[iHow1]

What this changes

Align the stable 0.1.0 security policy and three developer evidence harnesses with the existing tier-aware persistence and recall contract. No product source, version, dependencies, or runtime defaults change.

• SECURITY.md: supported release channels, default governed persistence, verified/unverified/flagged tiers, explicit review and durable-write modes, and honest local trust boundaries.
• activation-proof: establish an isolated Cursor MCP connection before requiring healthy diagnostics; assert reachability without claiming host UI activation.
• recall-quality: explicitly stage reviewed candidates once; retain off-topic and stale-memory assertions and report the current default-recall comparison.
• autopromote-precision: retain all 14 safety fixtures, 4 precision-ceiling fixtures, and 18 coverage fixtures; check tier decisions and real persistence separately.
• bench/README.md: remove obsolete binary allow/reject and usefulness claims.

Risk and verification

Bounded documentation/manual-harness repair under docs/release-governance.md. The product src tree and package identity are unchanged from base ee9bd24. No additional full local product suite is claimed; repository CI remains the baseline gate.

Local results on the exact five changed file hashes now in b496ea0:

• npm run build; npm run typecheck; npm run secret-scan: PASS.
• 53 targeted tests (auto-promote, governance-safety, recall-quality, prompt-recall-calibration, package-completeness, release-evidence): PASS.
• node scripts/activation-proof.mjs: PASS.
• node --experimental-strip-types bench/recall-quality.mjs: PASS (0/4 off-topic leakage; stale fixture excluded).
• node --experimental-strip-types bench/autopromote-precision.mjs: 14/14 tier/persistence contracts PASS.
• trust-eval: 9/9; evaluation-bench full: 84 cases, quality and safety gates PASS.
• Fresh tarball install: packaged proof PASS, benchmark 11/11, isolated MCP connection and redacted doctor diagnostics PASS; zero production dependencies.

Clean committed artifact manifest: HEAD b496ea0, tree a4e75c2f6d30edd3fc0a8c6513e89cafc2551d2b. Tarball SHA-256: 83adbf194ca7171093b2e035daddd14cc98ff54b7356e1d4ae97f999614feadf, identical to the locally installed and exercised development artifact. A clean artifact manifest is not publication authorization or a merged integrated RC.

Checklist

• npm run build passes locally
• Full local suite (not rerun for this bounded repair; 53 affected tests passed and PR CI runs npm test)
• Packaged governed-loop proof passes
• No customer data, private memory content, real local paths or credentials in changed files
• Synthetic data only in fixtures
• No private product code or internal strategy
• DCO sign-off on the commit
• License and trademark boundary unchanged
• No product behavior change requiring a changelog entry

No tag, npm publication, website deployment, or live-runtime upgrade is included.

[iHow1]

Independent agent review completed for the five-file patch at b496ea0: no actionable findings. The review checked isolation, preservation of original fixtures, tier/persistence assertions, and alignment of SECURITY.md with existing implementation. This is an AI-assisted local review, not a GitHub human approval.

Artifact integrity: committed tarball is byte-identical (cmp exit 0) to the previously installed and exercised tarball; all six checksum entries pass. SHA-256: 83adbf194ca7171093b2e035daddd14cc98ff54b7356e1d4ae97f999614feadf.

CI run: https://github.com/iHow1/ihow-memory-core/actions/runs/33975481338