docs: Prometheus /api/v1 paths, native image CPU baseline, bounded vector.fuse/rerank (#509, #501, #494, #498) - #510

Closes #509, closes #501, closes #494, closes #498.

#509: Prometheus endpoints need the /api/v1 prefix

The Prometheus paths were documented as /ts/{database}/prom/... and the server serves them under /api/v1/ts/..., so every one of them 404'd. Fixed in concepts/timeseries.adoc (both prometheus.yml snippets, the PromQL, labels and series examples, the endpoint table). The same mistake turned up in the time-series tutorial (two curl calls), reference/time-series/chapter.adoc, and an upgrade.adoc note that wrote POST /ts/query and GET /ts/latest. I checked the routes in HttpServer.java: the prom handlers are registered in the same /api/v1 block as /ts/{database}/write.

#501: minimum CPU for the native image

native-image.adoc gets a "CPU Requirements" section. It lists the baselines (x86-64-v3, armv8.1-a), the features they need, and the CPUs and VM CPU models they rule out. It also quotes the startup error, explains how to check a host, and gives two ways out: the JVM image, or a build with -Dnative.march=compatibility. The build section now shows the --march / -Dnative.march overrides and warns against -march=native. install-docker.adoc links to the new section. Source: the "CPU requirements" section of the engine's docs/native-image.md.

#494: cost of vector.fuse / vector.rerank

Neither function had an entry in reference/extended-functions/vector.adoc, so I added both. The vector.fuse entry has the full options table, including limit, which was not documented anywhere. Both entries say what drives the cost and what changed in 26.11.1.

concepts/vector-search.adoc has a new subsection, "Bounding the cost of a full-text leg", with the bounded form: rerank(fuse(neighbors, SEARCH_INDEX subquery LIMIT :fetch, { limit: :candidates })). The existing three-way example and the two hybrid examples in full-text-index.adoc now put LIMIT 100 on the full-text leg.

I checked the text against SQLFunctionVectorFuse and SQLFunctionVectorRerank at engine 9d9aa076ab, and ran each query shape on arcadedata/arcadedb:26.11.1-SNAPSHOT (built 2026-10-26): the bounded nested query with parameters, a direct SEARCH_INDEX source with fusion: 'LINEAR', fuse with limit, and the rerank example. The same run confirmed that a SEARCH_INDEX subquery with no ORDER BY returns matches by descending $score, so a LIMIT on it keeps the best ones.

#498: Cypher and Gremlin on TIMESERIES types

Nothing left to do. The Query Languages table in concepts/timeseries.adoc, added in 8f6e139, already lists SQL, PromQL, MongoDB QL and GraphQL as readers and says Cypher and Gremlin refuse a TimeSeries type. This PR just closes the issue.

Also in this PR

The typos pre-commit hook rewrote ANDed to "ANDead" in two existing lines of full-text-index.adoc as soon as that file was staged. _typos.toml now whitelists ANDed as an identifier.

Checks

docs-validator.py reports 3 broken refs, the same 3 as main. The migrate.sh + Antora build adds no warnings for the touched pages, and the new cross-page links (#hybrid-search-bounded, #native-image-cpu-requirements) resolve.

🤖 Generated with Claude Code