diff --git a/docs/core/hardening-review-v3.md b/docs/core/hardening-review-v3.md
new file mode 100644
index 0000000..1303e74
--- /dev/null
+++ b/docs/core/hardening-review-v3.md
@@ -0,0 +1,339 @@
+# Betterbase Backend/Core Hardening Review (v3)
+
+## 1) Objective (final outcome)
+
+Deliver a **backend-first security and reliability hardening baseline** for Betterbase, focused on:
+
+- `packages/server` runtime and auth boundaries,
+- database/migration safety,
+- Docker/container defaults and edge proxying,
+- operational protections for production readiness.
+
+> Out of scope by request: dashboard/frontend quality review.
+
+---
+
+## 2) Scope map (what is reviewed)
+
+### Included
+- API/auth/device/admin flows (`packages/server/src/**`)
+- Migration and schema lifecycle (`packages/server/src/lib/migrate.ts`, `packages/server/migrations/**`)
+- Container/runtime and local infra defaults (`Dockerfile`, `docker-compose.yml`, `docker/nginx/nginx.conf`)
+
+### Excluded
+- Dashboard UX, styling, and frontend correctness (`apps/dashboard/**`)
+
+---
+
+## 3) Plan (how this review is executed)
+
+1. **Synchronize branch state** before any edits (fetch/pull when tracking exists).
+2. **Static backend attack-surface review**:
+   - auth/session/device/token issuance,
+   - admin routing and privilege checks,
+   - secret handling and env defaults,
+   - SQL use patterns and dynamic identifier safety.
+3. **Docker and reverse-proxy hardening pass**:
+   - insecure defaults,
+   - exposed ports/services,
+   - transport/auth boundaries.
+4. **Reliability pass**:
+   - migration atomicity,
+   - logging correctness,
+   - operational behavior under failure.
+5. **Produce v3 documentation artifact** with severity, impact, evidence, and fixes.
+
+---
+
+## 4) Acceptance criteria vs non-acceptance criteria
+
+### Accepted criteria
+A review is accepted only if it:
+
+- Prioritizes **critical/high backend risks first**,
+- Lists concrete findings with **file-level evidence**,
+- Proposes remediation that is actionable and ordered,
+- Separates scope (core/backend) from excluded dashboard concerns,
+- Defines objective pass/fail criteria for closure.
+
+### Not accepted
+A review is not accepted if it:
+
+- Focuses mostly on frontend/dashboard problems,
+- Gives generic best-practice lists without code evidence,
+- Omits severity and business impact,
+- Doesn’t define what “done” looks like.
+
+---
+
+## 5) Problem classes targeted in this review
+
+- **Authentication hardening** (brute-force, token abuse, device flow safety)
+- **Authorization hardening** (scope enforcement, privilege boundaries)
+- **Input/output security** (HTML injection, unsafe interpolation)
+- **Secrets/config hardening** (weak defaults, exposure risks)
+- **Operational resilience** (migration safety, observability correctness)
+- **Container/proxy hardening** (network exposure, transport/auth controls)
+
+---
+
+## 6) Findings (backend/core only)
+
+## F-01 (HIGH): No brute-force/rate limiting on admin login and device verification
+
+**Evidence**
+- `/admin/auth/login` verifies credentials directly with no throttling/lockout/captcha/backoff. (`packages/server/src/routes/admin/auth.ts`)
+- `/device/verify` form POST verifies admin credentials with no anti-automation controls. (`packages/server/src/routes/device/index.ts`)
+
+**Impact**
+- Enables online password guessing and credential stuffing against administrative entry points.
+
+**Recommended fix**
+1. Add IP + account-based rate limits (e.g., sliding window).
+2. Add progressive delays/temporary lockouts after repeated failures.
+3. Emit structured security events for failed auth attempts.
+
+---
+
+## F-02 (HIGH): API key scopes are stored but not enforced
+
+**Evidence**
+- API keys include `scopes` at creation and persistence. (`packages/server/src/routes/admin/api-keys.ts`, `packages/server/migrations/008_api_keys.sql`)
+- Auth middleware validates key hash and expiration only; it does not evaluate `scopes` for route/action authorization. (`packages/server/src/lib/admin-middleware.ts`)
+
+**Impact**
+- A supposedly limited key can act as full-admin bearer credentials, breaking least-privilege expectations.
