@smthrs/engine-store 0.0.0-stage → 1.0.0-rc.3

package/src/internal/ActionPersistence.ts

@@ -0,0 +1,2687 @@
+/**
+ * Durable action dispatch at the engine encoded seam.
+ *
+ * Governing design: `packages/smithers/flows/plan/docs/concepts/step-keys.md`.
+ *
+ * @since 0.1.0
+ */
+
+import { Sha256 } from "@smthrs/crypto"
+import { FlowEngine } from "@smthrs/engine"
+import { Action, FlowRuntime } from "@smthrs/flow"
+import * as CacheEnvironment from "@smthrs/flow/CacheEnvironment"
+import type { FileBoundary } from "@smthrs/flow/FileBoundary"
+import { Journal, type JournalEvent, Redaction } from "@smthrs/journal"
+import { Jj } from "@smthrs/kernel"
+import { DerivedKey } from "@smthrs/keys"
+import { AttemptStore, Ownership, RunStore } from "@smthrs/run-store"
+import { CacheStore } from "@smthrs/step-cache"
+import * as Cause from "effect/Cause"
+import * as Clock from "effect/Clock"
+import * as Context from "effect/Context"
+import * as Effect from "effect/Effect"
+import * as Exit from "effect/Exit"
+import * as Metric from "effect/Metric"
+import * as Option from "effect/Option"
+import * as Schema from "effect/Schema"
+import * as EngineStoreMetrics from "../EngineStoreMetrics.ts"
+import * as Inconsistency from "../Inconsistency.ts"
+import * as ReplayOnly from "../ReplayOnly.ts"
+import * as StepBoundary from "../StepBoundary.ts"
+import * as StepSandbox from "../StepSandbox.ts"
+import * as WorkspaceSandbox from "../WorkspaceSandbox.ts"
+import * as AttemptAdmission from "./AttemptAdmission.ts"
+import * as CacheAdmission from "./CacheAdmission.ts"
+import * as CacheAgeHistory from "./CacheAgeHistory.ts"
+import * as CacheAgeVerdicts from "./CacheAgeVerdicts.ts"
+import * as CacheOutputPolicy from "./CacheOutputPolicy.ts"
+import * as CachePublication from "./CachePublication.ts"
+import * as CallFacts from "./CallFacts.ts"
+import * as CopiedRecord from "./CopiedRecord.ts"
+import * as EffectRecords from "./EffectRecords.ts"
+import * as HostReflection from "./HostReflection.ts"
+import * as JournalRecords from "./JournalRecords.ts"
+import * as ProvenanceSlot from "./ProvenanceSlot.ts"
+import * as SandboxedExecution from "./SandboxedExecution.ts"
+import * as StepFacts from "./StepFacts.ts"
+
+/**
+ * The boundary declaration an action may carry alongside its input.
+ *
+ * Aliased to `FileBoundary` rather than re-declared so the dispatch path and
+ * the `@smthrs/flow` declaration cannot drift apart; it is a distinct name only
+ * because "metadata" is how the action input refers to it.
+ *
+ * @since 0.1.0
+ * @category models
+ */
+export type BoundaryMetadata = FileBoundary
+
+/**
+ * Everything one dispatch of a durable action needs: the opaque `action`
+ * body, which `attempt` this is, the step `key` it is cached under, its trust
+ * `tier`, and the boundary it declared, if any.
+ *
+ * The `key` is a digest of the caller's declaration, so two dispatches sharing
+ * one carry the same claim about their inputs — which is what makes the
+ * attempt row and the cache row addressable by it.
+ *
+ * @since 0.1.0
+ * @category models
+ */
+export interface ActionInput {
+  readonly action: unknown
+  readonly attempt: number
+  readonly key: string
+  readonly tier: Action.Tier
+  /**
+   * The caller declares this sealed step's recorded result is one of many
+   * legitimate results; a same-key divergence is expected, not a hermeticity
+   * violation.
+   */
+  readonly nondeterministic?: true | undefined
+  readonly metadata?: BoundaryMetadata | undefined
+  /**
+   * The engine's `SnapshotBoundary` preparation for a compensable attempt:
+   * restore the retry sequence's original handle, then take this attempt's.
+   * Evaluated only when the attempt actually executes, after the journal and
+   * cache lookups, and its handle is persisted in the attempt row before the
+   * body runs.
+   */
+  readonly snapshot?: Effect.Effect<unknown> | undefined
+}
+
+/**
+ * The attempt stopped without settling — it is durably parked, not failed.
+ *
+ * The distinction matters to the driver: a suspended attempt keeps its row and
+ * its attempt number, so resuming continues the same attempt rather than
+ * burning a new one against the retry budget.
+ *
+ * @since 0.1.0
+ * @category errors
+ */
+export class AttemptSuspended extends Schema.TaggedError<AttemptSuspended>()(
+  "@smthrs/engine-store/AttemptSuspended",
+  {
+    code: Schema.Literal("attempt_suspended"),
+    runId: Schema.String,
+    keyDigest: Schema.String,
+    attempt: Schema.Int
+  }
+) {}
+
+/**
+ * The attempt was not admitted, so its body never ran.
+ *
+ * `outcome` names which admission check refused — a superseded fence, a live
+ * same-key attempt, an already-settled row. Because the body did not execute,
+ * this failure is always safe to surface without compensation.
