[korexil]

All four stages were measured again on 2026-10-16 in one sitting (01:00-01:37), 3 runs each after a warm-up, same text question, with a check that drops any record holding more than one answer (all 12 runs were valid). Raw records: bench/results/stages-2026-10-16.txt.

• README (all three languages): before/after is now 4.4 -> 2.3 s first audio and 4.59 -> 0.09 s total gaps (3-run averages). The 10-14 numbers (6.2 -> 2.1 s, 4.22 -> 0.17 s) came from 2 runs per stage; the origin stage's first audio varies a lot (3.5-6.0 s) and the earlier two runs were both at the high end. The gap figure is now credited to the native engine and turn-taking together.
• docs/benchmarks.md: new table, method notes (order 1, 0, 2, 3; text questions, so first audio excludes VAD/ASR; a gap counts even at 0.01 s; only gaps of 0.1 s or more have timestamps in the raw file), and a short note on the 10-14 run.
• Figures regenerated; the stages figure now sizes each row by its number of runs. Social preview shows the new numbers.
• README: on Linux, CI only builds the CPU version of the C bridge and checks its ABI; the CUDA engine and the full pipeline have never run there.