[no-ci] Register pytest.mark authorship markers, add matching AGENTS.md instructions

[rwgk]

Summary

This adds repository-wide AGENTS.md guidance for marking the provenance of newly added unit tests using pytest markers, and registers those markers in the pytest config roots.

The marker system is intentionally minimal. The design goal is to keep agents and humans choosing from no more than three clear states:

• @pytest.mark.agent_authored(model="<model>") for tests authored by an agent and not yet materially reviewed or rewritten by a human.
• @pytest.mark.human_reviewed for agent-authored tests that a human has materially reviewed or rewritten.
• @pytest.mark.human_authored for tests authored by a human, or rewritten enough that the authorship is primarily human.

The guidance keeps markers local to the tests they describe: markers should be placed immediately above each test function, with class-level markers only when every test method in the class has the same provenance. It also says not to use module-level pytestmark for authorship provenance, because that is easy to miss in large files and makes later per-test provenance changes ambiguous.

The pytest marker registrations include sync comments so future edits keep the copies aligned across pytest config roots.

Relationship to #2232

This is an alternative to #2232. PR #2232 uses lightweight comment tags; this version uses formal pytest markers so tests can be selected by provenance with pytest -m, for example pytest -m agent_authored.

Purpose

The purpose of these markers is to preserve ground-truth test provenance at the moment when it is easiest and most reliable to capture.

Provenance can be useful when interpreting a test. If a test was carefully authored or materially reviewed by a human, we may reasonably give it more evidentiary weight when investigating ambiguous behavior, evaluating an API change, or deciding whether an assertion reflects deliberate intent. This is not a guarantee that the test is correct, nor does an agent-authored marker imply that a test is inferior; it is additional context that would otherwise be lost.

Recording the generating model also makes the annotations actionable over time. As newer models become available, we could run automated passes over tests written by older models to review assumptions, strengthen assertions, identify missing cases, or remove low-value duplication. The provenance markers provide a practical way to select those cohorts and record when meaningful human review has occurred.

The accumulated metadata may also support useful empirical analysis. In the future, we could examine whether test provenance correlates with factors such as defect frequency, regressions detected, coverage, flakiness, or maintenance cost. We do not yet know which correlations will prove meaningful, but such analysis will only be practical if we capture the ground truth while it is readily available. Reconstructing provenance later from commit history or pull-request discussion would be incomplete and unreliable.

For now, the markers are informational. They do not automatically change CI behavior or assign a quality score. Their purpose is to preserve lightweight, machine-readable evidence that future reviewers, tools, and studies can use.

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.