+ *
+ * @since 0.1.0
+ * @category errors
+ */
+export class AttemptAdmissionRejected extends Schema.TaggedError<AttemptAdmissionRejected>()(
+  "@smthrs/engine-store/AttemptAdmissionRejected",
+  {
+    code: Schema.Literal("attempt_admission_rejected"),
+    keyDigest: Schema.String,
+    outcome: Schema.String
+  }
+) {}
+
+/**
+ * Two different runs recorded results under the same step key.
+ *
+ * The key is a digest of the declaration, so a conflict means the declaration
+ * does not fully describe what the step depends on: same key, different
+ * answer. `recordedRunId` names the run that got there first, so the
+ * divergence can be investigated rather than silently resolved.
+ *
+ * @since 0.1.0
+ * @category errors
+ */
+export class CacheConflictDetected extends Schema.TaggedError<CacheConflictDetected>()(
+  "@smthrs/engine-store/CacheConflictDetected",
+  {
+    code: Schema.Literal("cache_conflict_detected"),
+    keyDigest: Schema.String,
+    recordedRunId: Schema.String
+  }
+) {}
+
+/**
+ * A cached output's bytes no longer hash to the digest recorded for them.
+ *
+ * Unlike a succeeded attempt row, a shared cache row is evictable: the entry is
+ * dropped and the next dispatch re-executes and re-captures cleanly (issue
+ * #164). Reported rather than swallowed so a failing disk is visible.
+ *
+ * @since 0.1.0
+ * @category errors
+ */
+export class CacheCorruptionDetected extends Schema.TaggedError<CacheCorruptionDetected>()(
+  "@smthrs/engine-store/CacheCorruptionDetected",
+  {
+    code: Schema.Literal("cache_corruption_detected"),
+    keyDigest: Schema.String,
+    path: Schema.String,
+    recordedDigest: Schema.String,
+    measuredDigest: Schema.String
+  }
+) {}
+
+/**
+ * Corrupt recorded evidence on a SUCCEEDED durable attempt row under the
+ * strict verdict (issue #171).
+ *
+ * Distinct from {@link CacheCorruptionDetected} on purpose: a corrupt shared
+ * cache row is evictable — the next dispatch re-executes and re-captures
+ * cleanly (issue #164) — but a succeeded attempt row records that this run's
+ * side effects already ran, so eviction/re-execution would violate
+ * exactly-once for irreversible actions. The corrupt boundary evidence is
+ * quarantined off the succeeded row instead: the driver parks the first
+ * detection in the `quarantine` waiting state, and the next explicit resume
+ * returns the durable outcome without re-materializing the poisoned evidence
+ * or re-executing the action.
+ *
+ * @since 0.1.0
+ * @category errors
+ */
+export class AttemptEvidenceQuarantined extends Schema.TaggedError<AttemptEvidenceQuarantined>()(
+  "@smthrs/engine-store/AttemptEvidenceQuarantined",
+  {
+    code: Schema.Literal("attempt_evidence_quarantined"),
+    keyDigest: Schema.String,
+    attempt: Schema.Int,
+    path: Schema.String,
+    recordedDigest: Schema.String,
+    measuredDigest: Schema.String
+  }
+) {}
+
+/**
+ * Extracts an {@link AttemptEvidenceQuarantined} carried anywhere in a flow
+ * result's cause — as a typed failure or squashed into a defect — so the
+ * driver can park the run instead of settling it `failed` (issue #171).
+ *
+ * @since 0.1.0
+ * @category errors
+ */
+export const evidenceQuarantined = (
+  cause: Cause.Cause<unknown>
+): AttemptEvidenceQuarantined | undefined => {
+  for (const reason of cause.reasons) {
+    const carried = Cause.isFailReason(reason)
+      ? reason.error
+      : Cause.isDieReason(reason)
+      ? reason.defect
+      : undefined
+    if (carried instanceof AttemptEvidenceQuarantined) {
+      return carried
+    }
+  }
+  return undefined
+}
+
+/**
+ * The one classification of a `replayOutputs` failure (issue #150): a
+ * `BoundaryCorruption` in the cause is on-disk corruption of recorded
+ * evidence — the store's strongest invariant violated — while anything else
+ * is a transient host refusal that stays retryable. The two previously
+ * journalled identically, so a failing disk corrupting many blobs was
+ * indistinguishable from a one-off EIO.
+ */
+const replayCorruption = (
+  cause: Cause.Cause<unknown>
+): StepBoundary.BoundaryCorruption | undefined => {
+  for (const reason of cause.reasons) {
+    if (Cause.isFailReason(reason) && reason.error instanceof StepBoundary.BoundaryCorruption) {
+      return reason.error
+    }
+  }
+  return undefined
+}
+
+/**
+ * The stable effect kind a compensation handler is registered under.
+ *
+ * The action name is the adapter's own identity and is what an engine
+ * composition registers its handler by; a dispatch whose action carries no
+ * name at all (the plan scheduler's synthetic node dispatch) falls back to a
+ * constant, which resolves to no handler and therefore assesses as blocking —
+ * the safe direction.
+ */
+const actionKind = (action: unknown): string =>
+  typeof action === "object" && action !== null && typeof (action as { name?: unknown }).name === "string"
+    ? (action as { readonly name: string }).name
+    : "flows/engine-store/action"
+
+/**
+ * Whether a failing execution failed because it broke its declared boundary.
