v0.2.2

LLM callout opt-in, scope confirmation, and cost transparency.

Added

• --enable-embeddings and --enable-enrichment CLI flags; both tools are off by default and must be explicitly opted in via mcp.json args.
• docgraph_enrichment action=pending returns a pre-written ASCII scope tree with per-folder file counts, token estimates, per-model cost table, and sensitive path flags before any writes occur.
• docgraph_embeddings action=pending now generates a one-time confirmation token; action=store requires the token before processing.
• engine= parameter for docgraph_similar: auto (default), tfidf, or neural. Neural mode returns a clear error when embeddings are not enabled.
• docgraph_status now includes a LLM Callout Tools section showing enabled/disabled state and pending counts for both tools.

Security

• Vector payload cap (2 MB) enforced before deserialisation in docgraph_embeddings action=store.
• Enrichment metadata payload cap (1 MB) enforced before deserialisation in docgraph_enrichment action=process.
• Confirmation tokens use crypto/rand (16 bytes, 32-char hex); embeddings and enrichment tokens are kept in separate stores to prevent cross-tool reuse; 30-minute TTL with lazy sweep.
• When 100% of pending documents are in sensitive-flagged paths, no token is generated and action=store/action=process are rejected.

Breaking Changes

• docgraph_embeddings and docgraph_enrichment are no longer registered by default. Add --enable-embeddings / --enable-enrichment to mcp.json args to restore access.
• docgraph_enrichment API changed: operation=pending/store → action=pending/process; action=process requires a confirmation_token from the preceding action=pending call.