GHA cache key changes between restoring and saving

[penelopeysm]

In both the preview and publish workflows, the _freeze folder is cached with a key that depends on the Manifest.toml file (i.e. the project environment):

restore

docs/.github/workflows/publish.yml

Lines 32 to 38 in ef4165e

	- name: Restore cached _freeze folder
	id: cache-restore
	uses: actions/cache/restore@v4
	with:
	path: |
	_freeze/
	key: ${{ runner.os }}-${{ hashFiles('**/Manifest.toml') }}

save

docs/.github/workflows/publish.yml

Lines 73 to 79 in ef4165e

	- name: Save _freeze folder
	id: cache-save
	uses: actions/cache/save@v4
	with:
	path: |
	_freeze/
	key: ${{ runner.os }}-${{ hashFiles('**/Manifest.toml') }}

However, as evidenced by recent CI runs e.g. https://github.com/TuringLang/docs/actions/runs/10670233519/job/29573828445, the restore key and save key are different. Presumably this means that somehow the Manifest.toml file is being changed during the process of quarto render. I'm not sure how or why this happens, considering that the Quarto notebooks don't modify the environment (they only call Pkg.instantiate). In any case it would be worth fixing this so that we can avoid running needless 1 hr 45 min doc generations when fixing typos.

I wonder if it's because Quarto installs some stuff like bits of its own Julia engine in the environment.

[penelopeysm]

Trying to isolate the issue on another branch (pysm/cat-manifest); check this log later. https://github.com/TuringLang/docs/actions/runs/10711186250/job/29699379263

[penelopeysm]

Manifest specifies Julia 1.10.4, but GHA runs with Julia 1.10.5

[penelopeysm]

Solution is to call Pkg.resolve() (not just instantiate()) before hashing the file.

Will work on a fix soon. I'll also add a GHA which updates the Manifest as necessary