docs: add REVIEW.md, the repo-local review prior

[minerva-sky]

One short file at the repo root that tells a reviewer (human, panel seat, or an agent on a tick) what a bad merge looks like in this repo, before they read the diff.

Six failure shapes, each grounded in an issue or PR this repo has already had:

• Sanitizer regex drift (fix(security): keep UUIDs intact through Sanitizer phone redaction #54)
• Secrets steered onto disk (CLI and README steer the OpenAI key into $PWD/.agentic.yml, which nothing reads and nothing gitignores #59, CLI stops steering the OpenAI key into .agentic.yml (#59) #66)
• Writes that escape the workspace (security: guard PersistentAgentStore path components and close AgentCapabilityRegistry#find open dispatch #28, security: keep Workspace artifact writes inside the workspace (symlink containment) #30, Rescue SecurityError alongside StandardError around artifact writes and task execution (#31) #50)
• Works in the dev bundle, breaks on require "agentic" (lib/agentic.rb forbids require_relative between siblings; 61 lines in 27 files do it anyway, all reintroduced by #7 #57, require "agentic" can't load ArtifactGraph or AuthenticationSystem outside the dev bundle: rgl and base64 are required by lib/ but declared nowhere a consumer sees #63, fix: ArtifactGraph on stdlib TSort; declare every runtime dependency lib/ requires #74, Load lib/agentic through Zeitwerk alone; drop 57 require_relative lines, add loader spec #78)
• One-Ruby green across the 3.2 to 4.0 matrix (deps: raise Ruby floor to >= 3.2 and test a real CI matrix (#15) #27, fix: test suite loads on Ruby 4.0; de-flake EventDispatcher timing spec #33, json 3 raises on duplicate keys in model output: adopt strict parsing as policy, guard the journal writer #76)
• Documented, wired to nothing (SecureErrorMixin#secure_error_class! has zero callers: LlmError#message and #inspect are never sanitized despite the mixin docs #60, PlanOrchestratorConfig has no reader: nothing accepts it, its spec tests the orchestrator, and its defaults describe an orchestrator that does not exist #69, Learning::CapabilityOptimizer: 613 lines in the 0.2.0 release notes that Learning.create never builds and no spec has ever run #70, CriticFramework and AdaptationEngine: the Verification Layer's feedback loop is checked off in the architecture doc and wired to nothing #71)

Plus a path:line evidence rule, a three-nit cap, and the skip paths (CHANGELOG, lockfile, coverage and run artifacts, examples).

Same shape as codenamev/ai-software-architect#129. Docs only, no code or CI change. Edit anything that reads wrong; the file is the point, not my wording of it.

🤖 Generated with Claude Code