+
+**Recommended fix**
+1. Introduce route-level scope requirements (e.g., `requireScopes(["projects:read"])`).
+2. Validate scopes in middleware before allowing handler execution.
+3. Add tests proving denied access when scope is insufficient.
+
+---
+
+## F-03 (MEDIUM-HIGH): Reflected HTML injection risk in device verification page
+
+**Evidence**
+- `userCode` query param is interpolated directly into HTML input value without escaping. (`packages/server/src/routes/device/index.ts`)
+
+**Impact**
+- Crafted links can inject markup/script contexts depending on browser parsing behavior, risking phishing/session abuse in admin verification flow.
+
+**Recommended fix**
+1. Escape HTML entities for all interpolated values.
+2. Prefer server-side templates with automatic escaping.
+3. Add CSP headers and disable inline script where possible.
+
+---
+
+## F-04 (MEDIUM): Migration execution is not wrapped in per-file transactions
+
+**Evidence**
+- Migration runner executes raw SQL then records applied filename, without explicit transaction guards around each migration file. (`packages/server/src/lib/migrate.ts`)
+
+**Impact**
+- Partial migration failures can leave inconsistent schema states that are hard to recover in production rollouts.
+
+**Recommended fix**
+1. Wrap each migration in `BEGIN/COMMIT` with rollback on failure.
+2. Abort startup if any migration fails.
+3. Add idempotency checks for unsafe DDL operations.
+
+---
+
+## F-05 (MEDIUM): Insecure local Docker defaults can leak into non-local usage
+
+**Evidence**
+- Static/dev credentials and wide host port exposure in compose defaults (`5432`, `9000`, `9001`, `8288`) plus default passwords. (`docker-compose.yml`)
+
+**Impact**
+- If reused outside isolated local environments, these defaults materially increase compromise likelihood.
+
+**Recommended fix**
+1. Gate local-only config via explicit profile names and warnings.
+2. Require strong secrets through `.env` for non-dev profiles.
+3. Limit host port publishing to required services only.
+
+---
+
+## 7) Closure criteria for this hardening cycle
+
+To mark this cycle complete, all must be true:
+
+1. Login/device endpoints enforce rate limits and lockout policy.
+2. API key scopes are enforced in middleware/route guards.
+3. Device verification HTML interpolation is safely escaped.
+4. Migration runner supports transactional per-file execution.
+5. Container docs/config clearly separate local unsafe defaults from production-secure defaults.
+
+---
+
+## 8) Implementation priority order
+
+1. **P0:** F-01, F-02
+2. **P1:** F-03
+3. **P2:** F-04, F-05
+
+---
+
+## 9) Summary
+
+This v3 review identifies backend/core hardening gaps primarily in auth abuse resistance, authorization scope enforcement, HTML output safety, migration resilience, and Docker default safety posture. Addressing P0/P1 items first materially reduces compromise risk while improving operational confidence.
+
+---
+
+## 10) AI reviewer operating manual (for the next reviewing agent)
+
+This section is intentionally prescriptive so another AI/code reviewer can continue this work with high consistency and low ambiguity.
+
+### 10.1 Review intent
+
+You are not only finding bugs. You are validating whether Betterbase can safely operate as:
+- an internet-facing backend API,
+- a multi-project control plane,
+- a migration-driven data platform,
+- and a containerized service.
+
+### 10.2 Required review mindset
+
+For every endpoint or subsystem, ask:
+1. **Can an attacker reach this?**
+2. **What identity context is assumed?**
+3. **What authorization boundary is enforced?**
+4. **What happens on repeated abuse/failure?**
+5. **How does this fail in production under partial outage?**
+
+If any answer is unclear, treat it as a review finding candidate.
+
+### 10.3 Evidence standard (non-optional)
+
+Every finding should include all of:
+- exact file path(s),
+- function/route names,
+- exploit or failure path in 2–5 steps,
+- business impact,
+- concrete mitigation order (quick patch + durable fix).
+
+Avoid generic “best-practice only” findings with no file-level proof.
+
+---
+
+## 11) Backend review blueprint by domain
+
+Use this sequence to avoid missing critical areas.
