Epic — REPL Hardening (post-0.20.1 verification): keyword drift, dead config, first-JOIN stall, -all bundle

[fupelaqu]

Tracking issue for the REPL Hardening maintenance epic (post-0.20.1 verification pass). Sources: #161, #162, #163.

Stories & PRs

Story	Issue	PRs	Status
REPL.1 — SQL keyword single source of truth (highlighter ⇄ completer ⇄ parser)	#161	#164	done (draft PR)
REPL.2 — connection config precedence: flag > env > file > default	#162	#165	done (draft PR)
REPL.3 — async DuckDB warm-up + eager extension init	#163 (fixes 1+2)	#166 + softclient4es-arrow#128 (issue softclient4es-arrow#127)	done (draft PRs)
REPL.4 — self-contained `-all` cli bundle via softclient4es-repl + installer rework	#163 (fix 4)	#167 + softclient4es-repl#2 (issue softclient4es-repl#1)	done (draft PRs)
REPL.5 — batch/cold-start spike (AppCDS)	#163 (fix 3)	in progress	running

Key outcomes so far

• First JOIN now ~1× subsequent ones (34–66ms across ES6/7/8/9) vs 2431ms / ~28× baseline; DuckDB native load moved to a logged daemon-thread warm-up.
• Extension discovery+init ≈54ms on the single `-all` jar vs 442ms on the 279-jar classpath.
• `-all` bundle E2E smoke green on ES 6.8.23 / 7.17.29 / 8.18.3 / 9.0.3 from ONE self-contained jar; Community 3-JOIN quota rejection and SHOW LICENSE verified; licences/NOTICE bundled.
• Config precedence fixed (flag > env > file > default) with a 12-AC matrix incl. live smoke; `method = "noauth"` builtin default removed (release-note item for 0.20.2).
• Parser-derived keyword registry with a two-way anti-drift test; highlighter/completer no longer drift.

Discovered along the way: #168 (es6 client closure excludes all log4j — plain es6 cli 0.20.1 crashes on first use; fix should ride 0.20.2).

Release train once the epic closes: elasticsql 0.20.2 → extensions → arrow 0.2.3 → softclient4es-repl (pins engine 0.20.2 + arrow-ext ≥ 0.2.3, releases LAST).