fix(agent): consult credential pool in _restore_primary_runtime to avoid stale snapshot key

[dusterbloom]

Summary

Fixes #25205

_restore_primary_runtime() blindly restores api_key from the init-time snapshot, bypassing any credential pool rotation that occurred during the previous turn. After a 401/429/402 triggers pool rotation, the next turn restores the exhausted key, immediately fails again, and cascades into unnecessary cross-provider fallback.

Root Cause

1. Init: resolve_runtime_provider() selects credential A → snapshotted in _primary_runtime["api_key"]
2. Turn N: API returns 401 → pool marks A exhausted → _swap_credential() updates self.api_key to credential B → but _primary_runtime is never updated
3. Turn N+1: _restore_primary_runtime() restores stale credential A from snapshot → immediate 401 again

The same bug exists in _try_recover_primary_transport().

Fix

Both _restore_primary_runtime() and _try_recover_primary_transport() now consult the credential pool after restoring snapshot state:

pool_entry = pool.current() or pool.select()
if pool_entry is not None:
    self._swap_credential(pool_entry)  # fresh key + client rebuild
else:
    # fall back to snapshot behavior (no pool available)

• pool.current() is tried first (same entry used last turn)
• pool.select() is the fallback (picks best available)
• Exceptions from pool access are caught gracefully — falls back to snapshot
• When no pool exists, behavior is unchanged

Testing

• All 31 existing test_primary_runtime_restore.py tests pass
• All 47 test_credential_pool.py tests pass
• 4 new tests in TestCredentialPoolBypass:
  • test_restore_uses_pool_current_entry — verifies pool entry used instead of snapshot
  • test_restore_falls_back_to_snapshot_when_no_pool — no-pool backward compat
  • test_restore_handles_pool_exception_gracefully — pool error → snapshot fallback
  • test_restore_prefers_current_over_select — current() called first

[alt-glitch]

Competing fix: PR #25206 also targets the same _restore_primary_runtime stale-key issue (#25205). Both consult the credential pool after snapshot restore. #25733 takes a broader approach (centralized snapshots + align_current_to_runtime). These three PRs overlap significantly on the restore path — likely only one should land.

Related: #15434 (same function, cooldown check) and #25277 (pool in switch_model).

[dusterbloom]

Closing as stale: branch has fallen behind main and is conflicting, with no active review. Will reopen with a clean rebase if the change is still wanted.

[teknium1]

Closing as a duplicate of #25206 (by @jmmaloney4, submitted first, who also reported the underlying issue #25205). Both PRs fixed the same bug — _restore_primary_runtime restoring a stale credential snapshot instead of consulting the credential pool.

Merged via #53913 (commit f0de4c6 on main). Thanks for the contribution and for independently catching this.