verify: real-git gr2 playground scenarios against org repos

[laynepenney]

Summary

Use real GitHub repos in synapt-dev to verify gr2 UX and workflow behavior against actual remotes instead of tempdir-only stand-ins.

Repos

• synapt-dev/gr2-playground-app
• synapt-dev/gr2-playground-api
• synapt-dev/gr2-playground-web

Why

The current prototypes prove metadata and command-shape ideas, but they do not yet pressure the UX against real git behavior:

• cloning actual remotes
• default branches
• multi-repo branch switching
• PR/review lanes against live repo state
• dirtiness and recovery around real checkouts

Verification scenarios

• bootstrap a workspace from a real WorkspaceSpec pointing at the playground repos
• create a feature lane spanning 2-3 repos
• create and switch a second feature lane while the first stays intact
• create a review lane without disturbing the feature lane
• verify repo status, lane show, and exec status stay honest about actual checkout state
• verify dirty-work detection and recovery behavior against real repos
• verify shared scratchpad flow alongside private lanes

Outcome

This should become the first real-user validation surface for gr2, and future prototype verification should be expected to hit it before we call a workflow solid.

[laynepenney]

Real-git harness landed on the prototype branch and completed a first end-to-end run against the three private playground repos.

Saved in:

• gr2/prototypes/real_git_playground.py
• draft PR: grip#554

First concrete findings from the run:

• apply currently materializes unit-local clones under agents/<unit>/...
• repos/<repo> paths remain non-git placeholders, which surface as block_path_conflict in repo status
• dirty local git state is detected correctly on unit-local repos
• lane metadata supports multiple concurrent feature/review lanes
• shared scratchpads can coexist beside private lanes
• real-git exec verification is still blocked on branch convergence with the shipped exec surface
• prototype lane metadata and shipped lane metadata are not yet unified enough for one combined exec harness

This is the kind of product data we wanted: not just that the idea sounds right, but exactly where the current UX and layout diverge from the intended model.