[tmm]

Large views, as the Space goal asks: store a row's columns by its table's layout instead of a per-row copy of every column name.

• A json::Object keeps its keys apart from its values (Arc<Keys> + Arc<[Value]>), and the rows of a table share one key list, their layout. This is the Rust counterpart of Swift keying rows by column index and of JS hidden classes.
• The parser follows a recent object's keys while the next object has the same ones, comparing raw key bytes, so a synced row allocates neither its key strings nor their list. On a mismatch it switches to a recent layout with the same keys so far, so tables that share a first column (id) still share.
• The poke handler and mutators put rows in their table's layout (Layouts, shared by a branch and its forks).
• Comparators and filters remember where their column was (get_hinted), and check the key there before using it.

Public API is unchanged: Object and OrderedMap keep their methods, and the new ones are pub(crate).

Numbers (medians of 3 interleaved runs on one 8-core Xeon orb)

	Rust before	Rust after	Swift (same machine)	JS
memory, 100k rows + view	181 MB	83 MB	127 MB	239 MB
materialize 10k rows	678 µs	565 µs	863 µs	2.06 ms
materialize 100k rows	20.9 ms	10.5 ms	16.8 ms	30.1 ms
100k rows as structs	48.3 ms	23.8 ms	37.4 ms
poke of 1000 rows	7.74 ms	5.01 ms	10.8 ms
first sync, issue list	17.0 ms	15.3 ms	42.9 ms
poke of 1 row	11.4 µs	12.4 µs	49.1 µs
optimistic update	15.3 µs	15.9 µs	39.5 µs	274 µs

At 100k rows, the cost per row compared with 10k went from 3.1× to 1.9×, about Swift's 2.0× on the same machine. It doesn't scale linearly yet.

Caveat: swift/BENCHMARKS.md's Linux column has Swift 2–3× faster than it ran here (7.07 ms for 100k rows), while JS ran exactly as fast as documented. Against that column, Rust is behind Swift at 100k rows (10.5 vs 7.07 ms) and level on structs (23.8 vs 24.8 ms). It's written up in rust/BENCHMARKS.md.

Small costs: a 1-row poke is about 1 µs slower, because sharing a row's layout compares its column names. pk = N, an optimistic update and a round trip are 3–4% slower. A table's layout is one stored row per table. Locally inserted rows whose column order differs from the server's switch the layout back and forth: results stay correct, and those rows share less.

Checks

• Conformance: 63,453 cases, 0 failures, 2,255 under documented exceptions (unchanged).
• Locally: fmt, clippy -D warnings, workspace tests, docs with -D warnings, and MSRV 1.88 all pass.
• New unit test: parsed objects with equal keys share them. It covers shorter, longer and reordered objects, a repeated key, an escaped key equal to a raw one, and integer-like keys.

Adds a zero-rust patch changeset. Releasing it needs tmm.

Amp-Thread-ID: https://ampcode.com/threads/T-01a1c6b9-fc99-73a1-bf70-e71fd484042c