v9.14.0

Changed

• General pack 0.22.0: implementation routing splits on decision density. The implementation-spawn and inline-edit nudges (messages and classifier rubrics) now send bounded, decision-light work — the plan, work order, or repeated pattern already made the decisions and what remains is execution — to gpt-5.6-sol via the codex:codex-wrapper agent, and ambiguous, decision-dense, or long-run implementation to opus at xhigh. This replaces the "well-scoped, clearly-bounded" criterion and matches the fleet Models table v5, which also notes sol's two lane gates: judgment degrades as mid-task decisions pile up, and surrounding-code conventions are matched less faithfully (backstopped by finder/refuter review plus capt-hook guards).