+ *
+ * The isolated execution path raises the boundary's own `UndeclaredWrite`, so
+ * a violation detected while the body ran classifies exactly like one detected
+ * at settle time: the row records `hardViolation` and the journal gets the
+ * violation record, rather than the failure passing as an ordinary action
+ * error the retry policy might happily retry.
+ */
+const declarationViolated = (cause: Cause.Cause<unknown>): boolean =>
+  cause.reasons.some((reason) =>
+    Cause.isFailReason(reason) &&
+    (reason.error instanceof StepBoundary.UndeclaredWrite || reason.error instanceof StepSandbox.UndeclaredRead)
+  )
+
+/**
+ * Whether a failed settle carries one of the boundary's own contract
+ * violations, as opposed to a host refusal. Classification for the
+ * `boundarySettlements` counter only — the journal record stays the source
+ * of truth.
+ */
+const settlementViolated = (cause: Cause.Cause<unknown>): boolean =>
+  cause.reasons.some((reason) =>
+    Cause.isFailReason(reason) &&
+    (reason.error instanceof StepBoundary.UndeclaredWrite ||
+      reason.error instanceof StepBoundary.MissingDeclaredOutput ||
+      reason.error instanceof StepBoundary.SurvivingDeclaredRemoval)
+  )
+
+/**
+ * What the action dispatcher is constructed with: the run it belongs to, the
+ * ownership fence it writes under, and the `execute` function that actually
+ * runs an action body.
+ *
+ * Everything durable — the attempt row, the cache row, the journal record — is
+ * this module's job; `execute` is the only part it delegates, which is what
+ * keeps the persistence discipline in one place regardless of what a flow
+ * runtime does with the body.
+ *
+ * @since 0.1.0
+ * @category models
+ */
+export interface Dependencies {
+  readonly runId: string
+  readonly owner: Ownership.OwnerId
+  readonly sourceId: string
+  /** Runs one action body. The only part of a dispatch this module delegates. */
+  readonly execute: (input: ActionInput) => Effect.Effect<unknown, unknown>
+  /** Repository authority reserved for engine bookkeeping. */
+  readonly engineJj?: Jj.Jj | undefined
+  /**
+   * Makes a retry of an irreversible action recognizable downstream. Its
+   * absence is what
+   * {@link Action.IrreversibleRetryRequiresIdempotencyKey} reports.
+   */
+  readonly idempotencyKey?: string | undefined
+  /**
+   * The incarnation-wide admission mutex (issues #102, #103). `EngineStore`
+   * passes one shared instance per store incarnation so every dispatch of a
+   * given attempt key in this process serializes through it; when omitted, a
+   * fresh mutex private to this `make` call is used — correct only when all
+   * same-key dispatches share the returned executor.
+   */
+  readonly admission?: AttemptAdmission.Service | undefined
+  /** Shared by callers that construct an executor for every dispatch in a run. */
+  readonly cacheAgeVerdict?: ReturnType<typeof CacheAgeVerdicts.make> | undefined
+  /**
+   * Present on a replay-only engine: every settled dispatch is reported to it,
+   * and a dispatch that would execute a body dies with `WouldExecute` first.
+   */
+  readonly replayOnly?: ReplayOnly.Service | undefined
+}
+
+const AttemptMeta = Schema.Struct({
+  tier: Schema.Literals(["sealed", "compensable", "irreversible"]),
+  // Only a genuine action idempotency key permits automatic released retry.
+  // Old rows without this evidence remain fail-closed (issues #2982/#3409).
+  keyed: Schema.optional(Schema.Literal(true)),
+  /**
+   * The sealed declaration admits multiple legitimate recorded results under
+   * this key. Absence remains the durable determinism claim.
+   */
+  nondeterministic: Schema.optional(Schema.Literal(true)),
+  boundary: Schema.optional(StepBoundary.BoundaryEvidence),
+  /**
+   * The prepare-time measurement matched the caller-declared read set the
+   * step key was derived from (issue #106). Only such completions may enter
+   * the shared cache: a stale declaration executes against the *real*
+   * content, so recording its result under the declaration's key would hand
+   * a later, genuinely accurate run the wrong value as a verified hit.
+   */
+  readSetVerified: Schema.optional(Schema.Literal(true)),
+  /**
+   * The row's boundary evidence failed integrity verification and was removed
+   * after its detailed corruption record reached the journal (issue #171).
+   * The succeeded outcome remains authoritative, but this row may never be
+   * converged into the shared cache again.
+   */
+  boundaryQuarantined: Schema.optional(Schema.Literal(true)),
+  /**
+   * The failed row records a boundary violation (a prepare or settle
+   * failure), so the failed replay branch can re-emit the `hardViolation`
+   * journal record idempotently after a crash in the finish→emit window
+   * (issue #109) — the violation kind is not recoverable from the persisted
+   * cause alone.
+   */
+  hardViolation: Schema.optional(Schema.Literal(true)),
+  snapshotId: Schema.optional(Schema.String),
+  /**
+   * The jj operation that recorded `snapshotId`, when the host reports one.
+   * A whole-repository rewind restores it to undo bookmark moves, rebases,
+   * `describe`, and `abandon` that a tree restore leaves in place.
