fix(delivery): unify typed claim diagnostics before writeback

docs/architecture/rfcs/typescript-control-plane-migration-v0.md

@@ -182,6 +182,18 @@ label; no legacy prediction is retained without a concrete display consumer.
   State-only refresh remains legal without a delivery claim; this patch does
   not require every status refresh to declare progress. Existing write-time
   enum rejection, settlement evidence, quota, and gate checks remain in force.
+- Follow-up claim validation shares one TS diagnosis with historical projection.
+  New writes reject progress paired with preparation-only work, primary outcome
+  paired with blocker work/typed blocked observation, and primary outcome paired
+  with an explicit follow-through requirement. Historical conflicts remain
+  readable as `unknown` plus `delivery_claim_conflicts`; persisted records and
+  settled receipts are never rewritten. Partial progress and valid blocker
+  writebacks remain legal. This is an intentional authoring/readback behavior
+  correction, not a change to small-delivery policy or a new evidence validator.
+  Refresh validates individual fields in their established order, then checks
+  the normalized claim before registry access or lock creation. Invalid input
+  therefore takes precedence over store errors, including in dry-run mode;
+  state-dependent admission and writeback still share the same runtime lock.
 - Legacy outcome-marker/hint configuration remains readable and preserves
   whether an outcome floor is configured. Its words no longer classify runs.
   No persisted history is rewritten and no new default-off flag restores the

docs/architecture/rfcs/typescript-control-plane-migration-v0.zh-CN.md

@@ -141,6 +141,14 @@ classification 保留为历史标签；没有明确展示消费者时，不保
 - 历史字段缺失或不受支持时保持 unknown。unknown 中断连续小规模／outcome-gap
   证据计数，不视为成功或推断出的失败。未配置 floor 且没有 outcome 时，保留
   `not_configured` 展示哨兵值。
+- 后续声明校验与历史投影共用一份 TS 诊断：新写入拒绝“进展 + 仅准备”、
+  “主要成果 + blocker／typed blocked observation”和“主要成果 + 显式继续义务”
+  的矛盾组合。历史冲突以 `unknown` 和 `delivery_claim_conflicts` 保持可读，
+  不改写持久化记录或已结算 receipt；合法局部进展和 blocker 写回不受影响。
+  这是有意的写入／读取行为修正，不调整小步交付策略，也不是新的证据验证器。
+  Refresh 先按既有顺序校验各字段，再以归一化结果检查组合语义，之后才读取
+  registry 和创建锁。因此非法输入优先于存储错误返回，dry-run 也一致；依赖
+  当前状态的准入与写回仍在同一 runtime 锁内完成。
 - 新交付声明通过现有 writer API 写显式 enum，例如
   `refresh-state --delivery-outcome ... --delivery-batch-scale ...`。
   纯状态刷新仍可不声明交付；本批不强迫每次刷新声明进展。既有写入 enum 校验、

examples/shared-goal-authority-e2e/mutants.py

@@ -203,11 +203,8 @@ def apply(source: str) -> str:
          move_guard_outside_lock("require_registry_source_write_allowed")),),
          WRITER_TEST + "test_waiting_override_writer_rechecks_registry_binding_inside_shared_state_lock"),
     Case("remove_refresh_cas", (("loopx/state_refresh.py", replacement(
-        '''            if current_state_text != expected_write_state_text:
-                raise ValueError(
-                    "active goal state changed while refresh-state was qualifying "
-                    "its semantic writeback; retry from the current state"
-                )''', "")),),
+        "if current_state_text != expected_write_state_text:",
+        "if False:  # DELIBERATE MUTANT: bypass stale-state rejection.")),),
          WRITER_TEST + "test_concurrent_public_refresh_preserves_the_newer_owned_paragraph"),
     Case("fence_unshared_state_lock", ((COORDINATION + "legacy_writer_fence.ts", replacement(
         "withFileMutationLock(statePath, () =>",

loopx/control_plane/effect_runtime_handlers.ts

@@ -96,6 +96,7 @@ import {
 } from "./work_items/action_portfolio.ts";
 import { projectQuotaPlanningHorizon } from "./work_items/planning_horizon.ts";
 import { projectDeliveryHistory } from "./work_items/delivery_history.ts";
+import { validateDeliveryClaim } from "./work_items/delivery_outcome.ts";
 import {
   evaluateTaskLeaseAcquireDecision,
   evaluateTaskLeaseWriteScopesOverlap,
@@ -410,6 +411,7 @@ export function createEffectRuntimeHandlers(
     ["work_item.planning_inventory.detail", projectTodoPlanningInventoryDetail],
     ["work_item.refresh_recommendation.resolve", resolveRefreshRecommendation],
     ["work_item.delivery_history.project", projectDeliveryHistory],
+    ["work_item.delivery_claim.validate", validateDeliveryClaim],
     ["goal.vision_checkpoint.evaluate", buildVisionCheckpoint],
     ["goal.vision_wait.coverage", projectVisionWaitCoverage],
     ["goal.shared_goal_alignment.project", projectSharedGoalAlignment],

loopx/control_plane/work_items/delivery_history.py

@@ -1,4 +1,4 @@
-"""Compact transport for the typed delivery-history read model, never a writer."""
+"""Compact transport for typed delivery diagnostics and pre-write validation."""
 from __future__ import annotations
 
