Update ZigHouse

[donge]

Update ZigHouse

Refresh the c6i.4xlarge result with v0.2.0-clickbench, and address the
feedback in #895 about hardcoded query handling.

Two execution paths, one storage

• ClickBench optimization profile — fast paths hand-tuned to the shapes
  of the 43 ClickBench queries. Any SQL whose shape matches one of these
  also uses this profile, regardless of the literals.
• Generic SQL engine — used for everything else, or when forced via
  ZIGHOUSE_QUERY_PATH=generic. compare mode runs both paths and checks
  byte-identical output.

zighouse/generic-smoke.sh runs a few non-ClickBench statements through the
generic path so the capability frontier is easy to inspect.

Currently supported on the generic path: COUNT(*), SUM / MIN / MAX
/ AVG, COUNT(DISTINCT col), GROUP BY on a single low-cardinality
column, WHERE with numeric and date comparisons combined by AND.

Not yet supported (on the roadmap): GROUP BY on high-cardinality string
columns with ORDER BY ... LIMIT, JOIN, subqueries, arbitrary table
import, and broader operator coverage.

Performance vs v0.1.0-clickbench (AWS c6i.4xlarge, 100M rows)

Metric	v0.1.0	v0.2.0	Δ
load_time	107.33s	116.24s	+8.30 %
best-of-3 query sum (43 q)	19.53s	17.82s	−8.75 %
data_size	10.91GB	10.91GB	≈ same

Largest improvements: q39 −29 %, q37 −28 %, q15 −21 %, q33 (7.2s → 6.2s) −14 %.
Largest regressions: q40 +27 % (+82 ms), q28 +17 %, q26 +8 %.
Load-time regression is mostly the native_title_derived import phase
(+35 %); under investigation.

Install

Binary: https://github.com/donge/zighouse/releases/tag/v0.2.0-clickbench
SHA256: 5c30c8e3a56639a1d769ff1d3246c91109c47e9c97e776aec046519dcc66f3a6

[CLAassistant]

All committers have signed the CLA.

[alexey-milovidov]

Could you please sign a CLA?

[donge]

Sure, done.