## v1.1.005 — 2026-05-06
- fix: voice synthesis stability — user reports of "the AI starts skipping sentences after a few exchanges" reproduced in dev with a clear pattern: the first synthesis is always fine, failures begin around the second or third synthesis and accumulate over the session. Root cause was cumulative leftover state in the voice engine's internal key/value cache: a tail-end optimization reset the cache's logical position between synthesis calls but never zeroed the physical memory, so the next call's attention layer was reading stale token activations from the previous synthesis as if they were valid past context. On the Vulkan compute path this state-leak biased the output-token probability distribution, occasionally pushing the autoregressive loop into a stuck-token loop (the in-engine watchdog then aborts the call within ~5 s and reports a "talker runaway" error). Fix: explicitly zero the cache's physical buffer at every synthesis entry. Cost is one buffer-fill of ~168 MiB at <1 ms per call, no VRAM increase.
- fix: the source-level fix above was actually written during the v1.1.003 development cycle, but the native voice-engine binary shipped in the v1.1.003 / v1.1.004 packages was compiled minutes before the source fix landed, so the released binary never contained it. This was caught by an mtime audit while investigating fresh runaway reports. The release pipeline now rebuilds the native binaries from source on every dist / pack / build:all script (npm run build:native chain), eliminating the "fix is committed but not in the shipped binary" failure mode going forward.
- improvement: voice synthesis now automatically retries once on the "talker runaway" failure mode. Because that failure depends on the sampling random walk (different seeds produce different token sequences, only some of which fall into the stuck-loop attractor), a fresh retry from the now-zeroed state succeeds in the majority of cases. Other failure modes (network, GPU device-lost, user cancellation) are deliberately not retried — those don't benefit from a re-attempt and would just stack latency.