monitor: focus filter spec + validation harnesses

[anish749]

Problem

The earlier monitor-semantic-filter-spec.md proposed a multi-vector intent filter (focus + positives + negatives + window fallback + geometry-derived threshold). Validation across three workspaces showed that simpler is strictly better under the metrics that matter for fan-in reduction (the agent on the receiving end self-filters; precision at the router buys nothing).

Solution

• New spec at docs/monitor-focus-filter-spec.md: pure focus-cosine threshold per filter, multi-route across instances. Six-line algorithm. Threshold default 0.22 with --mode recall preset at 0.18. Section 2 of the spec records the full validation methodology, all metrics, and per-component rejection rationale.
• Validation harnesses under experiments/: ground-truth labelling via Claude judge, threshold sweeps reporting recall / wake-rate-on-noise / sessions-woken-per-msg, focus-prose pairwise geometry analysis, embedder comparison (MiniLM vs bge vs e5).
• workstreams.example.json documents the config schema; the actual workstreams.json and labels-*.json are gitignored (workspace-specific data).

Context

Supersedes the previous spec but does not delete it; the new spec links to it explicitly. Reviewer should read Section 1 first; Section 2 is the methodology trail in case anything looks underjustified.