+   */
+  snapshotOperationId: Schema.optional(Schema.String),
+  /**
+   * The engine `SnapshotBoundary` handle taken before this attempt executed.
+   * Wrapped so an opaque `null` or `undefined` handle stays distinguishable
+   * from no handle; `handle` is absent when the handle was `undefined`.
+   */
+  boundarySnapshot: Schema.optional(Schema.Struct({ handle: Schema.optionalKey(Schema.Unknown) })),
+  // The incarnation that admitted the running row. Since issues #102/#103
+  // the adoption decision rests on the admission permit rather than this
+  // nonce — a live same-key fiber of this process would be holding the
+  // permit, which distinguishes a dead fiber from a live dispatch in a way
+  // the recorded incarnation cannot — but the field is kept as durable
+  // forensic evidence of which incarnation last drove the attempt.
+  admittedBy: Schema.optional(Ownership.OwnerId),
+  /**
+   * How far this attempt's effect boundary got, in EXECUTABLE state.
+   *
+   * The journal's boundary records already say this, and adoption cannot use
+   * them: `@smthrs/journal` puts every payload it writes through redaction,
+   * deliberately, so a recorded `output` may never be re-entered as a flow's
+   * result (issue #72), and the journal offers no read addressed by producer
+   * identity — only a full scan of the run. So the same fact is written here,
+   * where the attempt's `outcome` column is already replayed verbatim, in the
+   * SAME transaction as the record that describes it.
+   *
+   * - `intended`: the body was dispatched and its outcome is not durably
+   *   known. Whether the effect reached the outside world is genuinely
+   *   unanswerable from durable state.
+   * - `parked`: the body reached a durable wait and suspended, so it never
+   *   finished and nothing crossed. Recorded explicitly because a park leaves
+   *   the same `running` row a crash does.
+   * - `succeeded`: the body completed and the row carries its result.
+   */
+  effectCrossing: Schema.optional(Schema.Literals(["intended", "parked", "succeeded"]))
+})
+
+type AttemptMeta = typeof AttemptMeta.Type
+
+const decodeMeta = (value: unknown): AttemptMeta | undefined => {
+  const decoded = Schema.decodeUnknownResult(AttemptMeta)(value)
+  return decoded._tag === "Success" ? decoded.success : undefined
+}
+
+/**
+ * The earliest boundary snapshot handle persisted for an action key's
+ * surviving attempts, walking `earliest` through `latest`. An attempt row
+ * without one (it crashed before its snapshot was persisted) is skipped.
+ *
+ * @since 0.1.0
+ * @category combinators
+ */
+export const earliestBoundarySnapshot = (
+  attempts: AttemptStore.Service,
+  runId: string,
+  stepKeyDigest: string,
+  range: { readonly earliest: number; readonly latest: number }
+): Effect.Effect<Option.Option<unknown>> =>
+  Effect.gen(function*() {
+    for (let attempt = range.earliest; attempt <= range.latest; attempt++) {
+      const row = yield* attempts.get({ runId, stepKeyDigest, attempt }).pipe(Effect.orDie)
+      const persisted = Option.isSome(row) ? decodeMeta(row.value.meta)?.boundarySnapshot : undefined
+      if (persisted !== undefined) return Option.some(persisted.handle)
+    }
+    return Option.none()
+  })
+
+const CauseJson = Schema.Struct({
+  reasons: Schema.Array(Schema.Union([
+    Schema.Struct({ _tag: Schema.Literal("Fail"), error: Schema.Unknown }),
+    Schema.Struct({ _tag: Schema.Literal("Die"), defect: Schema.Unknown }),
+    Schema.Struct({
+      _tag: Schema.Literal("Interrupt"),
+      fiberId: Schema.optional(Schema.NullOr(Schema.Number))
+    })
+  ]))
+})
+
+/**
+ * Maximum nesting copied from one live failure value.
+ *
+ * The bounds here are the store's own admission limits, less the envelope
+ * `persistCause` wraps around a reduced value, and never tighter: a value the
+ * store would have persisted must not be thrown away here. `AttemptStore`
+ * admits 128 levels, and a reduced value sits three levels down, inside
+ * `{ reasons: [{ error }] }`.
+ *
+ * @since 1.0.0
+ * @category constants
+ */
+export const maxInertJsonDepth = AttemptStore.maximumJsonDepth - 4
+
+/**
+ * Maximum JSON values copied from one persisted cause.
+ *
+ * The budget is spent across the cause object, its reasons array, every reason
+ * envelope, and every reduced leaf. A cause that cannot fit collapses to one
+ * bounded reason before it reaches `AttemptStore`. Own-key enumeration has a
+ * separate allowance of the same size so ignored keys cannot make traversal
+ * unbounded.
+ *
+ * @since 1.0.0
+ * @category constants
+ */
+export const maxInertJsonNodes = AttemptStore.maximumJsonNodes - 8
+
+/**
+ * Maximum UTF-16 units copied from one persisted cause, keys included.
+ *
+ * Spent across every caller-supplied key and string leaf. This bounds text the
+ * reduction inspects and retains, not the serialized byte length: numeric
+ * spellings, punctuation, and the fixed cause envelope are outside this
+ * character count and are bounded separately by {@link maxInertJsonNodes}.
