TOIN skip_compression_recommended is never applied: toin_publish drops it, recommendations loader is unwired #1775

Verified on v0.27.0 (deployed) and current main.

• PR-B5 retired the request-time hint API; recommendations flow via headroom.cli.toin_publish -> recommendations.toml -> crates/headroom-core/src/transforms/recommendations.rs.
• toin_publish.py publishes only strategy_hint/confidence/observations. ToolPattern.skip_compression_recommended is dropped (_eligible_rows/_select_strategy).
• recommendations.rs module doc, current main: the dispatcher "does not consume this surface yet — PR-F3 is responsible for wiring it". So even published hints are dead ends.
• Deployment data (0.27.0, 17,052 compressions lifetime): 228/1243 patterns flagged skip_compression_recommended, all still compressing. Of 241 patterns with >=10 compressions, 45 have >60% retrieval rate, several >100% (same item retrieved repeatedly). 91% of all retrievals are full-content, so for these patterns the wire cost is compressed copy + retrieve call + full original — more than never compressing.
• Those same patterns carry strategy_success_rates {"kompress": 1.0}, so toin_publish emits strategy_hint = "kompress" for exactly the patterns TOIN flagged to skip.

Asks:

1. Wire the dispatcher to the RecommendationStore (PR-F3).
2. Publish the skip signal in the TOML schema.
3. Make strategy scoring retrieval-aware (a pattern with ~100% full-retrieval rate should not score kompress 1.0).