docs: reference the 0.22.0 train (core 0.22.0, drivers/sidecar 0.3.1, installer pin v0.22.0)

[fupelaqu]

Post-release doc sweep — 0.22.0 train

Points every version-bearing surface at the just-released train: core 0.22.0 · extensions 0.3.1 · arrow 0.3.1 · jdbc 0.3.1 · repl bundle 0.22.0. Companion sweep: softclient4es-web#47.

Verified published BEFORE writing, not after

Every version this PR writes was probed on JFrog first — the 12 per-ES jdbc-driver / adbc-driver / arrow-flight-sql jars at 0.3.1, plus softclient4es-jdbc-driver_2.13, community-extensions_2.13, arrow-extensions_2.13 at 0.3.1 and softclient4es8-java-client_2.13 at 0.22.0. All 200.

The failure this guards against is silent: a version with no -all bundle makes the installer fall back to the plain artifact, which carries no arrow extensions and therefore no cross-index JOIN — the docs would demonstrate the path that loses the headline feature.

Swept

README.md (driver table, Maven/Gradle/sbt coordinates, the Scala dependency block) · documentation/client/{jdbc,adbc_driver,arrow_flight_sql}.md (jar tables + coordinates) · documentation/client/repl.md (install examples) · documentation/client/download_analytics.md · install.ps1 (help example) · install.cmd.

install.cmd — the release ritual, both halves

PS1_REF v0.21.0 → v0.22.0 and PS1_SHA256 recomputed together, exactly as the RELEASE RITUAL comment above them requires: a bumped tag with a stale hash is a hard failure for every install.cmd-only user.

The new hash is the SHA-256 of install.ps1 at tag v0.22.0. The method was validated first by recomputing v0.21.0's hash, which reproduced the currently-pinned e1275bd2…03912ae exactly — so the new value is right by the same procedure, not by assertion.

The file is CRLF by .gitattributes (-text). My first attempt rewrote it in text mode and silently converted all 165 line endings, showing up as a whole-file diff. Redone in binary: CRLF count 165 before and after, zero bare LF, and the diff is exactly 3 lines.

The --list-versions capture GAINED a line

repl.md's captured installer output is a quoted product output, and the sweep's first pass renamed • 0.21.0 → • 0.22.0 — which would have told users 0.21.0 no longer exists while leaving Total: 5 stale.

It is now re-captured for real (bash install.sh --list-versions --es-version 8, ANSI stripped) and the committed block is byte-identical to the live output: 0.20.1 … 0.21.0, 0.22.0, Total: 6. Verified against the artefact, not reasoned about.

Deliberately NOT bumped — these date a feature, they don't track the current version

file	line	text
scroll.md	51, 147, 319	sliced PIT (0.21.0+) / "Default since 0.21.0"
common_principles.md	491, 538	"Paged row extraction … 0.21.0+"
repl.md	54	Windows Java bootstrap "Since 0.20.4"

Bumping any of these tells a 0.21.0 user they lack a feature they have. Worth noting for the next sweep: the X+ spelling does not match a /since|before/ guard — the guard has to catch the family, not one wording. These were caught by reading the survey, not by the regex.

Verification

• 12 + 4 JFrog probes, all 200, before any edit.
• Quote-balance check on the diff (git diff | grep '^+' | awk '{n=gsub(/"/,...)}') — clean; a wrong capture group in a scripted sweep silently eats a closing quote.
• git grep for stragglers across all tracked files (excluding blogs/issues/bmad): zero remaining 0.21.0 / 0.3.0 artefact references.
• The committed --list-versions block diffed against live output: identical.

🤖 Generated with Claude Code