#1773 bench: historical Q8 TCP parity and timings on three Tesla P4s

Descrição

Preserves the 2026-10-09 same-host Q8 transport control: all nine 128-token local/TCP pairs matched exactly, as did all 12 within-arm repeat checks. TCP decode throughput was 21.6–22.6% lower for these cases. This is a historical hardware report, not a new runtime fix or current-main benchmark.

Updated the report branch against main 61b3fb5d without changing raw receipts. Independently rechecked prompt hashes, output IDs, the decode-rate formula and every median/min/max in the summary. No new inference runs are claimed.

Full report and reproduction: bench/results/2026-10-09-r730-q8-tcp-parity/README.md.