[datlechin]

Summary

• Firebase request lines read back per Unicode scalar, the unit they are written in. A grid delete or edit of a top-level node or root collection whose name starts with a combining mark (\u{FE0F}general) no longer runs against general.
• Query names and values (documentId, updateMask) that start with a mark read back whole.
• Percent decoding keeps a leading U+FEFF instead of going through Foundation, which drops it on some systems.
• A leading { in a path is escaped, so a QUERY path is never read as its JSON body.
• Every / split in the Firebase core goes per scalar: RTDB paths and URLs, Firestore paths and document names, collection group targets, __name__ checks in the driver, IN lists.
• A Firestore update mask whose field starts with a mark no longer parses to an empty mask, which sent a PATCH that replaced the whole document.
• The query classifier reads Firebase text with the driver's parser (FirebaseStatement.parse), so PUT \u{FE0F} null is destructive, as the driver runs it. The app links TableProFirebaseCore for this, as it does the Weaviate and Qdrant cores.
• The classifier treats .priority and other . keys as metadata and unwraps .value, as the Realtime Database stores JSON, so PUT users {".priority": 1} ranks with DELETE.

Test plan

• FirebaseStatementTests (TableProFirebaseCoreTests): round trip of names with leading combining marks, variation selectors, prepend characters, ZWJ and BOM, query items, QUERY paths, update masks, IN lists, Firestore and RTDB paths and request URLs. Fails on main.
• QueryClassifierNonSqlTests: the classifier on the same names, and metadata-only bodies. ExternalStatementGateTests: those writes are refused where deletes are. Fail on main.
• Existing Firebase classifier and gate tests pass unchanged.