+
+### 11.1 Identity and authentication
+- Admin login (`/admin/auth/login`)
+- Setup flow (`/admin/auth/setup`)
+- Device flow (`/device/code`, `/device/verify`, `/device/token`)
+- Token issuance/validation (`signAdminToken`, `verifyAdminToken`)
+- API key lifecycle and bearer handling
+
+**What to catch**
+- brute force/stuffing paths,
+- token replay windows,
+- missing audience/issuer constraints,
+- no rotation/revocation strategy,
+- secret leakage in logs/errors.
+
+### 11.2 Authorization and tenancy boundaries
+- `requireAdmin` middleware behavior for JWT vs API keys
+- project-scoped routes and project attachment middleware
+- actions that should require scoped permissions but do not
+
+**What to catch**
+- privilege escalation across routes,
+- “scope exists in DB but not enforced” patterns,
+- project hopping / IDOR possibilities,
+- routes inheriting global admin access unintentionally.
+
+### 11.3 Data and SQL safety
+- dynamic schema/table identifier usage
+- raw SQL interpolation (especially schema-qualified queries)
+- migration sequencing and rollback resilience
+
+**What to catch**
+- unsafe identifier injection opportunities,
+- inconsistent transaction handling,
+- ordering collisions in migration naming,
+- partial-apply states causing startup drift.
+
+### 11.4 Output and interface safety
+- HTML-rendering endpoints (device verification page)
+- plain-text/HTML response interpolation
+- error message leakage
+
+**What to catch**
+- reflected/stored injection risks,
+- security header gaps (CSP/XFO/etc.),
+- differential error behavior that aids attackers.
+
+### 11.5 Container, proxy, and runtime posture
+- `docker-compose.yml` defaults
+- nginx routing and trust headers
+- image pinning and exposed ports
+
+**What to catch**
+- weak default credentials reused outside dev,
+- unintended public service exposure,
+- over-trusting forwarded headers,
+- missing TLS/security hardening assumptions in docs.
+
+---
+
+## 12) Severity rubric for this project
+
+Use this rubric consistently for triage:
+
+- **Critical**: direct compromise of admin control plane, tenant data exposure, or remote code/data integrity impact with low attacker cost.
+- **High**: realistic unauthorized access path, major authz bypass, or high-probability account compromise vector.
+- **Medium**: meaningful security/reliability weakness requiring additional conditions.
+- **Low**: hygiene/documentation/observability issues with limited direct exploitability.
+
+When uncertain between two severities, choose the higher one and justify assumptions.
+
+---
+
+## 13) Acceptance gates for a “secure-enough” next milestone
+
+The next implementation PR(s) should not be accepted unless all gates pass:
+
+1. **Auth abuse gate**
+   - login + device verification endpoints have measurable rate limiting and lockout/backoff.
+2. **Authorization gate**
+   - API key scopes are enforceable and tested on representative privileged routes.
+3. **Injection gate**
+   - user-controlled values rendered in HTML are escaped or template-engine protected.
+4. **Migration safety gate**
+   - each migration file executes atomically with rollback on failure.
+5. **Runtime hygiene gate**
+   - local-only insecure defaults are explicitly isolated from production paths.
+
+---
+
+## 14) Reviewer handoff format (for future AI agents)
+
+When handing off to another reviewer/engineer, provide:
+
+1. **Top-5 unresolved risks** (ordered by exploitability × impact),
+2. **Proof snippets** (route/file/line references),
+3. **Proposed patch plan** split by:
+   - immediate hotfixes (hours),
+   - short-term hardening (days),
+   - structural improvements (weeks),
+4. **Regression checks** that must run after each fix,
+5. **Known blind spots** (areas not fully reviewed).
+
+This prevents review drift and keeps findings actionable.
+
+---
+
+## 15) Suggested follow-up implementation roadmap (non-doc work)
+
+1. Implement centralized rate-limit middleware for auth-sensitive endpoints.
+2. Introduce explicit scope policy map and enforce in `requireAdmin`/route guards.
+3. Add safe HTML escaping helper for any server-rendered templates.
+4. Refactor migration runner to transactional execution with deterministic failure logging.
+5. Add a production-hardening compose profile and deployment checklist.
+
+These implementation tracks should be handled in dedicated code PRs with tests.