Feedback worker model tiering cost

Justin said a session running 10 workers all on Fable 5 was using more usage/tokens than needed, and asked for a leaner process. New standing rule for /spawn and worker dispatch:

• Sonnet — build-from-spec workers (most WDA/app-build work)
• Haiku — mechanical sweeps: renames, alias tables, config edits
• Fable 5 — reserved for orchestration/merge reconciliation and genuine root-cause debugging only

Also leaner rounds: briefs shrink to ledger pointers + a repo-map of the worker's seam instead of letting each worker re-explore the codebase from scratch; worker self-QA drops to typecheck + one smoke check, with the full visual/Playwright pass happening once post-merge instead of per-slot.

Why: repeated codebase exploration was consuming a large chunk of each worker's 100-270k token budget, and Fable 5 usage credits are worth conserving for tasks that actually need frontier reasoning.

How to apply: when spawning workers via /spawn or dispatching WDA-style build tasks, default to Sonnet, use Haiku for mechanical work, and only reach for Fable 5 when the task is architecture, merge conflict resolution, or root-cause debugging.

[auto-memory session 69928070-35c5-4f35-a5cd-09393fe53e64, confidence 0.85, mode direct]