Benchmarks: the served ArcadeDB arms leave -Xms unset, like the embedded arms - #456

Closes #455

The served ArcadeDB arms now pass ARCADEDB_OPTS_MEMORY=-Xmx{heap} instead of -Xms{heap} -Xmx{heap}, so they leave the JVM's initial heap unset like the embedded arms since #429. G1 and the 50% heap budget stay, so both deployments run one JVM configuration again, and it matches the maintainers' guidance for Docker (ArcadeData#9167, arcadedb-docs "JVM settings in a fixed memory budget"). Embedded evidence (ab.py, 10 clean pairs, heap-pin-ab-20261022): unset versus pinned peak RSS about 1.4 versus 2.4 GB, committed heap about 1 GB versus 4096 MiB, total wall pinned 1-7% worse, GC count and time better pinned, vector build pinned 2-9% worse, ingest and graph pinned 2-11% better. The decision is DECISIONS #200.

What changes: runner.py (five served arms, comments), export_web.py (the generated image-defaults sentence reads the main arm's initial heap from the rows, so rows measured before this change keep the old wording and new rows say the main arm sets no initial heap size; the jvm_heap page declaration now says Elasticsearch and Neo4j pin and ArcadeDB, embedded and served, does not; the served main arm is named in the "starts with no initial heap size" memory clause from its own JVM report), fairness_check.py comments, PROTOCOL, FAIRNESS, CAMPAIGN, PAGE-SPEC and the bindings protocol page. PROTOCOL section 7 also gets a table of each JVM engine's heap setting, whether the initial heap is pinned, the vendor source with its sentence, and what the memory column shows. Verified against the current docs: Neo4j recommends initial equal to max, Elasticsearch requires the minimum and maximum to be the same, and QuestDB's configuration page has no JVM heap recommendation, so Neo4j, Elasticsearch and QuestDB are unchanged. Their pinned heaps read toward the cap in the memory columns, which is the cost of the vendors' recommended setups and is reported as such.

Existing rows are not edited; the ArcadeDB served arms run in the Nov 5-12 pass, after this change, so the comparator subset is unaffected. Tests: test_instrument_method.py asserts no ArcadeDB arm pins -Xms and the served arms keep -Xmx{heap}, test_jvm_defaults_arm.py covers both wordings of the generated sentence and every served arm's env, and test_index_before_load_note.py follows the new declaration. The branch is rebased on #461 (Neo4j and Elasticsearch follow their vendors' memory rules, DECISIONS #202), and the "Beyond the heap" rows for the Neo4j page cache and the Elasticsearch heap ceiling now say match.

Untimed served memory check, one container per run, no speed claim (probe and raw rows in the notes repo, DECISIONS #200):

UNTIMED. arcadedata/arcadedb:26.10.2, -m 8g, -Xmx4g, G1, 400000 documents + query mix, MiB, runs in order (median).

JVM read-back (InitialHeapSize, MaxHeapSize, GC, -Xm flags): {'pinned': [('4294967296', '4294967296', 'G1', '-Xms4g -Xmx4g')], 'unset': [('134217728', '4294967296', 'G1', '-Xmx4g')]}
runs ok 6 of 6

Tests, run after the rebase on the head of this branch (.venv python -m pytest, benchmarks/experiments): test_instrument_method.py, test_jvm_defaults_arm.py, test_index_before_load_note.py, test_instrument_selector.py, test_overrides.py and test_strict_durability.py, 232 passed. I did not run the full benchmark suite.

🤖 Generated with Claude Code