task: Implement Kiro spec downstream-cache-affinity-hinting

[matdev83]

Implement the approved Kiro SDD introduced by #565 and aligned to the lean-core target by merged spec-only PR #569.

Spec: .kiro/specs/downstream-cache-affinity-hinting/

The canonical spec is fully approved and ready_for_implementation: true, but it carries an explicit execution gate: production implementation begins only after core-feature-ownership-full-closure is implemented/certified and its final ownership census/core-admission ratchets are present on main.

Outcome

Deliver provider-facing downstream prompt-cache locality hinting without regrowing generic core:

• feature-owned HMAC derivation/fill-only synthesis/telemetry under the downstream-cache-affinity feature;
• standard construction through the post-closure standard featurehost;
• reuse existing candidate-aware PlaneAttemptTransforms, with no cache-specific executor stage/extension plane/core metrics callback;
• direct OpenAI Responses prompt_cache_key forwarding/validation repair;
• non-lossy provider-profile lifecycle repair or reuse of the already-landed single owner;
• typed provider-profile projection and the frozen xAI/Mistral/Fireworks/RunInfra/OpenRouter behavior matrix;
• executable connector feature negotiation without a cache-specific protobuf value DTO/protocol bump;
• strict separation from secure-session authority, continuation, cache-residency control and keep-warm.

Dependencies

• Blocked by task: Implement Kiro spec core-feature-ownership-full-closure #572 — spec: align downstream cache affinity with lean core architecture #569 explicitly requires the completed/certified full core feature-ownership closure as the production implementation baseline. Both trackers are in the same post-split milestone, so this is the correct hard dependency to model.
• Do not model the earlier OSS Base pre-OSS slimming task as another blocker here; milestone sequencing and task: Implement Kiro spec core-feature-ownership-full-closure #572's own execution baseline already preserve that prerequisite transitively.
• Bulk provider expansion may land the shared provider-profile lifecycle repair earlier. Reuse the single authoritative implementation rather than creating a competing bridge; this is coordination/revalidation, not a cross-milestone dependency.

RED execution gate

Before changing production behavior, prove the post-closure request pipeline still has no later writer that creates/replaces PromptCacheKey after PlaneAttemptTransforms and that the key survives request hooks, candidate adaptation and final raw-session scrub. If that premise is false, stop and repair the SDD instead of adding a cache-specific late core stage.

Do not close this work order until all spec tasks and provider/connector/profile certification are complete on merged main.