[PIR][WP22] D²ACCI-pattern stage-level memory diagnostic gate + DCR metric (ADR-325)

Epic: #837 · Wave 3, Phase W3-2 · Proposed ADR-325 · Paper: D²ACCI — A Dual-Loop Diagnostic Protocol for Evidence-Preserving Agent memory (arXiv:2608.17756, Grade A, all figures exact-match). ⚠️ No upstream code or benchmark artifact exists today — this is a reproduction-from-description build, materially larger scope than an adapt-from-repo WP. Size accordingly.

Goal

Build a stage-level diagnostic gate over the five memory-pipeline stages — ingestion, extraction, retrieval, filtering, generation — in crates/ruvector-agent-memory, plus a graded DCR (Diagnostic Coverage Ratio) observability metric attached to every memory-mutation's promotion record.

Extends merged ADR-307 (three-level persistent memory, WP3/WP4), ADR-320 (MemFuse-pattern atomic observation + causal graph, WP18), ADR-282/ADR-306 (research-gate paired-bootstrap promotion gate).

The key architectural distinction — STAGE-level, complementary to research-gate's OUTCOME-level gate

research-gate today answers a single outcome-level question: "did this whole mutation beat its parent?" It does not localize which of the five memory-pipeline stages changed behavior when a memory-system mutation's end-to-end score moves. This WP adds exactly that fault-localization layer, and it sits underneath research-gate, layered, not a replacement — the outcome-level paired-bootstrap statistics remain the promotion authority; the diagnostic gate adds stage-localizable attribution beneath them.

Every memory-mutation promotion record carries a graded DCR metric (per the paper).

Acceptance criteria

• Each of the 5 stages is independently instrumented so a behavior change is stage-localizable — the promotion record identifies which stage moved, not just that the end-to-end score moved.
• Promote a memory strategy only on: paired-improvement (research-gate's existing outcome-level delta) AND no protected-regression AND stage-localizable provenance. Every memory strategy is treated as an immutable, versioned policy artifact.
• DCR metric computed and attached to every memory-mutation promotion record.
• The diagnostic layer is wired under research-gate's existing paired-bootstrap statistics, provably not replacing or weakening the outcome-level gate.

Preprint-reproduction rule

D²ACCI's reported numbers are hypotheses, not the acceptance bar. With no upstream code to check against, the reproduction is entirely from the paper description — and promotion still requires an independently recomputed research-gate paired-bootstrap delta over the pre-WP baseline, never the paper's own figures.

Scope / files

crates/ruvector-agent-memory/src/{ledger.rs, observation.rs, fusion.rs, scoring.rs}, plus scripts/research-gate/research_gate.py (new diagnostic-trace consumer, layered under the existing paired-bootstrap statistics).

Depends on

WP4 (#840, TARL ledger — ledger.rs, observation.rs must exist) · WP18 (#865, MemFuse causal-graph fusion — fusion.rs)

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.

Start with crates/ruvector-agent-memory/src/{ledger.rs, observation.rs, fusion.rs, scoring.rs} and scripts/research-gate/research_gate.py, then read ADR-307, ADR-320, ADR-282 and ADR-306 to understand the existing promotion layers. Done means all five stages are independently localizable, every promotion record carries DCR, and the diagnostic layer remains underneath the existing paired-bootstrap gate without weakening it.