+ *
+ * @since 1.0.0
+ * @category constants
+ */
+export const maxInertJsonCharacters = 256 * 1024
+
+type InertJsonValue =
+  | null
+  | boolean
+  | number
+  | string
+  | Array<InertJsonValue>
+  | { [key: string]: InertJsonValue }
+
+const inertJsonRejected = Symbol("inertJsonRejected")
+const inertJsonOmitted = Symbol("inertJsonOmitted")
+const arrayBufferViewMarker = "[ArrayBufferView]"
+
+type InertJsonReduction = InertJsonValue | typeof inertJsonRejected | typeof inertJsonOmitted
+
+/** What one cause's reduction has already spent of its shared bounds. */
+interface InertJsonBudget {
+  nodes: number
+  characters: number
+  enumeratedKeys: number
+}
+
+/** Charges JSON values against the cause-wide node allowance. */
+const spendNodes = (budget: InertJsonBudget, nodes: number): boolean => {
+  if (budget.nodes + nodes > maxInertJsonNodes) return false
+  budget.nodes += nodes
+  return true
+}
+
+/** The store rejects unpaired surrogates, so an ill-formed string is refused. */
+const isWellFormedInertString = (value: string): boolean => {
+  for (let index = 0; index < value.length; index++) {
+    const unit = value.charCodeAt(index)
+    if (unit >= 0xd800 && unit <= 0xdbff) {
+      const low = value.charCodeAt(++index)
+      if (!(low >= 0xdc00 && low <= 0xdfff)) return false
+    } else if (unit >= 0xdc00 && unit <= 0xdfff) return false
+  }
+  return true
+}
+
+/**
+ * Charges one string, a key or a value, against the cause's character budget.
+ * Answers whether the string is admitted.
+ */
+const spend = (budget: InertJsonBudget, value: string): boolean => {
+  budget.characters += value.length
+  return budget.characters <= maxInertJsonCharacters && isWellFormedInertString(value)
+}
+
+/** Charges every own key before its type or descriptor can make it ignorable. */
+const spendEnumeratedKey = (budget: InertJsonBudget, key: PropertyKey): boolean => {
+  budget.enumeratedKeys += 1
+  if (budget.enumeratedKeys > maxInertJsonNodes) return false
+  return typeof key !== "string" || spend(budget, key)
+}
+
+/**
+ * Projects one failure value into the inert JSON data `AttemptStore` admits.
+ *
+ * A `Fail` reason raised by a service rather than by `Action.executeEncoded`
+ * carries a live class instance — `StepBoundary.UnsupportedBoundary`, a
+ * `StepSandbox` refusal, any `Data.TaggedError` — and a `Die` reason carries
+ * an arbitrary defect. `@smthrs/run-store`'s attempt boundary refuses a
+ * non-plain object outright rather than serializing it, so the projection has
+ * to happen on the write side, here, where the durable shape is already owned.
+ *
+ * The reduction reads only own data-property descriptors. It never invokes a
+ * getter or `toJSON`, and a host-detected proxy is rejected before any of its
+ * traps can run. An array-buffer view becomes the fixed
+ * `"[ArrayBufferView]"` marker before its indexed own keys can be materialized;
+ * its bytes and attached properties are deliberately omitted. JSON leaves,
+ * arrays, and enumerable string-keyed object data otherwise keep their ordinary
+ * JSON shape. Unsupported object members are omitted and unsupported array
+ * members become `null`, matching JSON's container rules. Every key returned
+ * by an own-key enumeration spends from a cause-wide allowance before its type,
+ * enumerability, or descriptor is inspected.
+ *
+ * Depth, node, and character bounds apply before the attempt store's admission
+ * boundary, and `budget` carries the allowances across every reason of one
+ * cause. The caller charges the root value as part of its reason envelope; this
+ * walk charges every descendant. A cycle, proxy, inspection refusal, or
+ * exceeded bound rejects the whole value as `null`, so persistence on the
+ * failure path neither executes user code nor leaves an attempt running because
+ * its error was too large for the store.
+ */
+const inertJson = (value: unknown, budget: InertJsonBudget): unknown => {
+  const active = new WeakSet<object>()
+
+  const reduce = (current: unknown, depth: number): InertJsonReduction => {
+    if (depth > maxInertJsonDepth) return inertJsonRejected
+    if (depth > 0 && !spendNodes(budget, 1)) return inertJsonRejected
+
+    if (current === null) return null
+    switch (typeof current) {
+      case "boolean":
+        return current
+      case "number":
+        return Number.isFinite(current) ? current : null
+      case "string":
+        return spend(budget, current) ? current : inertJsonRejected
+      case "undefined":
+      case "symbol":
+      case "function":
+      case "bigint":
+        return inertJsonOmitted
+      case "object": {
+        if (HostReflection.host.isProxy(current) || active.has(current)) return inertJsonRejected
+        if (ArrayBuffer.isView(current)) return arrayBufferViewMarker
+        active.add(current)
+        try {
+          if (Array.isArray(current)) {
+            const lengthDescriptor = Object.getOwnPropertyDescriptor(current, "length")
+            /* v8 ignore next 3 -- A non-proxy Array always has an own data-property uint32 length. */
+            if (
+              lengthDescriptor === undefined || !("value" in lengthDescriptor) ||
+              typeof lengthDescriptor.value !== "number"
+            ) return inertJsonRejected
+            const output: Array<InertJsonValue> = []
+            for (let index = 0; index < lengthDescriptor.value; index++) {
+              const descriptor = Object.getOwnPropertyDescriptor(current, String(index))
+              const reduced = reduce(
+                descriptor !== undefined && "value" in descriptor ? descriptor.value : undefined,
+                depth + 1
+              )
+              if (reduced === inertJsonRejected) return inertJsonRejected
+              output.push(reduced === inertJsonOmitted ? null : reduced)
+            }
+            return output
+          }
+
+          const output: Record<string, InertJsonValue> = {}
+          // This engine-owned error's native Error.message is non-enumerable,
+          // but is part of the durable schema used to classify a landing
+          // conflict. Preserve that one data field, never getters, stacks, or
+          // arbitrary hidden properties. It spends the same bounded budget.
