[Feature]: Complete adoption recovery and idempotency coverage

[Baziar]

Problem statement

Adoption writes project metadata, workspace binding, registry/contract state,
Model/Graph evidence, and agent outputs. Transactional rollback now protects
many writes, but the complete multi-surface transition matrix is not yet one
explicit release gate.

Proposed solution

Current foundation

• linked adoption without moving/copying source;
• managed output ownership and lifecycle transaction rollback;
• re-adoption detection refresh and generated-agent collision protection;
• project-local entry, workspace binding, canonical aggregation, and freshness sealing;
• isolated real-repository qualification;
• parent-plus-nested adoption coverage where an unrelated stale Goal cannot
  roll back a valid project binding.

Remaining work

Test adoption as a state machine:

unmanaged -> linked -> synchronized -> intelligence-current
          -> no-op re-adopt | changed re-adopt | unavailable -> relinked

Acceptance criteria

• Default-workspace and explicitly selected workspace adoption are covered.
• Repeating an unchanged adoption is an observable no-op.
• Runtime/manifest changes invalidate only required downstream evidence.
• Same-basename projects retain distinct identity.
• Interruptions before/after project-local commit preserve a recoverable state.
• Missing or moved linked source reports an exact relink/recovery action.
• Registry, contract, Model, Graph, Doctor, and agent index agree after every transition.
• A stale or invalid Goal blocks only the project scope it applies to.
• Native Linux, macOS, and Windows transition fixtures pass.

Completion evidence

Publish a transition table with expected writes, rollback points, JSON verdict,
and recovery command for each scenario. No transition may leave split-brain
membership while reporting success.

Alternatives considered

The following alternatives or adjacent concerns are intentionally outside this issue:

• unadopt, detach, or project removal commands; they are not part of the
  current product and require a separate product/safety decision;
• deleting user source;
• unrelated workspace archive/delete behavior.

Expected impact

Adoption becomes safely repeatable and recoverable without split-brain registry,
workspace, intelligence, or agent state.

[github-actions]

Hey @Baziar — welcome to Workspai! 👋

Thanks for taking the time to open your first issue. Clear reports and focused ideas help us make Workspai more useful for developers and AI tools working across real software workspaces.

A maintainer will review the scope and labels. If you would like to work on this issue yourself, leave a comment and we’ll help you find the relevant contracts, code, and tests.

Glad to have you here.