Cut 1.0-alpha

[jdatcmd]

What

Closes #367. VERSION goes to 1.0-alpha, [Unreleased] becomes [1.0-alpha]
dated 2026-08-04.

Every item on the checklist is closed:

item	how
open bugs	#358, #362, #363, #364, #376 all closed
docs accuracy pass	#371
CHANGELOG	#372
single-engine benchmarks	#374
cross-engine benchmarks	#377
five-major matrix	green, 112 suites on each of PG15-19

Two decisions worth a reviewer's eye

The extension's default_version stays 1.0-dev, and the changelog now says so.
That field governs ALTER EXTENSION UPDATE. Moving it means renaming
pgcolumnar--1.0-dev.sql and adding an upgrade script that does not exist, and
breaking installs to make a version string tidy is the wrong trade for a pre-release.
SELECT extversion FROM pg_extension therefore reports 1.0-dev on a 1.0-alpha
build. Stated in the changelog preamble rather than left to be discovered.

A known-limitations section, in the changelog rather than only the tracker. An
alpha that ships its sharp edges in issues alone is an alpha that hides them:

• the grouped parallel arm declined on expression grouping keys (The grouped parallel arm loses to the serial node because the Finalize is priced off estimate_num_groups on an expression key #369), both
  settings off by default;
• the fetch cache's real bound, 4 x (cap + retained indexes + groupBuffer), and
  the measured 62 MB entry (#361's fetch cache is not bounded by 4x the cap: one entry measures 62 MB against 32 MB on a wide varlena table, and gets 1.14x #364), including why holding the bound costs 47% in
  time;
• the index-fetch penalty bounded rather than modelled against the consumer (The index-fetch penalty prices a path by its own row estimate, not the rows its consumer pulls: 57x regression on DISTINCT ON (last point per host) #376);
• point-lookup cost growing with table width, because a fetch decodes the attribute
  prefix.

Verification

docs_style, smoke and harness_selftest pass on the cut tree. No source change,
so the five-major matrix result for a2027ea stands.

🤖 Generated with Claude Code

[ChronicallyJD]

Reviewed. The cut is right and the two decisions you flagged are both the right call —
especially keeping default_version at 1.0-dev and saying so rather than renaming
the SQL file and inventing an upgrade script to make a string tidy. I verified the
known-limitations entries against my own measurements: 62 MB against a 32 MB cap is
what I measured on #364, and the 47% is 127,699 ms against 187,521. Both correct.

Putting the sharp edges in the changelog rather than only the tracker is the right
instinct, and I would have argued for it if you had not done it.

One blocker: README contradicts the file it cites

README.md:27:

It is pre-release; the version marker is 1.0-dev, recorded in VERSION.

This PR makes VERSION say 1.0-alpha. So the README asserts the contents of a file
that is in the same commit and disagrees with it. That is the first thing a reader
sees, and it points straight at VERSION to check.

README.md:12 has the same problem in the badge:

<a href="VERSION"><img src="badges/version.svg" alt="Version 1.0-dev"></a>

and badges/version.svg itself renders version: 1.0-dev, so the badge needs
regenerating along with the alt text.

Suggested wording for line 27, which also carries your default_version point to
where people will actually hit it:

It is pre-release; the version marker is
1.0-alpha, recorded in VERSION. The extension's own default_version is still
1.0-dev, so SELECT extversion FROM pg_extension reports that.

Not blocking

The 1.0-dev references in PROVENANCE.md are all about the v1.0-dev tag and
the practice line it preserves. Those are historical fact and should stay exactly as
they are — I checked each one rather than pattern-matching the string.

pgcolumnar.control and pgcolumnar--1.0-dev.sql are the deliberate pair you
described, and they agree with each other.

Everything else checks out: VERSION, the [1.0-alpha] - 2026-08-04 heading, and the
checklist table all match what actually landed. Happy to see this merge once the README
matches VERSION; I can push that commit if you would rather not do another round.