+          const tag = Object.getOwnPropertyDescriptor(current, "_tag")
+          const conflictMessage = tag !== undefined && "value" in tag &&
+            tag.value === WorkspaceSandbox.MaterializationConflict.identifier
+          for (const key of Reflect.ownKeys(current)) {
+            if (!spendEnumeratedKey(budget, key)) return inertJsonRejected
+            if (typeof key !== "string") continue
+            const descriptor = Object.getOwnPropertyDescriptor(current, key)
+            /* v8 ignore next -- A key returned by Reflect.ownKeys exists on a non-proxy object. */
+            if (descriptor === undefined) continue
+            if ((!descriptor.enumerable && !(key === "message" && conflictMessage)) || !("value" in descriptor)) {
+              continue
+            }
+            // `message` and `stack` are observability text, never a field replay
+            // classifies on, and a git or HTTP failure embeds its remote URL,
+            // credentials included. They meet the journal's own rules; every
+            // other member, `_tag` included, stays byte-exact for replay.
+            const member = (key === "message" || key === "stack") && typeof descriptor.value === "string"
+              ? Redaction.redact(descriptor.value)
+              : descriptor.value
+            const reduced = reduce(member, depth + 1)
+            if (reduced === inertJsonRejected) return inertJsonRejected
+            if (reduced === inertJsonOmitted) continue
+            Object.defineProperty(output, key, {
+              value: reduced,
+              enumerable: true,
+              configurable: true,
+              writable: true
+            })
+          }
+          return output
+        } finally {
+          active.delete(current)
+        }
+      }
+    }
+    /* v8 ignore next -- The typeof switch covers every JavaScript value category. */
+    return inertJsonRejected
+  }
+
+  try {
+    const reduced = reduce(value, 0)
+    return reduced === inertJsonRejected || reduced === inertJsonOmitted ? null : reduced
+  } catch {
+    return null
+  }
+}
+
+/**
+ * Encodes a live `Cause` into the plain tagged-reason JSON `rehydrateCause`
+ * decodes. Persisting the `Cause` object itself left the durable shape to
+ * whatever the ambient serializer produced (`Cause.toJSON` emits
+ * `{_id, failures}`, a structural walk emits `{reasons}`), so a change in
+ * the store's encoding silently broke failed-attempt replay. The write side
+ * now owns the shape explicitly. {@link inertJson} makes every live leaf
+ * inert and bounded before the store sees it, against one budget shared by
+ * the fixed envelope and every reason. The loop stops at the first reason that
+ * cannot fit and returns one `Die(null)` reason, which {@link rehydrateCause}
+ * decodes on replay, so even an adversarial reason count remains admissible.
+ */
+const persistCause = (cause: Cause.Cause<unknown>): typeof CauseJson.Type => {
+  // The cause object and its `reasons` array are the two fixed envelope nodes.
+  const budget: InertJsonBudget = { nodes: 2, characters: 0, enumeratedKeys: 0 }
+  const reasons: Array<typeof CauseJson.Type.reasons[number]> = []
+  for (const reason of cause.reasons) {
+    // One object, its `_tag` string, and its error, defect, or fiber-id value.
+    if (!spendNodes(budget, 3)) {
+      return { reasons: [{ _tag: "Die", defect: null }] }
+    }
+    if (Cause.isFailReason(reason)) {
+      reasons.push({ _tag: "Fail", error: inertJson(reason.error, budget) })
+    } else if (Cause.isDieReason(reason)) {
+      reasons.push({ _tag: "Die", defect: inertJson(reason.defect, budget) })
+    } else {
+      reasons.push({ _tag: "Interrupt", fiberId: reason.fiberId ?? null })
+    }
+  }
+  return { reasons }
+}
+
+/**
+ * Presents a payload this module already reduced to JSON as the attempt
+ * store's `JsonValue`.
+ *
+ * `@smthrs/run-store` types the `meta`, `error`, and `outcome` columns as
+ * `JsonValue` and admits every write through its own inert-JSON boundary,
+ * which refuses anything that boundary rejects with `AttemptStoreError`. The
+ * values handed here are already reduced: `AttemptMeta` is a schema whose
+ * optional fields are omitted rather than written as `undefined`,
+ * {@link persistCause} passes every leaf through {@link inertJson}, and an
+ * action `outcome` is the encoded result the caller's own codec produced.
+ * TypeScript still refuses the assignment because an optional property's type
+ * includes `undefined` and a JSON index signature does not, and because a
+ * schema-level `unknown` leaf cannot be proven JSON statically. The store's
+ * admission is the check; this is the type-level statement of what it checks.
