[isayev]

The fast legs moved to uv last change; these are the jobs that were left, and
the ones where the install actually cost something. Six install steps convert:
the four slow shards, no-aimnet, and both docs.yml jobs.

Result

RUN WALL 165s            (baseline 243s / 253s / 304s over the last three runs)
  +165s  slow tests (NNP integration)                   (7s)
  +155s  slow shard 1 of 4 (end-to-end pipeline)        (152s)  <- binds the run
  +145s  slow shard 3 of 4 (workers, compile, thermo)   (142s)
  +145s  slow shard 2 of 4 (main and smiles2mols)       (142s)
  +145s  slow shard 4 of 4 (everything else)            (129s)
  + 83s  pytest py3.12                                   (80s)
  + 71s  pytest py3.11                                   (68s)
  + 41s  Build Documentation                             (39s)
  + 37s  no-aimnet                                       (35s)
  + 11s  ruff                                             (9s)

Per install step:

job	pip	uv
slow shard 1	72 s	7 s
slow shard 2	78 s	8 s
slow shard 3	63 s	10 s
slow shard 4	66 s	6 s
no-aimnet	86 s	7 s
Build Documentation	50 s	7 s
Check Links	85 s	8 s

Built head to head in one runner, with nothing else differing: 81-90 s with
pip, 9-10 s with uv.

No job's install is its largest step any more. The gate's binding constraint is
now shard 1's own test step (120 s of a 152 s job), so the next lever is the
shard balance -- 120 / 109 / 105 / 74 test-seconds -- which this change
deliberately leaves alone.

The two flags

--index-strategy unsafe-best-match: uv's default stops at the first index
that has a package at all, and the PyTorch index carries a requests (2.28.1)
far below this project's floor, so the default resolves nothing and the install
fails outright. --system: a runner has no virtual environment and uv declines
to touch an interpreter it was not pointed at.

The resolution is unchanged, checked per job

A temporary workflow (run 41546099135, removed from this branch's head) built
each converted environment twice in one runner -- the pip command the job used
to run and the uv command it runs now -- and froze both with the same pip
binary:

extras	job(s)	pip	uv	identical	torch	nvidia_*/triton
ase,ani,dev	slow shard 1-4	118	118	yes	2.13.0+cpu	none
ani,ase,dev minus aimnet	no-aimnet	117	117	yes	2.13.0+cpu	none
ase,ani,docs	build, linkcheck	106	106	yes	2.13.0+cpu	none

For no-aimnet the absence is the point, so it was asserted rather than
inferred: on both environments aimnet is missing from the freeze and python -c "import aimnet" fails. The 118 to 117 drop is aimnet and nothing else on
both sides, so uv pip uninstall left its transitive dependencies in place
exactly as pip uninstall -y did.

cache: pip goes with them

uv does not read pip's HTTP cache, so the restore was 6-10 s of setup feeding a
tool that never looks at it. Nothing writes those entries now either; the five
stored ones (~3.3 GB of the repository's 10 GB budget) age out on their own.
The only cache a run still depends on is the 31 MB model-weight entry the slow
shards restore, which cache-keepalive.yml holds open -- that file needs no
change and gets none.

uv is pinned

UV_VERSION: "0.13.6" as one workflow-level value per file, referenced as pip install "uv==${UV_VERSION}" at each of the five sites. Unpinned, uv was a
resolver taking whatever version was current at run time in front of every gate
here -- the same fails-on-luck mode that pins ruff exactly and SHA-pins every
action. The comment that declined a marketplace action partly to avoid another
pinned dependency now says why that is an argument against the action, not for
leaving the installer floating.

0.13.6 because that is the version every number above was measured with,
confirmed from the logs of all four runs cited. It cannot go in the dev
extra: uv is what installs the project, so it exists before any extra does.

The pin costs nothing. On the pinned head (run 41548392879) the gate is 181 s
and every install is 6-9 s, unchanged; the 165 to 181 s difference is entirely
shard 1's test step (120 s then 137 s), which is the tier's ordinary noise. Two
runs put the gate at 165-181 s against a 243-304 s baseline.

Not converted

lint installs ruff and nothing else in 2 s, which is less than pip install uv would add. publish.yml is out of scope, and separately its pip install of
the built wheel is deliberately the way a user would install it.

Two comments that had gone stale are corrected while the files are open: the
fast legs' note on cache: pip claimed the other jobs still wrote that cache,
and docs.yml described tests.yml's cron as weekly for the model cache,
which went monthly when cache-keepalive.yml took that job over.

Check Links is skipped on a pull request, so its install was exercised by
dispatching the workflow on this branch (run 41547117953, green). Its pip
baseline is the 2026-11-02 monthly cron, run 41402781229.