fix: clear stale MoveFailed once the cluster is balanced

[eightynine01]

What

reconcileRebalance raises Degraded=True with reason MoveFailed when a move
command fails, but the only path that lowers it again is the active-move settle
branch. When the plan later becomes empty — the cluster reached balance on its
own, or the move stopped being necessary — the function returns phaseRunning
and leaves the condition standing.

The balanced branch now clears the condition, but only when the reason is the
one this function owns. DrainBlocked and ImmutableFieldChanged are raised
elsewhere; lowering them here would swallow those signals.

Why

Observed on a production cluster from 2026-08-22 to 2026-08-26: Degraded=True
for four days while every collection reported green and the shard named in the
condition message was Active on both peers. The message pointed at a consensus
commit timeout that had long since stopped mattering.

A condition that can be raised but never lowered is worse than no condition,
because it hides the next real fault behind a red light the operator has learned
to ignore.

Tests

internal/controller/rebalance_degraded_test.go:

• TestClearMoveFailed_자기사유만회수 — the owned reason is cleared, backoff reset.
• TestClearMoveFailed_타사유는불가침 — DrainBlocked / ImmutableFieldChanged
  survive untouched, backoff preserved.
• TestReconcileRebalance_균형이면Degraded회수 — the real entry point, with a
  balanced observation, clears a pre-existing MoveFailed.

Verified the guard fails without the fix:

--- FAIL: TestReconcileRebalance_균형이면Degraded회수
    균형 도달 시 MoveFailed 회수 실패: status=True reason=MoveFailed

go vet ./... clean, gofmt clean. TestControllers needs envtest binaries and
fails identically on an unmodified tree in this environment.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VbPL3ACkTc18bnwYUd7G3w

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.