+ */
+const attemptPayload = (value: unknown): AttemptStore.JsonValue => value as AttemptStore.JsonValue
+
+/**
+ * Rebuilds the persisted failure of a `failed` attempt row so replay can
+ * rethrow the original domain error (issue #59). The row's `error` column
+ * holds the {@link persistCause} encoding of the failing `Cause` — a plain
+ * object whose `reasons` carry the tagged `Fail`/`Die`/`Interrupt` material.
+ * `Fail` errors are already schema-encoded by `Action.executeEncoded`, so
+ * their `_tag` survives and `RetryPolicy` non-retryable matching applies on
+ * replay exactly as it did on the live attempt. Live reasons are rebuilt
+ * unconditionally from the tagged material; anything unrecognizable becomes
+ * a defect carrying the raw persisted value.
+ */
+const rehydrateCause = (error: unknown): Cause.Cause<unknown> => {
+  const decoded = Schema.decodeUnknownResult(CauseJson)(error)
+  if (decoded._tag === "Success" && decoded.success.reasons.length > 0) {
+    return Cause.fromReasons(decoded.success.reasons.map((reason) => {
+      switch (reason._tag) {
+        case "Fail":
+          return Cause.makeFailReason(reason.error)
+        case "Die":
+          return Cause.makeDieReason(reason.defect)
+        case "Interrupt":
+          return Cause.makeInterruptReason(reason.fiberId ?? undefined)
+      }
+    }))
+  }
+  return Cause.die(error)
+}
+
+/**
+ * The {@link module:CacheEnvironment.CachePolicy} the dispatched declaration
+ * annotated, if any.
+ *
+ * The action stays opaque to this module — it is `unknown` on
+ * {@link ActionInput} on purpose — so the policy is read out of its annotation
+ * bag rather than off a field. An action without one, or a caller that hands
+ * the seam a bare object (every unit test of this module does), simply has no
+ * policy and keeps the pre-policy behavior.
+ *
+ * @since 0.1.0
+ * @private
+ */
+const declaredCachePolicy = (action: unknown): CacheEnvironment.CachePolicy | undefined => {
+  const annotations = (action as { readonly annotations?: unknown } | null | undefined)?.annotations
+  return Context.isContext(annotations) ? CacheEnvironment.cachePolicyOf(annotations) : undefined
+}
+
+/**
+ * Constructs the encoded action executor. The action itself stays opaque;
+ * the supplied dispatcher is the only physical execution point.
+ *
+ * @since 0.1.0
+ * @category constructors
+ */
+export const make = (deps: Dependencies) => {
+  /**
+   * Runs one action body, and tells whoever drove this dispatch that it ran.
+   *
+   * The report is the only way a node above this knows the difference between
+   * work and replay. Every execution path funnels through here, so a body
+   * that starts says so exactly once; a path that returns a durable record
+   * instead reports `replayed` at its own return. A dispatch that reports
+   * nothing is read as work, which is the safe reading: a node never claims
+   * it rebuilt nothing when it might have rebuilt something.
+   *
+   * The digest travels with the report because it is what joins the two
+   * halves of a node's evidence: the attempt records this dispatch writes are
+   * keyed by it, and the node record above has no other way to name them.
+   */
+  const execute = (input: ActionInput, stepKeyDigest: string) =>
+    Effect.andThen(
+      Action.reportDispatch({ outcome: "executed", stepKeyDigest, attempt: input.attempt }),
+      deps.execute(input)
+    )
+  const admission = deps.admission ?? AttemptAdmission.makeUnsafe()
+  const recordedAgeVerdict = deps.cacheAgeVerdict ?? CacheAgeVerdicts.make(deps.runId)
+  // The lineage every record this executor writes addresses itself to.
+  // An action is a node inside its run's root lineage, not a lineage of
+  // its own: a lineage segment is minted only where a separate run is.
+  const lineageId = FlowEngine.Lineage.root(deps.runId)
+  return Effect.fn("ActionPersistence.execute")((input: ActionInput) => {
+    // What a replay-only engine reports for this dispatch: the digest once it
+    // exists, and whether the dispatch stopped at the execution gate.
+    const replay: { stepKeyDigest?: string; gated: boolean } = { gated: false }
+    return Effect.gen(function*() {
+      yield* Effect.annotateCurrentSpan({
+        runId: deps.runId,
+        attempt: input.attempt,
+        tier: input.tier
+      })
+      const attempts = yield* AttemptStore.AttemptStore
+      const cache = yield* CacheStore.CacheStore
+      const journal = yield* Journal.Journal
+      const runs = yield* RunStore.RunStore
+      const policy = declaredCachePolicy(input.action)
+      // SCOPE NARROWS THE CACHE ADDRESS. The engine derives a cross-run key
+      // from the inputs and the environment, which is the widest reach a
+      // content address can have. A declaration that wants a narrower one
+      // folds the identity it is narrowing to into the address of its cache
+      // row, so a sibling run derives a different address and never finds the
+      // row: narrowing by construction rather than by a read-time filter that
+      // a second reader could disagree with. `shared` is the unnarrowed
+      // default and folds nothing, so every key predating the policy keeps
+      // its bytes.
