[PIR][WP23] DeAR-pattern decentralized capability grounding + thought-map navigation (ADR-326)

Epic: #837 · Wave 3, Phase W3-3 · Proposed ADR-326 · Paper: DeAR — Decentralized Agentic Reasoning via Capability Grounding (arXiv:2608.17282, Grade B+ — headline mechanism + benchmark count confirmed verbatim; one behavioral detail unconfirmable from the abstract). ⚠️ No upstream code today — the abstract states code ships only "upon acceptance." This is a reproduction-from-description build, materially larger scope than an adapt-from-repo WP.

Goal

Add three mechanisms to the Latent Communication Fabric:

1. Decentralized capability grounding — C(i,j,q) capability vectors with query-dependent peer specialization and no central role-assigner (local peer selection).
2. Shared thought-map — a RuVector thought-graph data structure agents navigate directly, rather than exchanging only point-to-point messages. LatentMesh transport.
3. Dead-end-triggered topology mutation.

Extends merged ADR-309/310/311 (LatentMesh greenfield fabric, causal-attribution gate, anomaly quarantine).

⚠️ CRITICAL — the continue-vs-restart behavior is UNVERIFIED (do NOT harden an acceptance criterion around it)

The design intent is that a detected dead-end triggers a topology update to CONTINUE, not restart. This "continue, not restart" reading is this program's working interpretation of the abstract's "topology update for adaptive error correction" — it is NOT a verified quotation from the paper (08-wave3-evidence-review.md §3, and the reason DeAR grades B+ rather than A).

• The ADR-326 Decision section MUST NOT treat continue-vs-restart as settled, and this WP MUST NOT harden its implementation around it as a fixed behavioral contract.
• Re-verify against the full paper (not available today) or the eventual code release before finalizing. If the full text or code contradicts the working interpretation, ADR-326's Decision section needs revision before WP23 locks in a specific behavioral contract.

Acceptance criteria

• Query-dependent peer specialization works with no central role-assigner (decentralized).
• Shared thought-map is directly navigable; agents traverse it rather than only exchanging point-to-point messages.
• Topology changes remain attributable (WP6 causal-audit gate) — every topology mutation traces to a cause.
• A misbehaving peer cannot trigger spurious topology churn (WP7 anomaly quarantine).
• Continue-vs-restart behavior is implemented behind an explicit "unverified — pending full-text/code confirmation" marker, not as a hardened contract.

Preprint-reproduction rule

DeAR's reported numbers are hypotheses, not the acceptance bar. No upstream code to check against; reproduction is from description. Promotion requires an independently recomputed research-gate paired-bootstrap delta over the pre-WP baseline.

Scope / files

crates/rvAgent/latentmesh-* (per ADR-305's WP5 greenfield crates — likely a new latentmesh-thoughtmap or an extension of latentmesh-align), coordinating with latentmesh-gate's causal.rs per the existing ADR-310 CI gate.

Depends on

WP5 (#849, greenfield latentmesh-* crates) · WP6 (#850, causal-audit gate) · WP7 (#853, anomaly quarantine)

Gates

Run npx @claude-flow/cli@latest security scan after landing (touches a topology-mutation surface).

Sources: 08-wave3-evidence-review.md (evidence grade + artifact status), 09-wave3-program-plan.md (WP/ADR mapping). Filed by the issue steward after coordinator approval of the Wave-3 breakdown.

Read the linked 08-wave3-evidence-review.md and 09-wave3-program-plan.md, then inspect the crates/rvAgent/latentmesh-* area and latentmesh-gate/causal.rs alongside WP5-WP7. The work is done when peer specialization, direct thought-map navigation, attributable and quarantined topology mutations, and the explicitly unverified continue-vs-restart behavior are implemented and the research-gate delta and required security scan pass.