Multi-model fleet tiering for the prompt surface (deferred from Spec 1252 / D4)

[waleedkadous]

Context

Issue #1252's proposed direction #2 — tier prompt content by consuming model capability (weaker models get fuller scaffolding, stronger models the compressed form) — was cut from Spec 1252 by architect decision D4 and deferred until the ownership map existed.

That prerequisite now exists: codev/resources/prompt-ownership.yaml (Spec 1252 Phase 6) enumerates every prompt surface with load semantics, every instruction class with an owner, and machine-checkable completeness (T12). Tiering becomes specifiable as an attribute of that map.

Prior design notes (from the cut Appendix C, for whoever picks this up)

• Two tiers to start: strong (compressed) / weak (fuller scaffolding); selector via an explicit promptTier key in .codev/config.json rather than model-string sniffing (breaks on every new model id); fallback to weak (failing toward more instruction is the safe direction).
• Scar rules are tier-invariant — full canonical text at every tier (codev/resources/scar-rules.yaml; D3).
• Behavioural guardrails from Spec 1252's M12 machinery apply: capture a baseline before tiered serving ships, compare after.

Refs #1252

🤖 Generated with Claude Code