[ev-br]

Implement UC-04 style fixed branch, cross-environment comparisons.

The PR is on top of gh-71.
Since the "stack" of PRs gets deep and unwieldy, here's the same PR sent against gh-71 branch on my fork: ev-br#6.
Also to simplify reviewing, the first commit adds the plan file, and the last commit removes.

The CLI spellings end up a bit different from the UC-04 itself, am happy to tweak further or discuss.
There's a demo in prototype/examples/numpy-venvs:

$ cd benchx/prototype/examples/numpy-venvs
$ ./run.sh 
== 1. the user prepares both environments; benchx builds nothing
  .venv-a: numpy 1.26.4
  .venv-b: numpy 2.0.2
== 2. the caller's loop: alternate the sides, round by round, one slot per call
precision: 3 processes x 3 values (config: /home/ev-br/proj/quansight/benchx/prototype/examples/numpy-venvs/benchx.conf.json)
  round 0, slot 0: side a
precision: 3 processes x 3 values (config: /home/ev-br/proj/quansight/benchx/prototype/examples/numpy-venvs/benchx.conf.json)
  round 0, slot 1: side b
precision: 3 processes x 3 values (config: /home/ev-br/proj/quansight/benchx/prototype/examples/numpy-venvs/benchx.conf.json)
  round 1, slot 2: side a
precision: 3 processes x 3 values (config: /home/ev-br/proj/quansight/benchx/prototype/examples/numpy-venvs/benchx.conf.json)
  round 1, slot 3: side b
precision: 3 processes x 3 values (config: /home/ev-br/proj/quansight/benchx/prototype/examples/numpy-venvs/benchx.conf.json)
  round 2, slot 4: side a
precision: 3 processes x 3 values (config: /home/ev-br/proj/quansight/benchx/prototype/examples/numpy-venvs/benchx.conf.json)
  round 2, slot 5: side b
== 3. compare: the sides differ in the numpy the collector read
run numpy-1793123241  profile environments  baseline numpy-1.26.4  contender numpy-2.0.2  [local-only]
  numpy-1.26.4: how built: {"summary":{"numpy":"1.26.4"}}
  numpy-2.0.2: how built: {"summary":{"numpy":"2.0.2"}}
  differs: coordinates.environment.system.deps.numpy, coordinates.environment.system.deps.sha256

IMPROVED (1)
  bench_sort                       wall-time  effect  -81.20%  noise   0.27%  rounds=3

NO CHANGE DETECTED (2)
  bench_matmul                     wall-time  effect  +10.34%  noise   5.34%  rounds=3
  bench_sum                        wall-time  effect   -0.50%  noise   0.23%  rounds=3
== 4. the check: declare numpy 2.0.2 for a run in the venv that has 1.26.4
precision: 3 processes x 3 values (config: /home/ev-br/proj/quansight/benchx/prototype/examples/numpy-venvs/benchx.conf.json)
  the results carry collector-mismatch; the run went ahead, with the numpy it found

[MaanasArora]

Thanks, this looks right on a high-level read, no objections to merging and iterating!

[ev-br]

Thanks @MaanasArora ! I'll merge, rebase the next PR in the "stack" (#89), and expand the discussion of the context collectors / identity design in gh-97. This will need some redesign, as you noted in #71 (comment)