Version Compat CI spent six runner slots on 287s of combined work. On a sampled main
commit each of the six waited 23 to 31 minutes to be admitted, in order to execute for
between 15 and 128 seconds. The queue is the cost, so the slots are the target.

Why these six can share a runner
------------------------------------------------------------------------
Not a general claim about the suite, a specific one about these tests: they fetch raw
source from raw.githubusercontent.com and grep it for symbols (tests/version_compat/
_fetch.py, has_def / first_match). Nothing is pip-installed, which is the whole point of
the canary and is also what makes co-location safe -- there is no venv for "transformers
4.57.6" and "transformers main" to fight over, because neither is installed.

The repo had already proved the shape: `daily-fresh-fetch` runs all of these in one
process on cron. This makes the PR path do the same thing, in parallel.

Measured on the full set, locally:

    serial   1606 passed, 182 skipped   182.8s
    -n 4     1606 passed, 182 skipped    27.8s
    -n 8     1606 passed, 182 skipped    75.2s

The pass counts are identical in all three, which is the number that matters: the bundle
tests exactly what the six jobs tested.

-n 4 rather than higher. -n 8 is 2.7x SLOWER, which for a fetch-bound suite is upstream
throttling and not core contention -- the job's own comment already notes the 5000/h
authenticated raw-fetch ceiling. 4 also matches the hosted runner's core count and the
-n 4 every other pytest job here uses. --dist loadfile keeps a suite on one worker.

What was deliberately NOT folded in
------------------------------------------------------------------------
zoo-imports-under-spoof and grpo-fake-run stay separate. Both install a torch stack, and
they install mutually exclusive TRL pins ('trl>=0.22,<0.26' against `pip install --upgrade
trl`), so one venv cannot hold both. A guard asserts they still exist, because folding
them into a job with no torch at all is the plausible next step and it would not fail
loudly.

Two pre-existing gaps this found
------------------------------------------------------------------------
The new guard derives coverage from the filesystem rather than from a list, and on its
first run it reported two suites that no pull_request job runs at all:

  test_import_leaves_torch_globals_alone.py   `import torch` inside a subprocess probe
  test_trl_vllm_generation_lora_patch.py      needs an installed TRL

Both pre-date this change and both are swept daily by the cron job, so they are checked
daily rather than never. Neither can join a bundle that installs only pytest. The second
belongs with grpo-fake-run, which already has that stack, but it is not in that job's list
either; it was left alone rather than wired in blind, because I could not verify it passes
in that environment first and an unverified suite would have put an unrelated red on the
change that noticed the gap. Recorded in CRON_ONLY with the reasons.

Check names
------------------------------------------------------------------------
Six checks become one. Removed:

  vLLM pinned-symbol matrix (>= 0.9.0 + main)
  TRL / GRPO pinned-symbol matrix
  PEFT pinned-symbol matrix (pyproject window + main)
  sentence-transformers pinned-symbol matrix
  bitsandbytes pinned-symbol matrix
  transformers pinned-symbol matrix (4.57.6 + 5.x + main)

Added:

  pinned-symbol matrix (vLLM, TRL, PEFT, ST, bitsandbytes, transformers)

Branch protection naming any of the six needs updating.

Verified
------------------------------------------------------------------------
Mutation-tested, each failing exactly one guard: drop a suite from the bundle; remove
-n 4; remove --dist loadfile; sweep the directory instead of naming files.
tests/studio 4584 passed. lint_workflow_triggers OK across 41 files. No dangling
references to the six removed job ids anywhere in the tree.