[wess]

Changes

• Record reported outcomes separately from observed checks and user judgments. Retain failed-check flags across passing retries until explicit review.
• Save project-scoped, reversible drift assessments tied to exact stored content.
• Queue one procedure suggestion after three independent verified sessions. Queue complete skill replacements for review, guard the original fingerprint, preserve edited copies and revision history, and require explicit global approval for shared changes.
• Propose retrieval regression labels for human review, then replay actual scoped recall with quality, noise, context and latency metrics.
• Keep collection local, bounded, event driven, and usable without a GUI. Redact sensitive evidence and prevent private skill text from entering revision history.
• Reserve memory write transactions before index work, fixing a reproduced SQLite snapshot-upgrade failure under simultaneous memory writes and bookkeeping.
• Release desktop and core as 1.14.0. Optional cloud sync remains deferred.

Verification

• Core: 564 tests, no-default-features check, strict clippy across all targets, formatting.
• Desktop workspace: build and 26 tests. Independent server: 18 tests. Bridge install/typecheck. Site build and coverage/link checks.
• Headless CLI/MCP integration covers reports, real checks, timeout handling, project isolation, review and concurrent approval, stale targets, edited installations, revision restore, drift restore, credential redaction, migration, retention, and actual recall evaluation.
• Optimized disposable benchmark: 10,002 memories, 5,000 feedback entries, 2,000 resolution entries; 100 samples. Saturated recall plus drift p95 4.16 ms, feedback p95 0.52 ms, confirmation p95 0.52 ms, peak RSS 21.9 MiB. Reproduce with cargo run --release --example learning --manifest-path crates/synapsecore/Cargo.toml -- 100.
• Idle MCP with learning enabled: eight-second sample, unchanged CPU time and database files, 14.5 MiB RSS. Learning adds no background loop or worker.

Behavior change

revise now queues a proposal instead of mutating installed skills. Its receipt preserves the original fields and adds proposal/state fields; revision remains zero until approval. Review with synapse learn proposal <id>, then apply with synapse learn approve <id> --confirm.