CreateRepo should wait for git-layer readiness; delete pool recycling machinery

[ggallen]

Blocked by #6820. Closes #6701.

What happens

CreateRepo returns as soon as the GitHub API accepts the creation request. Callers that immediately perform git operations (create tree, commit, branch) hit eventual-consistency errors because the auto_init commit hasn't propagated:

• 422 Tree SHA does not exist
• 422 base_tree is not a valid tree oid
• 409 Git Repository is empty
• 404 Not Found

The current workaround — awaitCreation in ensure.go — polls GetRepo, which only proves API visibility, not git-layer readiness. The poll succeeds but the repo isn't actually usable yet.

4 behaviour test failures in a 2-day window (Aug 31) hit these errors, all during test setup on freshly created repos. See also #5671 (closed), #5774 (closed), #6701 (open, superseded by this issue).

What should happen

CreateRepo should not return until the repo is ready for git operations. Polling GetRef("heads/main") (or equivalent) after the create call confirms the auto_init commit and default branch ref have propagated. Once that succeeds, the repo is usable — git objects exist and permissions are settled.

Every existing CreateRepo caller immediately performs git operations on the result, so this benefits all call sites with no behavior change.

Cleanup

Once #6820 lands (ephemeral single-org repos with --per-repo-wif-repos "*" on the preview mint), the pool and recycling machinery becomes dead code. This is a significant simplification.

Pool recycling (already identified)

• Delete repopool_external_mint.go — unused factory, bot-generated, never called
• Delete resetRepo, awaitCreation, awaitDeletion, ensureRepoExists and related constants (resetMaxAttempts, resetRetryDelay)

Channel-based pool in composedDriver

With --per-repo-wif-repos "*" on the preview mint, repo names no longer need to be coordinated between the factory and the driver. Each repo is unique, created once, and deleted after its test. The pool pattern is unnecessary:

• Remove names chan string, capacity field, Capacity() method
• AllocateRepo generates a fresh bt-{uuid} name on the spot instead of pulling from a channel
• DeallocateRepo no longer pushes names back to the channel — the repo is deleted, the name is never reused
• Remove DefaultPoolSize, envPoolSize(), BEHAVIOUR_POOL_SIZE env var support
• Remove buildEphemeralRepoList and the prefix parameter threading from factory → composed driver
• If concurrency limiting is still needed, use a sync.Semaphore instead of a channel

Ensurer simplification

Each repo is unique and never reused, so caching and deduplication are unnecessary:

• Remove ensured map and singleflight.Group from repoEnsurer
• Remove InvalidateCache method and the ensurer interface's InvalidateCache contract
• Collapse the ensurer interface into a simple function: CreateRepo → installFullsend → awaitWorkflowReady

Finalize cleanup

• Simplify Finalize's lease reclamation — outstanding repos should be deleted, not pushed back to a pool
• Mint teardown remains unchanged

Summary

The ensure flow simplifies from the current multi-step reset/create/validate/cache cycle to: CreateRepo (with git-readiness wait) → installFullsend → awaitWorkflowReady. The composed driver becomes a thin wrapper around name generation, outstanding-repo tracking, and mint lifecycle.

[fullsend-ai-triage]

🤖 Finished Triage · ✅ Success · Started 9:12 PM UTC · Completed 9:18 PM UTC

Commit: e362047 · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $0.97

[fullsend-ai-triage]

This issue is blocked by an open pull request that must land first.

Prerequisite: PR #6820 migrates behaviour tests to single-org ephemeral repos. The cleanup section of this issue (deleting repopool_external_mint.go, resetRepo, awaitCreation, awaitDeletion, ensureRepoExists, and related constants) explicitly depends on that PR — the pool recycling machinery only becomes dead code once ephemeral repos replace the 12-org pool model.

The CreateRepo readiness fix (polling GetRef("heads/main") instead of GetRepo) can also be validated more cleanly once the new ephemeral flow is in place.

PR #6820 is currently labeled ready-for-review — once it merges, this issue should be unblocked and ready for implementation.

Proposed test case:

Test: CreateRepo blocks until git-layer is ready
- Call CreateRepo with auto_init=true
- Immediately call GetRef("heads/main") on the result
- Assert the ref exists and points to a valid commit SHA
- Then perform a git operation (create tree/commit) — assert no 422/409/404
- Edge case: verify timeout behavior when ref never appears

Blocked by:

• ci(#6815): migrate behaviour tests to single-org ephemeral repos #6820

---

Labels: Bug in e2e test infrastructure causing eventual-consistency failures with high impact (4 failures in 2 days), explicitly blocked by PR #6820.

[fullsend-ai-triage]

🤖 Finished Triage · ✅ Success · Started 10:20 PM UTC · Completed 10:24 PM UTC

Commit: e362047 · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $0.68

[fullsend-ai-triage]

PR #6820 is still open and remains the prerequisite for this work. It is currently labeled ready-for-merge and needs-human, suggesting it is close to landing but requires human intervention.

Once PR #6820 merges, this issue becomes unblocked:

1. The CreateRepo readiness fix (polling GetRef("heads/main") instead of GetRepo) can be validated against the new ephemeral repo flow
2. The pool recycling machinery (repopool_external_mint.go, resetRepo, awaitCreation, awaitDeletion, ensureRepoExists, channel-based pool in composedDriver, ensurer caching) becomes dead code and can be deleted

No other changes since the prior triage — the blocking relationship and all identified problems remain as documented.

Blocked by:

• ci(#6815): migrate behaviour tests to single-org ephemeral repos #6820