Add a local label filter for presence silence and no-cue intent

[tylerreckart]

Summary

A local Ollama model can score a closed set of single-letter labels and skip two completions Arbiter already makes. The filter ships dark: it runs only when ARBITER_DECISION_MODEL is an ollama/… id and the matching margin is greater than 0. A skip also requires a renormalized peak of at least 0.9.

• Presence. Before run_presence_review, score the first line of the task plus tool names. An extreme silent skips the review and emits presence with detail=filter plus filter_peak, filter_margin, and filter_label. A peaked context still runs the review. The filter cannot inject a note.
• Intent. Only when the heuristic kind is unknown. An extreme real kind that maps to a roster agent routes with source=decision and the canned brief, and the advisor completion does not run. Multi-cue text, a confident cue, unknown, a flat distribution, and a missing label fall through to the existing tagged completion, which is still the only writer of seeds.

Missing letters in top_logprobs are not filled with zero. Hosted models are not called. The advisor gate is unchanged.

Tests

• unit_label_score: renormalize, missing-label fail-open, OpenAI logprob body, margin parse
• unit_intent: peaked no-cue route skips the LLM; flat, unknown, multi-cue, confident cue, zero margin, and heuristic mode do not
• unit_presence: the filter snapshot keeps tool names and drops result bodies; only an extreme silent skips
• unit_api_client: the Ollama body gains logprobs only when the filter asks

orchestrator.cpp and api_server.cpp were syntax-checked. The full arbiter binary was not linked here (OpenTUI).