CMI v0.9.2

Fixed

• Enforced the default MCP read-only boundary by removing scan mutation from safe tool discovery and rejecting direct scan calls before generated CMI caches can be written unless write mode is explicitly enabled.
• Made graph freshness cover the complete source candidate set plus resolver, workspace, scan, and ignore inputs so new files and configuration drift block impact analysis instead of returning stale confidence.
• Separated source/project freshness from semantic memory review: source refresh now yields fresh-source, while reviewed-current requires explicit review provenance that remains current relative to source refresh.
• Made unsafe, unreadable, symlinked, or oversized durable memory fail closed as blocked evidence rather than silently appearing empty; search and refresh cannot use blocked memory as trusted input.
• Made an existing invalid, unsafe, or oversized findings registry fail closed with CMI_FINDINGS_BLOCKED so normal finding/session mutations cannot overwrite corrupted evidence.
• Hardened CLI automation contracts: unknown options fail, JSON-mode errors remain machine-readable, and blocked impact exits non-zero while preserving structured output.
• Aligned memory-review JSON Schema conditions with runtime provenance validation.
• Reduced representative heuristic parser errors by ignoring JavaScript import-like text in comments/ordinary strings and resolving reviewed Python relative-import and Rust crate:: cases.
• Aligned cmi doctor with the Unified Evidence Health model so blocked graph, impact, durable-memory, or overall evidence states fail diagnostics instead of being reported as warning/pass states.

Added

• Regression coverage for safe MCP direct-call bypass attempts, source-set and resolver drift, semantic-review separation, blocked durable memory, corrupt findings no-overwrite behavior, strict CLI contracts, graph/schema freshness, parser edge cases, and blocked doctor diagnostics.
• docs/INDEPENDENT_REVIEW_REMEDIATION.md documenting the finding-by-finding remediation contract and the empirical gaps that remain intentionally open.
• docs/EMPIRICAL_VALIDATION.md defining a paired controlled comparison protocol for plain Codex/Git/source search versus Codex + CMI without converting synthetic or caller-attested data into productivity claims.

Evidence limits

• CMI remains heuristic/advisory rather than compiler-grade semantic analysis.
• Evaluation repository/reviewer provenance remains caller-attested unless a stronger external assurance process is used.
• This release improves trust boundaries and testable correctness; it does not claim that CMI productivity value is independently proven across real repositories.