benchmarks: the pass-2 roster counts one engine version in its three spellings (#258) - #259

…s, in a layer of its own (DECISIONS #258, #259)

OfficialSession (lean_http.py) against arcadedb-driver 0.4.0, read from the wheel (sha256 00fd23e3...):
- db handles are cached per (root, auth, timeout, database); srv.db() per call threw away the per-database URL cache that
  arcadedb-drivers PR #168 (our #108 finding) relies on.
- the per-call timeout was accepted and dropped (driver default: no timeout). The driver has no per-call timeout, only
  ArcadeDBServer(timeout=httpx.Timeout), so each distinct timeout gets a server of its own (the lanes use 30, 60, 300, 600). httpx timeouts
  surface as lean_http.Timeout, other transport errors as ConnectionError, as for LeanSession.
- command with limit goes through the driver (0.3.0 added it; present in 0.4.0).
- OFFICIAL_PIN is 0.4.0 and OfficialSession refuses any other installed version, so a stale image cannot run 0.2.0 under a 0.4.0 pin.

The image: the served ArcadeDB client cells ran in the shared dbbench:client. Moving arcadedb-driver in PKGS[client] would change that
build arg, rebuild the image under a new digest, and split every comparator's measured rows (the stage scripts run build_images.sh on every
stage). So PKGS[client] keeps 0.2.0 (commented) and the driver arrives in dbbench:client-arcadedb: Dockerfile.client-arcadedb, FROM
dbbench:client plus one pip layer of ARCADEDB_DRIVER_PKGS (driver, httpx, httpcore, h11, anyio, idna, certifi, attrs, typing-extensions,
websockets, all ==, PyPI latest of 2026-11-07). runner.py points every served arcadedb arm (client_server, name arcadedb*) at it; e4 and the
embedded arms keep dbbench:arcadedb. engine_pin.FLOATING_OK knows the name; make_2611_stages derives the target from BACKENDS.

Docs: pin strings 0.2.0 -> 0.4.0 in COMPARATORS.md, CAMPAIGN.md row 82, PROTOCOL.md (labelled: official for /query and /command, lean for
bulk loads and streams), the stamp in three tests and the publish fixture rows. COMPARATORS.md: the two stale 'newest stable' sentences
(pymongo, SurrealDB SDK) are replaced by a Client holds table (each lagging client, newer version and PyPI date, why held).

Tests (stubbed driver and httpx, no network, no docker): handle caching, timeout forwarding, limit path, timeout and transport errors, the
pin and the refusal, the registry swap, the build script text, one build of the layer under a fake docker.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…lient each served row used (DECISIONS #256, #258, #259)

export_web: one global condition, generated from engine_commit_equiv.json when a printed row carries server_engine_commit_equiv (release pair, both commits, file and line counts from the registry; a row claiming an equivalence the registry does not hold refuses the export). The served-client sentence now says the official driver is the headline for query and command statements, that bulk loads, the probe, transactions and streams stay on the lean client in the headline rows too, and names the lean and PostgreSQL wire twins only from their rows. Only the page generator, its new test and PAGE-SPEC.md change: nothing the runner measures.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>