Add Replay ("Policy") performance tests (TaskCompletionRateTest)

[abrichr]

Feature request

We need to extend #314 to include some useful tests and generate an automated report.

This involves:

• Create recordings of three tasks:

1. Open a calculator and perform a short calculation
2. Open a spreadsheet (e.g. https://github.com/OpenAdaptAI/OpenAdapt/blob/cb70f35985eeb579fd3e13b20a9839b10729921d/tests/assets/excel.png), open a time tracking app (e.g. https://clockify.me), copy a week's worth of data from the spreadsheet into the app, and save/submit the data in the app. (e.g. https://www.youtube.com/watch?v=omP11q-o_0I)
   Alternatively if browser events are not yet available (see Add Chrome browser event in database during recording #744), replicate something similar with two different spreadsheets open simultaneously (one for reading, one for writing).
3. Open powerpoint and create a short presentation.

• Save them as fixtures
• Add automated tests to run a replay (with configurable strategy, defaulting to VanillaReplayStrategy) and evaluate the outcome. Outcome evaluation can be implemented with WindowEvent data.
• Add a script to log the outcome results to stdout and/or to a file.

Motivation

Scientific rigor and reproducibility.

[abrichr]

@seanmcguire12 your assistance would be greatly appreciated!

[abrichr]

@KrishPatel13 outcome evaluation for web apps will depend on finishing #364

[abrichr]

Save a fixture with recording.task_description = "test: calculate 2x3" that is just like the video currently on the website.

Test 1: Run the VanillaReplayStrategy with empty instructions (or give it instructions like replay the recording verbatim). Use openadapt.window to assert that the calculator display area contains the expected value 6.

Test 2: Run the VanillaReplayStrategy with instructions like calculate 9-8+7. Use the same API to assert that the calculator display area contains the expected value 8.

Parameterize the replay strategy and iterate over all of them. Produce a report with the results.

[abrichr]

@seanmcguire12 please submit a PR with your work-in-progress 🙏

[abrichr]

Closing as part of the architecture transition to modular packages. The legacy codebase has been frozen at v0.46.0.

Replay performance tests (TaskCompletionRateTest) are being implemented in the new openadapt-evals package.

For the new modular architecture, please open issues in the relevant package repository:

• GUI capture: https://github.com/OpenAdaptAI/openadapt-capture
• ML/training: https://github.com/OpenAdaptAI/openadapt-ml
• Benchmarks: https://github.com/OpenAdaptAI/openadapt-evals
• Privacy: https://github.com/OpenAdaptAI/openadapt-privacy

See PR #960 for details on the new architecture.