fix(vad): clamp cursor when resizing speech buffer

[gudcks0305]

Problem

Reducing max_buffered_speech or prefix_padding_duration while a VAD stream has buffered speech can shrink its sample array below its write cursor. The next speech event then raises ValueError: data length must be >= num_channels * samples_per_channel * sizeof(int16) and terminates the stream. Shrinking and immediately growing the buffer also leaves discarded samples counted as valid audio.

Change

Keep the cursor on the stream and clamp it when resizing the buffer, in both inference.VAD and the Silero plugin. Existing event emission, prefix retention, and reset paths use the same cursor. Add deterministic regressions for both options, buffer growth, shrink-then-grow, subsequent speech, and explicit reset.

Load the optional Silero plugin only for its fixture cases, so its absence does not prevent the bundled VAD regressions from running. Skip missing modules while allowing a general ImportError from a broken installed plugin to propagate.

Validation

• Before the fix: 8 regression cases failed; 4 growth controls passed.
• After the fix, with Silero installed: all 12 regression cases passed. They use synthetic PCM and mocked VAD predictions while exercising the real stream and rtc.AudioFrame buffer paths.
• A process-local import-hook simulation of missing Silero: 6 bundled VAD cases passed and 6 Silero cases skipped. A simulated general ImportError propagated as 6 fixture errors with a nonzero exit status, while the 6 bundled cases still passed.
• Related audio recognition, turn detection, and PCM buffer suites: 199 tests passed, including those regressions and local resampling tests.
• Existing native VAD executor lifecycle test: passed.
• Repository Ruff formatting/lint and focused mypy on both changed production modules: passed.

The local environment installed core Agents plus the Silero extra. Full --unit --audio_eot collection stopped at missing optional provider packages, and the project-wide type check was blocked by missing optional packages and typing stubs. Focused mypy checked the two changed production modules with --follow-imports=silent. No cloud/provider integration tests were run.

AI assistance was used to develop and test this change.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

🔍 Devin Review: 1 flag

Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)