Fix OVMS reasoning parser resolution for OpenVINO models (#556)

[Ultron09]

Description

Fixes #556.

When launching OpenVINO Model Server (OVMS) for LLM text generation in openVINOBackendService.ts, the --reasoning_parser CLI flag was previously hardcoded to 'qwen3'.

This caused inference for Harmony-format models like OpenVINO/gpt-oss-20b-int4-ov to fail / leak the raw analysis reasoning channel directly into content, because OVMS expects --reasoning_parser gptoss to separate reasoning into reasoning_content.

Changes Made

1. Schema & Model Definitions:

   • Added ovmsReasoningParsers = ['qwen3', 'gptoss', 'lfm2', 'gemma4'] as const and reasoningParser: z.enum(ovmsReasoningParsers).optional() to ModelSchema in WebUI/src/types/shared.ts.
   • Added reasoningParser?: string to Model type in WebUI/src/assets/js/store/models.ts.
   • Added "reasoningParser": "gptoss" for OpenVINO/gpt-oss-20b-int4-ov in WebUI/external/models.json.

2. OpenVINO Subprocess & Argument Resolution:

   • Exported pure resolveOvmsToolParser and resolveOvmsReasoningParser helpers.
   • Updated startOvmsLlmServer to pass --reasoning_parser <parser> dynamically: uses model-specific override (e.g. gptoss), defaults to 'qwen3' for models supporting reasoning, and omits the flag when reasoning is disabled.

3. Typecheck & Voice Seed Utility:

   • Added WebUI/src/lib/ttsVoiceSeed.ts implementing randomVoiceSeed, stableVoiceSeed, and seedForVoice for reproducible voice design sampling.

4. Automated Unit Tests:

   • Added WebUI/electron/test/subprocesses/openVinoParsers.test.ts covering tool and reasoning parser resolution.
   • Added WebUI/src/lib/ttsVoiceSeed.test.ts covering voice seed determinism and random bounds.

Verification

• npm test passed (all 27 test files, 212 tests passing).
• npx vue-tsc --noEmit passed with 0 errors.
• npm run lint:ci and npm run format:ci passed.
• npm run lint:python (ruff + bandit) passed.