feat(realtime): adapt semantic endpointing controls

[kvnloo]

What and why

Mirror the provider-neutral turn-detection contract in Talk's wire boundary: map OpenAI server/semantic VAD through mint and live update paths, advertise Grok native/server support and fail fast on semantic mode, advertise Gemini native-only support and fail fast on server/semantic modes. The semantic turn-detection names are probed as an optional capability, so pre-semantic core heads (today's #101808) keep the full core lane with provider-native turn detection instead of losing it. Ships a deterministic, content-free endpointing benchmark module.

Pairs with TheSmokeDev/hermes-agent#4 (stacked on NousResearch/hermes-agent#101808).

How to test

• Full suite: 1580 passed, 56 skipped, 5 xfailed
• tests/test_core_provider.py: 8 passed (was 7), 38 skipped without a core checkout
• Ruff passes on every changed Python file
• Wheel build succeeds and includes talk_endpointing_benchmark.py
• Benchmark smoke reports endpoint p50/p95/max, playback latency, timeout, cutoff/split, and false activations without retaining audio or transcripts

Platforms

• Linux (CI + developer box)
• Windows/macOS: untouched paths only; no platform-specific code in this PR

Live receipt

No live-provider run in this PR: provider lanes are exercised through contract-shaped unit coverage and the benchmark smoke module. Live-lane receipt stays with hermes talk check (#97) on a configured box.

Checklist

• Failing behavior reproduced first (old head dropped the whole core lane with a bare ImportError)
• Regression test binds the #101808-shaped contract and fails without the fix
• Full suite green (1580 passed, 56 skipped, 5 xfailed)
• Ruff clean on every changed file
• CHANGELOG carries the endpointing bridge entry, not just the benchmark
• No credentials, transcripts, or audio retained in tests or artifacts

[TheSmokeDev]

Thanks Kevin — this is a clean design: neutral RealtimeTurnDetection, exact GA semantic_vad shape, and honest Grok/Gemini refusals. One blocker before I can take it: talk_core_provider hard-imports RealtimeSemanticEagerness / RealtimeTurnDetection / RealtimeTurnDetectionMode inside the contract try, so on today's #101808 head (without your hermes-agent#4) the whole core lane disappears — core_contract_available() goes False with a bare ImportError, and the widened _REAL_CONTRACT_MARKERS makes every contract test skip instead of fail. Could you probe those three names as an optional capability (own try/except, getattr on validate_setup / setup.turn_detection, default to native), and add a test that binds the #101808-shaped contract? Also a CHANGELOG line for the endpointing bridge itself, not just the benchmark. Happy to merge right after. — SmokeDev

[kvnloo]

Addressed in 412c673 (full suite 1580 passed, 56 skipped, 5 xfailed, ruff clean): the three semantic names are now imported in their own try/except with a turn_detection_available() probe, mode sets degrade through _contract_turn_modes, and a missing setup.turn_detection defaults to provider-native via getattr. test_old_head_contract_keeps_core_lane_with_native_only_turn_detection binds a synthetic #101808-shaped contract (API v2, no turn-detection names) and proves the lane stays up degraded with native-only advertisement and a clean ValueError on semantic. Also added the bridge CHANGELOG entry and remapped the body to the repo template.