[tmm]

Fixes a bug found by the bug sweep, with option (a) chosen by the user.

Bug: JSONValue.object was a [String: JSONValue]. Swift strings compare by Unicode canonical equivalence, so keys like "\u00e9" and "e\u0301" merged and a value was lost. That dropped data from pokes (row values, primary keys, update merges, client IDs) and gave custom-query and AST hashes that differ from JS.

Fix: a new public JSONObject value type (case object(JSONObject)):

• keys compare and hash by UTF-16 code unit;
• keys keep JS order: integer-like keys ascending, then insertion order; the parser keeps the last value of a duplicate key in the first key's place, as JSON.parse does;
• the serializer writes keys in that order, matching JSON.stringify; query hashes still sort keys, as JS does;
• dictionary literals, subscript(key), ordered iteration, mapValues, merge and .dictionary keep it ergonomic. JSONValue stays 17 bytes.

Breaking (minor changeset, with a migration note): objectValue, ZeroObjectReader/Writer.object and AST.Bound.row are JSONObject; .object(dictVariable) needs JSONObject(dict). Mutator and start APIs still take dictionaries. zbugs needed a one-line change.

Benchmarks (release, Linux): no regression; materialize 100k rows 7.61 → 6.20 ms, put 1000 changed rows 9.41 → 8.60 ms, memory for 100k rows 127.7 → 126.2 MB.

Checked: 173 Swift tests including all conformance vectors and 6 sweep wire tests generated from JS, swift format lint --strict, ClientE2ETests against a real zero-cache, and zbugs-swift builds its tests. TestHost can't build on Linux; macOS CI covers it.