[PIR][WP20] Pre-compiled pipeline-sharded multi-node ruvLLM serving (pipeline-shards pattern, ADR-323)

Part of the RuV Perpetual Intelligence Runtime (PIR), Wave 2. Epic: #837. Plan: docs/research/perpetual-intelligence-runtime/07-wave2-program-plan.md · evidence: 06-wave2-evidence-review.md (both land on main with the feat/pir-wave2-adrs PR).

ADR mapping

Proposed ruvector ADR-323 (extends merged ruvector ADR-314, KV-cache cross-model migration, WP13 #844) — pre-compiled pipeline-sharded serving for ruvLLM on heterogeneous multi-node fleets: pipeline parallelism across nodes with pre-compiled per-stage graphs and speculative decoding. Ships independently of ADR-314 (multi-node topology, not cross-model cache reuse) but lives in the same Cross-Model Cognition Migration context and the same crates/ruvllm surface.

⚠️ Why ADR-323, not ADR-322: ruvector deliberately skips 322. This program has already hit the "wrong repo owns this ADR number" failure mode four times, and ruvector ADR-305/306/310/312/313 repeatedly cite ruflo ADR-322/322A/322B/322C as the flywheel-receipt/promotion authority. Minting an unrelated ruvector ADR-322 would put two heavily-cited, differently-scoped "ADR-322"s in concurrent use across sibling repos — exactly the ambiguity ADR-305 §4 exists to prevent. Rationale in 07-wave2-program-plan.md; confirm 323 is still free against the live docs/adr/INDEX.md at kickoff.

Evidence

Pipeline shards (arXiv:2608.19147) — "Pre-Compiled Pipeline Shards for Distributed LLM Inference on Intel AI PC Fleets," submitted 2026-08-19. Grade A — the strongest-evidence item in this wave; all figures exact match. Artifact available now: github.com/labscommunity/pipeline-sharded-inference-paper — the strongest artifact package of the six (full reproduction/ directory with claims/hardware/results docs + scripts) — adapt the released reproduction package.

⚠️ Number-conflation flag: the 1.79× figure and the 4-node/70B interactive-serving result are two separate results in the source paper. Target the verified 1.79× on an internal 2-node/8B-class same-family setup; treat the 4-node/70B result as a separate, never-1.79×-labeled target.

Scope

• Implement pre-compiled pipeline-sharded multi-node serving in crates/ruvllm.
• Surfaces: crates/ruvllm (kv_cache.rs, paged_attention.rs, serving/kv_cache_manager.rs) — the same files WP13 (#844) touches; coordinate to avoid merge conflicts with the open WP13 PR (#854).

Candidate-mutation rule (applies to all of Wave 2)

Candidate mutation, not trusted prior art. Promotion requires this program's own research-gate paired-bootstrap-recomputed delta over the pre-WP baseline — never the paper's self-reported numbers.

Acceptance criteria

• Pipeline-sharded multi-node serving implemented in crates/ruvllm (ADR-323)
• Internal 2-node/8B-class benchmark records the measured speedup; the 4-node/70B configuration tracked as a separate target, never labeled 1.79×
• research-gate-measured delta over the pre-WP baseline before promotion
• No merge conflicts with WP13's serving-surface changes (coordinated)
• ADR-323 merged with repo-qualified citations per ADR-305 §4, citing "ruvector ADR-323 (extends ruvector ADR-314)"

Depends on: WP13 (#844) — soft dependency only (shared crates/ruvllm surface, no hard blocker). Phase W2-4 — can start anytime resourcing allows, same independent-parallel-track pattern as WP13 in Wave 1.

Start with docs/research/perpetual-intelligence-runtime/07-wave2-program-plan.md, 06-wave2-evidence-review.md, and the live docs/adr/INDEX.md to confirm ADR-323. Then inspect crates/ruvllm/kv_cache.rs, paged_attention.rs, and serving/kv_cache_manager.rs alongside WP13 #844 and PR #854. Done means coordinated multi-node serving, separate 2-node/8B and 4-node/70B benchmark records, a research-gate delta, and the merged ADR.