.

Stock experiment IDs: clean 20260919T175448Z and double-loss 20260919T180025Z. The DATA fault was a restored mute on the virtual BlackHole16 input; the ACK fault was an exact sample erasure in the established reverse relay. Actual receiver input is recorded, and both virtual levels and Wine input settings were restored.

Kestrel double-loss and erased-retry negative controls exercise both paths. An erased lower retry draws a NAK and preserves the pending frame; losing its ACK solicits positive state and advances exactly once. No16-symbol frame is accepted as this positive answer.

The final Kestrel short double-loss pairs (20260919T180917Z BW2300 and 20260919T181125Z BW2750) pass94/225 bytes exactly. Longer393-byte double-loss runs (20260919T181535Z and20260919T181658Z) pass393/225 bytes exactly and exercise288/249 before the remaining base frames. Each has independently decoded CRC-clean DATA, intact phased query/reply, verified receive erasure, empty pending/queue at completion, unchanged runtime hashes and completed cleanup. Earlier393-byte clean runs establish the normal transfer baseline; the phase change is exercised only after a full retry.