feat(admin): assign existing upstreams to pools

[erwill2]

Summary

Adds a complete admin workflow for attaching an existing upstream identity to a Pool, and fixes account reactivation so a newly attached pending Pool assignment can become operational.

These changes are combined because they share the same assignment lifecycle invariant: an existing Pool/identity slot must be safely promoted to an active, eligible assignment rather than duplicated or left pending.

Operator workflow

• Unfiltered Upstreams view now includes non-deleted identities available to Pool management, including identities with no active Pool assignment.
• An explicit Pool filter continues to show only accounts assigned to that Pool; this is based on the filter value rather than Pool-count equality, so it is correct even when the instance has only one Pool.
• Adds an Assign to Pool card action and an accessible dialog using the existing admin dialog/footer/action primitives.
• The identity is held in server-side LiveView state and the submitted Pool ID must resolve through the operator's currently visible Pool list.
• The mutation is routed through a scoped admin workflow that rechecks pool.operate capability and current identity visibility at execution time.

Assignment lifecycle and race handling

• Creates a new assignment directly as active / active health / eligible.
• Reuses an existing active assignment idempotently.
• Promotes an existing pending row instead of leaving the new attachment non-operational.
• Restores a soft-deleted row in the same Pool/identity slot rather than violating the unique key.
• Rejects deleted upstream identities.
• Serializes assignment through an upstream-identity row lock and locks an existing assignment row FOR UPDATE, preventing concurrent admin clicks or delete/reactivate overlap from producing duplicate/stale outcomes.
• Clears cooldown and disabled timestamps when restoring an assignment.

Coordinated side effects

The scoped workflow also:

• records upstream_account.assign_pool in the audit log;
• broadcasts an upstream_assignment_assigned invalidation event;
• best-effort enqueues a manual catalog sync so the target Pool does not retain a stale model catalog;
• logs only a sanitized enqueue failure code if catalog scheduling fails.

Reactivation fix

Account reactivation now treats a pending Pool assignment as reactivatable. This covers the attach-after-detach sequence where the preserved/source assignment was removed and the newly attached target assignment is still pending. Deleted assignments remain deleted.

Validation

• mix compile --warnings-as-errors - passed
• Focused workflow/lifecycle/LiveView command - 5 passed, 0 failures
  • scoped authorization, audit event, broadcast, and catalog enqueue
  • denied target-Pool access has no database/job/audit side effects
  • active/deleted/pending assignment reuse and deleted-identity rejection
  • pending assignment promotion during account reactivation
  • unassigned account visibility, explicit Pool-filter exclusion, dialog, and database assignment
• mix test test/codex_pooler/audit_test.exs - 4 passed, 0 failures
• Strict Credo on all changed source and test files - no issues
• git diff --check - passed

Scope boundaries

• No database migration is required.
• No production deployment or data cleanup was performed.
• Partial catalog-sync resilience and stale-refreshing recovery are separate PRs.
• The proposed gateway side_effects.ex Oban-state change is intentionally excluded; duplicate gateway reconciliation is tracked separately in issue Bug: investigate duplicate gateway reconciliation and token-refresh runs #166.
• Routing-details/circuit-state UI and other unrelated admin changes are excluded.

[erwill2]

cleaning this up to resolve conflicts