[PIR][WP13] Implement closed-form KV-cache cross-model migration in ruvllm (arXiv:2608.03893)

Epic: #837 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md

Goal

Implement the closed-form linear KV-cache mapper for same-family model migration in crates/ruvllm (kv_cache.rs, paged_attention.rs, serving/kv_cache_manager.rs), with the nonlinear MLP fallback for degrading pairs, plus a routing gate that predicts transfer quality before migrating — so the system never migrates blind.

Why this ships first

Grade-A evidence, cheapest to build, and it depends only on ruvllm. No dependency on WP0–WP12 — it runs as a parallel track and can start immediately, landing while the main branch is still in Phase 0/1.

Evidence

arXiv:2608.03893 — "Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse," submitted 2026-08-04. The strongest-evidence item in the whole program review, exact match on every figure:

• The mapper runs 2.7–25x faster than re-prefill.
• Six pairs tested across three model families.
• The linear mapper retains 73–98% of the receiver's standalone-prefill accuracy on four pairs, while two degrade sharply.
• A nonlinear MLP variant recovers up to +37pp HellaSwag accuracy on the failing pairs.

The 2-of-6 failure rate is the point: it is exactly why the quality-prediction gate is a requirement rather than a nice-to-have.

Acceptance criteria

• Reproduces the paper's reported 2.7–25x speedup over re-prefill on an internal same-family model pair. Report our own measured numbers, never the paper's restated as ours.
• The routing gate correctly refuses or downgrades for a known-degrading pair.
• Closed-form linear mapper implemented for same-family migration.
• Nonlinear MLP fallback implemented for pairs the linear mapper degrades on.
• No code path migrates without consulting the gate. Enforced structurally, with a bypass-attempt test.
• Per-pair accuracy retention measured and recorded; degrading pairs are identified automatically rather than hand-maintained in a list that will drift.
• Degradation is observable at runtime — a migration whose realized quality falls below prediction is logged and counts against the gate's calibration.
• Published as @ruvector/kv-migrate per the existing @ruvector/* release flow.

Scope boundary

Cross-provider routing (Anthropic/OpenAI/local) is not in scope — that role sits closer to claude-flow's ADR-026 3-tier routing. This is same-family KV-cache transfer within ruvllm's serving runtime.

Dependencies

None. Independent parallel track, can start immediately.

Start with docs/research/perpetual-intelligence-runtime/03-program-plan.md and inspect crates/ruvllm/kv_cache.rs, paged_attention.rs, and serving/kv_cache_manager.rs, then compare the paper's mapper and failure cases with the existing serving flow. Done means the gate, linear mapper, fallback, structural bypass test, measurements, observability, and @ruvector/kv-migrate release criteria are demonstrated without adding cross-provider routing.