Display observation (Cyd, on-cabinet): held F0000000 through the clean
handshake and four NAK-then-ACK sub-threshold cycles, flipped exactly at
the 5th teardown to E0000105. That reading recalibrates the counters:
$3184 counts give-up cycles (not per-retransmit) and $3185 counts
teardowns of any imperfect reply cycle — the counter behind the original
lone E0000001. Also recorded: 5 retransmits per cycle (not 4), reply-
await arms only after the complete frame, NAK forces exactly one counted
retransmit, and a late ACK cannot rescue a cycle once retries begin.
Tool predictor updated to the confirmed semantics.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two-phase test against a DTR-reset board: phase A forces sub-threshold
retransmits (NAK the completed reply frame -> exactly one counted
retransmit, then ACK the retry), where the display must hold F0000000;
phase B withholds ACKs for a full give-up cycle to cross the threshold,
with the expected E0 readout predicted from observed traffic.
Bench findings while building it (9600, e0t5 chip): the reply-retry
machine sends 5 retransmits per cycle (not 4); responses are honored
only after the complete reply frame (mid-frame ACK/NAK/RESTART is
ignored); once the first retry fires the cycle runs blind to give-up;
a host NAK triggers an immediate counted retransmit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CydandClaude Fable 5