research: add fail-closed MemorixBench-Transfer artifact

[AVIDS2]

Scope

• Adds the fail-closed MemorixBench-Transfer research artifact, public fixtures, reproducible cohort summaries, and analysis tooling.
• Adds controlled Workset delivery profiles for preregistered product-surface ablations.
• Adds a development-only real Claude Code hook-capture controller with isolated settings/home, exact precursor-patch attestation, exact single-model routing, and strict PostToolUse Edit/Write capture validation.
• Documents a local Pi Coding Agent product-usability observation without promoting it into the benchmark or a product-effect claim.
• Adds anonymous NIER paper source and an allowlisted anonymous supplement builder.

Evidence boundary

The committed public cohorts are descriptive only. They do not establish a general Memorix efficacy claim, a product ranking, cross-agent transfer, or a completed private-oracle study. The local Pi observation is usability feedback only; local headless Claude candidates that did not emit the required hook event produced no admitted result. research/EVIDENCE-STATUS.md and research/CLAIMS.md make the boundary explicit.

Verification

• npm run lint
• npm test
• npm run build
• uv run --directory research pytest -q
• frozen 12-case registry validation
• source and materialized public-artifact audits

Remaining external gates

Independent human case review, a KVM-backed worker/vault deployment with trusted relay/runtime attestations, and any venue portal submission remain outside this draft.