[haolpku]

DataMind can leave stale graph relationships after a document edit, and successive tool calls can read different source generations while an update is publishing. StoreAgent also needs ingestion guidance with durable attribution of the Skill it followed.

This change builds on the optional backend/versioning support merged in #28:

• Store typed graph evidence with source identity, content version, locator and extraction method. Replace complete document revisions; reconcile deleted sources and preserve previous edges if extraction fails. Retain immutable input evidence for inline text and explicit assertions.
• Publish immutable workspace generations using copied sources, file-level compute checkpoints and atomic head updates with a monotonic epoch. Pin KB/KG/DB/raw reads to one generation; support retained-generation rollback, bounded transient read retries, concurrent publication conflict detection and recovery after process termination.
• Use a SQLite receipt ledger for durable records and cross-worker revision allocation. Add workspace-ingest/source-update Skills, allowlisted procedural/input reads for StoreAgent, Skill version/content-hash attribution, and explicit local Codex workspace update/query tools.
• Include workspace consistency, failure recovery and Skill integration tests in the required CI gate introduced with feat: KB/DB versioning and CocoIndex-based incremental workspace ingest #28.

Validation: 425 local tests passed with real Dolt; 45 local tests skipped where MySQL/other external services were unavailable. Required cloud checks run Python 3.11/3.12 package and local Codex plugin suites, 62 optional-provider/publication/recovery contracts, 130 real MySQL/Dolt tests with no skips, and an isolated installed-wheel smoke test. The aggregate gate requires every job to succeed.

The snapshot coordinator supports concurrent readers/builders on one host with local NetworkX, SQLite and Chroma/LanceDB. It reuses file-level model work and materializes a new full index snapshot. Remote snapshot adapters, multi-host deployment, automatic snapshot GC and exactly-once external append transactions remain outside this implementation. Published workspace queries use the explicit API; live-profile APIs retain their existing semantics.

Live Agent SDK/model integrations, enterprise load tests and research quality/cost experiments have not been run. See docs/KNOWLEDGE_LIFECYCLE.md and docs/TESTING.md for API examples, supported deployment boundaries and recovery granularity.