[tae898]

Closes #577

This implements DECISIONS #268 and #285 in the generators. Every served ArcadeDB headline cell now comes from the benchmark's own lean http.client client, and the vendor's driver is neither printed nor named on the page or in the paper tables.

What changed:

• make_paper_tables.load_canonical sets the driver's rows aside before any dedupe (by role headline or a client name containing arcadedb-driver) and keeps the lean rows. runs_paper.csv, the paper tables T2 to T5, the gates and the claims therefore hold lean rows only. The include_twins switch and the role in the dedupe key are gone. The dense multipass overlay builder applies the same rule to a driver overlay file.
• export_web.py loses the twin split, _attach_twin, client_twin and the twin_rows parameters. _arcadedb_client_note now names lean_http (http.client, keep-alive) with its source benchmarks/experiments/lean_http.py (also added to source_urls of every table that prints a served ArcadeDB row) and says the other engines are measured through their vendors' own Python clients, so client weight differs between the served engines. The exporter refuses a frozen row, a table, or a multipass overlay file that names another client.
• The arm without a lean job, arcadedb_imgdefaults_server on the documents transaction table, has no cell any more. The table says it was not measured through the benchmark's own HTTP client and declares the absence. I did not invent a number for it.
• page_check.py has a banned-term gate for arcadedb-driver and official ... driver/client in any payload string, with a REMOVE WHEN note (upstream ships the transport fix, ArcadeData/arcadedb-drivers /next: the dense table (l3d) cannot land: page_check and the generators refuse the October route #188 merged and released).
• test_export_client_twin.py is rewritten for the new behaviour, test_publish_2611_e2e.py and fixtures/publish_2611/build_fixture.py follow (the fixture keeps both clients in the raw rows, one arm with driver rows only, and the committed runs.jsonl.gz is regenerated with --patch).
• PROTOCOL.md, CAMPAIGN.md (rows 72 and 82), COMPARATORS.md and the lean_http.py docstring describe the lean client as the headline without naming or comparing the driver. The pin string in COMPARATORS.md stays because the client image still installs the package.

Which numbers move: every served ArcadeDB cell (documents, graph, sparse, dense, time series, cross-model, lifecycle, restart) is now the lean client's median. Latency columns drop and throughput columns rise on the one-row paths, memory and CPU columns of the served arms change, best-in-column marks and rank statements can change, and claims_check.py pins and the paper's prose numbers need a re-read after the next freeze. I have not run the freeze on real rows. The default of lean_http.Session() is unchanged, so a running pass does not split its rows.

Nothing of the page is published yet. Not in this PR: runner.py, the harness, the bindings guide table in bindings/python/docs/guide/server.md, and the public methodology line in the dbbench repo.

🤖 Generated with Claude Code