Exclude superseded producer epochs from maintenance readiness

[zzylol]

Why

After a producer partition restarts, the coordinator could still use its old barrier to report readiness and return both old and new epoch inputs in the same batch.

What

Select one active epoch per partition from both staged inputs and watermarks. Ready batches contain only that epoch, and newly arriving superseded-epoch inputs are rejected before checkpoint mutation.

How

Readiness and barrier completeness use the same epoch map. Observing a new staged epoch invalidates the prior epoch's readiness immediately; the new epoch must supply its own inputs and barriers. Existing identical retries remain idempotent. Historical metadata is retained.

Before this PR

Stage epoch 1 inputs and barriers, then stage epoch 2 inputs without barriers: the coordinator incorrectly returns ready, including stale contributions.

After this PR

That batch waits for both epoch 2 barriers and contains exactly the two epoch 2 inputs. Restart preserves the decision. A new epoch 1 input cannot mutate staged state after epoch 2 is observed.

Verification

• The new regression fails on the original implementation at the premature-readiness assertion.
• All 6 coordinator tests pass after the fix with actual sketchlib refactor(legacy_expr): step β — thread Schema through legacy planning stack #139 (2.40 s).
• Independent source review passed for 78001618.

Limits

This fixes the existing coordinator's readiness contract; it does not wire continuous ingestion or claim automatic maintenance completion. Superseded metadata is retained for audit/retry, not garbage-collected. Recovery of uncommitted older-epoch work needs an explicit valid completion/replay proof; this change does not infer one.