Ground YouTube comment replies in per-comment context

[Foundup]

What changed

• Extracts the exact video title, video ID, and video URL from each YouTube Studio inbox row and propagates the receiving channel ID for personality routing.
• Preserves the original comment when semantic guidance is added, applies commenter history and verified VoiceMemory to every tier, and makes #FFCPLN context-sensitive instead of automatically treating it as a troll marker.
• Makes the local model route the default: LM Studio first with a 3-second bound, then embedded general-purpose Qwen. Grok is now an explicit legacy opt-in.
• Uses resolve_general_model_path() instead of the coding-role model and passes embedded Qwen the same full grounded prompt as the service route.
• Records the actual provider beneath skill routing and fails closed when only template, BanterEngine, stats-only, unknown, or stale output is available. Legacy generic replies require YT_ALLOW_GENERIC_REPLY_FALLBACK=true.
• Adds focused regression coverage and updates module/test documentation.

Why

The active daemon showed Grok unavailable, LM Studio consuming most of the 10-second budget, embedded Qwen receiving a reduced context-free prompt, and generic templates such as “Appreciate the comment!” being posted and mislabeled as contextual output. Studio rows also lost their video context before generation.

OpenRouter / RedDog decision

RedDog's production Fusion calls are autonomous and do not require 012 to approve individual calls. They consume policy and signed runtime/model bindings and persist durable, content-free provider-call evidence.

This PR does not import the private advisory_model_once.py bridge because the currently exposed production runners are typed for repository audits, backend-architect decisions, and bounded artifact generation—not short-form YouTube replies. The next RedDog slice should expose a provider-neutral reply runner that consumes preconfigured daemon policy and returns reply text plus a durable call receipt. 012 remains out of the per-comment loop.

Validation

• python -m pytest modules/communication/video_comments/tests/test_reply_context_pipeline.py modules/communication/video_comments/tests/test_ytr1_runtime_hardening.py modules/communication/video_comments/tests/test_classifier_pipeline.py -q — 23 passed
• python -m py_compile on the three changed runtime modules
• git diff --check
• WSP-97 repository evidence receipt — compliant, 0 violations