Specify frame-item attribution hooks so hosts can evaluate retrieval quality

[macanderson]

Context

The Context Frame spec's sixth question — "Why was each item included, and can its effect be evaluated later?" — currently has provenance (where an item came from; #12 covers integrity) but no specified way to close the loop: nothing ties a frame item to downstream outcome so a host can score whether including it helped. Concrete consumer: stella already A/B-suppresses recall on a fraction of turns to measure recall value, entirely outside the protocol.

Task

• Specify a stable per-item attribution id and an optional host→provider feedback surface (item id → outcome signal), so retrieval quality is measurable across implementations.
• Keep it optional/capability-gated: providers that don't want feedback simply don't advertise it.
• Coordinate with Define token_cost semantics: a canonical counting rule so budget honesty is actually checkable #8 (token_cost) — cost + outcome together enable value-per-token ranking.

Acceptance

• Spec section + schema for attribution ids and the feedback capability; one worked example showing a host scoring an item's inclusion.