fix(pre_build): a fully-ignored dataset/ must not abort the release

[Jammy2211]

The bug

pre_build.sh runs under set -e. Line 75 was:

[ -d dataset ] && git add dataset/

git add dataset/ exits non-zero when every path it matches is ignored
(The following paths are ignored by one of your .gitignore files: dataset).
Under set -e that aborted the release on the first workspace processed
(autofit_workspace), before anything was committed, pushed or dispatched.

Why it went unnoticed

Latent since #150, which dropped the git add -f — correctly, since the -f
was force-committing simulated datasets meant to be generated at runtime (#126).
But it left the fully-ignored case unhandled.

The live path has not executed since:

• the 2026.7.9.1 release (2026-07-09) predates fix: honor .gitignore for dataset/ staging in pre_build (drop -f) [#126 leg 1] #150 (merged 2026-07-13);
• the nightly driver has stopped at Stage 4 validation every night since.

So this is the first time the step has run post-#150.

Blast radius — 7 of 11 release workspaces

Measured with git add -n dataset/ in each repo:

Workspace	.gitignore	old	new
autofit_workspace	dataset/	exit 1	exit 0
autofit_workspace_test	dataset/	exit 1	exit 0
autogalaxy_workspace_test	dataset/	exit 1	exit 0
autolens_workspace_test	dataset/	exit 1	exit 0
HowToFit / HowToGalaxy / HowToLens	dataset/	exit 1	exit 0
autolens_workspace	dataset/** + 13 ! allowlist	exit 0	exit 0
autogalaxy_workspace	dataset/** + 5 ! allowlist	exit 0	exit 0
euclid_…_pipeline, autolens_assistant	n/a	exit 0	exit 0

The fix

Tolerate the fully-ignored case; staging nothing is the correct outcome there.

Only the failure path changes. Where git add dataset/ already succeeded,
the command is byte-identical, so autolens_workspace's 13 allowlisted real
datasets and autogalaxy_workspace's 5 still stage exactly as before. #150's
leak fix is preserved — no -f is reintroduced.

Found by a human-authorized manual release --force for 2026.7.15.1, which is
blocked on this.

🤖 Generated with Claude Code