grust-ladybug: bulk loads COPY from a temporary CSV; lbug 0.20.4

[alexy]

Stacked on #6 (turso-mvcc-concurrency); only the last five commits are this PR.

Why

LadybugDB never loaded cit-Patents in the strain harness: the load did not finish inside the 7,200 s budget at 9.7 GB RSS (2026-09-12-quegee), while the maintainer loads the same graph from CSR parquet in about five seconds. The engine was not the problem; the adapter's bulk path was.

bulk_load.rs registered each chunk's rows as an Arrow table and ran COPY t FROM (MATCH (a)-[r:scratch]->(b) RETURN …). Measured with the new copy_bench example (eigen, 1 M nodes / 5 M edges, lbug 0.20.4):

path	nodes	edges
registered Arrow + COPY FROM (MATCH …) (before)	1.05 M rows/s	4.1 k rows/s at 500 k; unfinished after 132 min at 5 M
COPY t FROM <registered arrow table>	binder rejects	binder rejects
COPY t FROM 'file.csv', this adapter's schema (STRING keys, JSON props)	2.2 M rows/s	1.28 M rows/s (5 M in 3.9 s)
file COPY, INT64 keys, no props	7.7 M rows/s	6.1 M rows/s
CSR registration (UINT64 offsets) + COPY FROM (MATCH …)	–	231 k rows/s

The relationship probe is superlinear in the engine; the engine's file COPY takes the same rows about 300× faster. The lbug Rust crate has no equivalent of the Python binding's COPY t FROM $df.

The second cost was the per-chunk read-back of every existing node id and (from, to) pair, kept to preserve put_graph's upsert semantics; the harness feeds large graphs in 5 M-edge chunks, so that read-back grew with every chunk and is where the 9.7 GB went.

What changes

• grust-ladybug: every bulk write (nodes, edges, the metadata index) is written to a temporary CSV with Arrow's writer (RFC 4180 quoting, ESCAPE='"') and loaded with COPY t FROM 'file'. A count query skips the read-back when the target table is empty, so a fresh load costs nothing beyond the copy. Upsert semantics are unchanged: rows whose id or pair already exists still take the per-row MERGE path.
• lbug 0.20.4 (crates.io, 2026-09-10; the Rust API is identical to 0.20.2).
• examples/copy_bench.rs: the measurement above, VARIANTS=ABCDE cargo run --release -p grust-ladybug --example copy_bench -- [nodes] [edges].
• Test: a round trip of properties with commas, double quotes, newlines, backslashes, a quote-only string, an empty string, unicode and nested JSON, then a second put over the same ids that must merge, not duplicate. 19/19 crate tests pass (eigen).

Not in this PR

• The read-back still runs for non-empty tables, so a multi-chunk load pays it from the second chunk on. The harness chunk size (AG_CHUNK_EDGES) is the lever; a per-load session that remembers what this process copied would remove it.
• A direct CSR / parquet import path (the maintainer's icebug-disk format, native in 0.21.x) would be a different load path and, in the harness, a separately tagged profile.
• The workspace gate still excludes grust-ladybug on Linux because of the bundled-zstd duplicate symbols when linked next to zstd-sys; unchanged by 0.20.4.

The strain harness pins this head (6af2006) for the Ladybug ladder that follows the Turso, memory and LanceDB reruns.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Sak2FQcdgs5pkL2ruKrSUZ