Close v0.11.0 review findings: Plan continuation and output-limit ordering #237

The independent whole-release review of PR #236 at 5388d096196b87a9fbb4fdff91cb0b6adfcd5741 identified two MEDIUM findings. This is the single correction-batch tracker required by the repository review policy. It tracks repairs to existing v0.11.0 contracts, not additional feature scope.

R1: current checkpoint Plan decisions before bootstrap

go/context/plan_preparation.go rejects valid checkpoint-3 records before applying a pending Plan decision. Simply allowing version 3 is insufficient: task schedules, held claims, child evidence, original native allowance and event versions must survive the decision and restart.

The real native in-provider Plan pause also leaves its original provider outcome unknown. The correction must retain that audit record and require exact original invocation release and durable pending-Plan/decision evidence before permitting a distinct subsequent invocation. Approval or a checkpoint marker alone must not authorize replay or new allowance. A completed-provider/out-of-provider Plan fixture is only a control, not proof of the actual native pause path.

Existing scope: #219, interacting with #183 and checkpoint-3 features.

R2: invalid output limit before classifier dispatch

go/adapter/eino/run.go resolves the optional route classifier before validating RequestOutputCeiling. Both a negative value and a value above the resolved reserve dispatch one classifier request before the named configuration refusal.

Validate the limit before physical generation while preserving valid one-token limits, fallback encoding, classifier bounds and normal routing.

Existing scope: #199 and #172.

Acceptance

• Actual model and native checkpoint-3 Plan accept/reject preserve exact state, task claims and original bounds through a fresh writer and subsequent adapter.
• Real native callback → Plan pause → exact release → restart → explicit decision → distinct invocation passes through the basic RunStore contract, without requiring optional HistoryRunStore.
• Missing, stale, wrong, duplicate, malformed or uncertain release/decision evidence remains closed; unrelated unknown operations remain blocked and normal allowance is consumed.
• Invalid output limits produce the named refusal with zero physical classifier/model calls; restoring the old order fails the intended assertion.
• Paired schemas/readers and focused positive/refusal/safeguard-removal checks pass. Published historical record formats remain frozen.
• Grouped exact-head CI passes and the original independent reviewer confirms the findings closed before merge.

The same fix round also corrects LOW report-2 feedback wording and finalizes the reviewed design-only ADR-0025 acceptance status. The original complete-release review and finding dispositions remain part of PR #236's review record; this tracker does not replace them. The review is explicitly same-vendor OpenAI/Codex under the approved exception.