[coretl]

Throwaway. Deleted once the question is answered.

Observations 1 and 2 both had the branch and the tag pointing at the same commit, so neither separates "GITHUB_SHA is main's HEAD" from "…is the commit that was built". This one does, and it is the highest-volume real case: if PR-triggered deploys all collide on an unchanging main HEAD, previews break even though tags work.

Two pushes, each carrying a different visible marker in docs/index.md:

1. ALPHA — does /pr-<n>/ appear at all?
2. BRAVO — does it update, at the same unchanged main HEAD?

Also checking github.sha in the probe's context job against workflow_run.head_sha — here they genuinely differ, so this is the empirical test of the docs-table claim.

🤖 Generated with Claude Code

[coretl]

Experiment complete — all three observations passed. Four consecutive deploys at the identical pages_build_version: b242374… (main push, tag, PR push 1, PR push 2) and every one of them updated the origin. workflow_run re-serves at a repeat SHA, and it stamps main's HEAD (this PR's head was 6d48fce…). Migration proceeding in a follow-up.