feat: compare version model evaluations in the CLI

[leeclemnet]

What does this PR do?

Add roboflow eval compare --project … --version … to inspect the shared accuracy/latency comparison, including frontier membership and exclusions, as a table or full JSON. The command uses the public read-only endpoint and starts no evaluations.

Type of Change

• New feature (non-breaking change that adds functionality)

Testing

• local cli test

help

roboflow eval compare --help

 Usage: roboflow eval compare [OPTIONS]

 Compare test-set accuracy and median latency without starting evaluations.

╭─ Options ────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ *  --project          -p      TEXT                  Project slug [required]                                          │
│ *  --version          -v      INTEGER RANGE [x>=1]  Dataset version number [required]                                │
│    --frontier-metric          TEXT                  Metric for frontier membership (uses the project default if      │
│                                                     omitted)                                                         │
│    --help             -h                            Show this message and exit.                                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

object detection version with default frontier metric (mAP@50:95)

roboflow eval compare -p chess-pieces-mjzgj-v3mep -v 26

MODEL                                                    mAP5095  MEDIAN LATENCY (ms)  FRONTIER  EXCLUSION
-------------------------------------------------------  -------  -------------------  --------  -------------------
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-nano-t6   53.3%    17.04
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-nano-t5   70.1%    47.19
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-nano-t4   76.2%    41.96
lee-sandbox/chess-pieces-mjzgj-v3mep-26-yolo26s-t3       81.8%    14.06                Yes
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-small-t1  78.6%                                   latency_unavailable
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-small-t2  78.5%    47.69

object detection version with non-default frontier metric (F1)

roboflow eval compare -p chess-pieces-mjzgj-v3mep -v 26 --frontier-metric f1

MODEL                                                    f1     MEDIAN LATENCY (ms)  FRONTIER  EXCLUSION
-------------------------------------------------------  -----  -------------------  --------  -------------------
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-nano-t6   83.4%  17.04
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-nano-t5   87.0%  47.19
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-nano-t4   93.0%  41.96
lee-sandbox/chess-pieces-mjzgj-v3mep-26-yolo26s-t3       95.9%  14.06                Yes
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-small-t1  96.8%                                 latency_unavailable
lee-sandbox/chess-pieces-mjzgj-v3mep-26-rfdetr-small-t2  97.4%  47.69                Yes

classification version with default frontier metric (F1)

roboflow eval compare -p car-colors-1smyc-aqetf -v 4

MODEL                                                               f1     MEDIAN LATENCY (ms)  FRONTIER  EXCLUSION
------------------------------------------------------------------  -----  -------------------  --------  ---------
lee-sandbox/car-colors-1smyc-aqetf-4-resnet50-t2                    88.3%  4.89                 Yes
lee-sandbox/car-colors-1smyc-aqetf-4-vit-base-patch16-224-in21k-t1  88.2%  4.60                 Yes
lee-sandbox/car-colors-1smyc-aqetf-4-resnet101-t3                   89.5%  6.68                 Yes

evaluation unavailable

roboflow eval compare -p chess-pieces-mjzgj-v3mep -v 15

MODEL                        mAP5095  MEDIAN LATENCY (ms)  FRONTIER  EXCLUSION
---------------------------  -------  -------------------  --------  ----------------------
chess-pieces-mjzgj-v3mep/15                                          evaluation_unavailable

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context

Deploy the platform comparison endpoint before releasing the Python package. Workspace availability remains controlled by the platform feature flag.