 from collections.abc import Mapping
@@ -70,3 +70,14 @@ def project_delivery_history(
             # Display-only annotation after the decision; narrative never enters TS.
             hint["latest_classification"] = _text(run.get("classification")).strip()
     return result
+
+
+def require_consistent_delivery_claim(record: Mapping[str, Any]) -> None:
+    """Reject contradictory authored claims before effects; add no new fields."""
+    if not any(record.get(key) for key in ("delivery_outcome", "delivery_turn_kind", "outcome_followthrough_required")):
+        return
+    result = effect_runtime_result("work_item.delivery_claim.validate", _run_facts(record))
+    if not isinstance(result, dict) or result.get("schema_version") != "delivery_claim_validation_v0":
+        raise RuntimeError("TypeScript delivery claim validation shape mismatch")
+    if result.get("valid") is not True:
+        raise ValueError("contradictory delivery claim: " + ", ".join(result.get("conflicts") or []))

loopx/control_plane/work_items/delivery_history.ts

@@ -1,7 +1,7 @@
 import type { JsonObject } from "../effect_program.ts";
 import { EffectRuntimeRequestError } from "../effect_runtime_errors.ts";
 import { requireJsonObject } from "../runtime_decode.ts";
-import { DELIVERY_OUTCOMES, isTurnScopedSettlementOutcome, type DeliveryOutcome } from "./delivery_outcome.ts";
+import { DELIVERY_OUTCOMES, diagnoseDeliveryClaim, isTurnScopedSettlementOutcome, type DeliveryOutcome } from "./delivery_outcome.ts";
 
 const TURN_KINDS = [
   "contract_only_preparation", "compact_evidence", "blocker_writeback",
@@ -117,6 +117,12 @@ export function projectDeliveryHistory(value: unknown): JsonObject {
   }
   const floorConfigured = input.outcome_floor_configured;
   const runs: DeliverySignal[] = input.runs.map(decodeRun).map((run) => {
+    const conflicts = diagnoseDeliveryClaim({ ...run });
+    if (conflicts.length > 0) return {
+      delivery_outcome: "unknown", delivery_batch_scale: batchScale(run.delivery_batch_scale),
+      delivery_turn_kind: "unknown", outcome_followthrough: null,
+      delivery_claim_conflicts: conflicts,
+    };
     const outcome = outcomeSignal(run.delivery_outcome, floorConfigured);
     const kind = turnKind(run, outcome);
     return {

loopx/control_plane/work_items/delivery_outcome.ts

@@ -1,4 +1,6 @@
 import { EffectRuntimeRequestError } from "../effect_runtime_errors.ts";
+import { requireJsonObject } from "../runtime_decode.ts";
+import type { JsonObject } from "../effect_program.ts";
 
 export const DELIVERY_OUTCOMES = [
   "surface_only",
@@ -17,6 +19,35 @@ export type DeliveryOutcome = (typeof DELIVERY_OUTCOMES)[number];
 export type MaterialDeliveryOutcome =
   (typeof MATERIAL_DELIVERY_OUTCOMES)[number];
 
+export type DeliveryClaimConflict =
+  | "progress_with_preparation_only"
+  | "primary_outcome_with_blocker"
+  | "primary_outcome_with_followthrough";
+
+/** Validate combinations, not prose or evidence truth. Historical readers use
+ * these same codes without rewriting already committed records or receipts. */
+export function diagnoseDeliveryClaim(raw: JsonObject): DeliveryClaimConflict[] {
+  const outcome = String(raw.delivery_outcome ?? "").trim();
+  const kind = String(raw.delivery_turn_kind ?? "").trim();
+  const conflicts: DeliveryClaimConflict[] = [];
+  if ((outcome === "outcome_progress" || outcome === "primary_goal_outcome")
+    && kind === "contract_only_preparation") conflicts.push("progress_with_preparation_only");
+  const observation = jsonObject(raw.progress_observation);
+  if (outcome === "primary_goal_outcome" && (kind === "blocker_writeback"
+    || (observation?.schema_version === "typed_progress_observation_v0" && observation.result_class === "blocked"))) {
+    conflicts.push("primary_outcome_with_blocker");
+  }
+  if (outcome === "primary_goal_outcome" && raw.outcome_followthrough_required === true) {
+    conflicts.push("primary_outcome_with_followthrough");
+  }
+  return conflicts;
+}
+
+export function validateDeliveryClaim(value: unknown): JsonObject {
+  const conflicts = diagnoseDeliveryClaim(requireJsonObject(value, "delivery claim"));
+  return { schema_version: "delivery_claim_validation_v0", valid: conflicts.length === 0, conflicts };
+}
+
 const STABLE_ID_PATTERN = /^[A-Za-z0-9][A-Za-z0-9._:/-]{0,127}$/;
 const PROGRESS_DELIVERY_OUTCOMES = new Set<DeliveryOutcome>([
   "outcome_progress",

loopx/history.py

@@ -110,6 +110,10 @@ def write_reserved_run_artifacts(
     render_markdown: Callable[[dict[str, Any]], str],
 ) -> None:
     from .control_plane.quota.usage_collector import ingest_usage_into_run_record
+    from .control_plane.work_items.delivery_history import require_consistent_delivery_claim
+
+    require_consistent_delivery_claim(record)
+    require_consistent_delivery_claim(index_record)
 
     # Producer call site for GH-C95: normalize any compact usage measurement onto
     # the durable run record and its index row before append. Fail closed here so