ci: judge a new branch by the whole branch, not its last commit

[jcschaff]

Fixes #2034.

The bug

The should-run gate skips the fast lane when a push changed only documentation. It picked its comparison base like this:

github.event.before, if it resolves   ->  the previous tip of this ref
otherwise HEAD^1                      ->  the previous COMMIT

On a push that creates a branch, github.event.before is all-zeros and does not resolve, so it fell through to HEAD^1 and judged the entire branch by its final commit. A branch ending in a docs commit skipped build, every Fast shard and Quarkus — and reported CI-Test-group-Fast as pass in about two seconds.

PR #2033 shipped exactly that way: four commits, three of them Java across vcell-server, a docs commit last.

It is the bad kind of failure — silent, indistinguishable from a pass, and since regression.yml does not run on pull_request either, the merge queue was the first place that code would ever be compiled. It also gets more likely with better commit hygiene, because ending a branch with a docs commit is a normal shape.

The fix

Base selection now distinguishes three cases rather than two:

situation	base	why
push to an existing ref	github.event.before	most precise answer there is. All-zeros now fails this test explicitly — it means the ref did not exist, so it is not a base
merge commit (pull_request, merge_group)	HEAD^1	the first parent is exactly the PR's contents, as before
new branch	git merge-base with the default branch	spans the branch however many commits it has, and reduces to HEAD^1 for a single-commit branch

Verified against a scratch repository

scenario	old	fixed
new branch, code then docs last	SKIP ❌	RUN ✅
new branch, genuinely docs-only	SKIP	SKIP ✅ optimisation preserved
push to existing branch, docs-only push	SKIP	SKIP ✅ before still wins
merge commit	RUN	RUN ✅ first parent still used

This PR is its own regression test

The last commit is deliberately docs-only. Under the bug this branch would have skipped everything and reported green; with the fix the lane runs, because the branch as a whole touches ci.yml.

So: if the fast lane runs on this PR, the fix works.

Also

A skip was indistinguishable from a pass, which is what let this go unnoticed. The gate now writes its decision and its comparison base to the job summary:

### CI lane: SKIPPED - nothing was compiled or tested
Reason: only documentation changed (docs/MESSAGING.md)
Base for the comparison: `a1b2c3d`

docs/BUILDING.md records that a green fast lane has two meanings and where to look to tell them apart.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D71LBYmQNf5J94wPqr81Jx