+      const scopeTag = policy?.scope === "run"
+        ? `run:${deps.runId}`
+        : policy?.scope === "flow"
+        // The flow tag comes from the executing instance, which the engine
+        // provides around every dispatch. A caller that reaches this seam
+        // without one — the unit tests of this module — cannot name a flow, so
+        // the key narrows to the run instead: never wider than the caller
+        // asked for.
+        ? `flow:${
+          Option.getOrUndefined(
+            Option.map(yield* Effect.serviceOption(FlowRuntime.FlowInstance), (instance) => instance.flow._tag)
+          ) ?? `run:${deps.runId}`
+        }`
+        : undefined
+      // THE STEP KEY DIGEST IS NOT NARROWED. It is the identity the rest of the
+      // engine addresses this dispatch by: `flows_attempts` rows, which
+      // `EngineStore.actionRetryOrigin` and `actionLatestAttempt` probe under
+      // `sha256(input.key)`, and the `stepKeyDigest` on every attempt record,
+      // which `PlanScheduler` maps back to the node that dispatched it. Both
+      // derive the digest from the key alone and neither can see an
+      // annotation, so narrowing this one would hide a scoped step's attempts
+      // from its own retry counter and its deviations from the reconciler.
+      const stepKeyDigest = yield* Schema.decodeUnknownEffect(Sha256)(input.key).pipe(Effect.orDie)
+      replay.stepKeyDigest = stepKeyDigest
+      // What the scope narrows is the CACHE ROW ADDRESS, which is the only
+      // thing a scope is about: an unscoped declaration addresses the row by
+      // the step key alone, and a narrowed one folds the identity it is
+      // narrowing to into the address, so a sibling derives a different one and
+      // never finds the row. The two digests are the same string for every
+      // declaration that narrows nothing.
+      const keyDigest = scopeTag === undefined
+        ? stepKeyDigest
+        : yield* Schema.decodeUnknownEffect(Sha256)(`${input.key}\u0000scope:${scopeTag}`).pipe(Effect.orDie)
+      yield* Effect.annotateCurrentSpan({ stepKeyDigest, keyDigest })
+      const attemptId = { runId: deps.runId, stepKeyDigest, attempt: input.attempt }
+      /**
+       * Lifecycle events take the journal's durable channel, fenced to the
+       * owning process: the write commits inside the SQL sink while
+       * `flows_runs` still records `deps.owner`, so a reclaimed (zombie) run
+       * fails with `fence_lost` — surfaced as self-interruption, matching the
+       * store-level fence outcomes — and a saturated lossy queue can never
+       * drop an attempt lifecycle record (issue #10).
+       */
+      const emitLifecycle = (record: JournalEvent.Input) =>
+        journal.emitDurable(record, deps.owner).pipe(
+          Effect.catch((error) => error.code === "fence_lost" ? Effect.interrupt : Effect.fail(error))
+        )
+      const callFacts = CallFacts.make(input.action, deps.runId)
+      const stepFacts = StepFacts.make(input.action, deps.runId)
+      const emitStepSettled = (outcome: unknown) =>
+        Effect.flatMap(
+          stepFacts.settled(outcome),
+          (record) => record === undefined ? Effect.void : Effect.asVoid(emitLifecycle(record))
+        )
+      const emitCallInvoked = Effect.flatMap(
+        callFacts.invoked,
+        (record) => record === undefined ? Effect.void : Effect.asVoid(emitLifecycle(record))
+      )
+      const emitCallSettled = (outcome: unknown) =>
+        Effect.flatMap(
+          callFacts.settled(outcome),
+          (record) => record === undefined ? Effect.void : Effect.asVoid(emitLifecycle(record))
+        )
+      /**
+       * Commits an attempt/cache state transition and the lifecycle records
+       * describing it in ONE write transaction.
+       *
+       * The store writes below (`attempts.put`/`patch`/`finish`, `cache.put`)
+       * and `emitLifecycle` all go through the same `DurableWriter`, so the inner
+       * writes become savepoints of this transaction: either the row and its
+       * journal entry are both durable, or neither is. Before this, a crash in
+       * the interstitial left an attempt row the journal could not explain —
+       * `attemptStarted` with no `attemptFinished`, forever. Temporal's
+       * mutable state and history events commit as one persistence request
+       * (`reference/temporal/service/history/workflow/transaction_impl.go`);
+       * this is that unit of work, scoped to one attempt.
+       *
+       * Nothing that is not storage work belongs inside: the action body,
+       * the Jj snapshot, and the boundary prepare/settle all stay outside so
+       * the write transaction is never held across a host call.
+       */
+      const atomically = <A, E, R>(effect: Effect.Effect<A, E, R>) => journal.transact(effect)
+      /**
+       * Commits an attempt's terminal row and the lifecycle records describing
+       * it as one unit, reporting whether the fenced write landed.
+       *
+       * Returning a boolean rather than interrupting inside keeps the
+       * transaction's failure paths to genuine storage failures: a lost fence
+       * is an ordinary "someone else owns this attempt now" outcome, and its
+       * self-interruption belongs to the caller, outside the transaction.
+       */
+      const settleAttempt = (
+        row: AttemptStore.FinishAttempt,
+        records: ReadonlyArray<JournalEvent.Input>
+      ) =>
+        atomically(Effect.gen(function*() {
+          const finished = yield* attempts.finish(row, deps.owner)
+          if (finished._tag !== "Finished") {
+            return false
+          }
+          // The checkpoint fact precedes its completion boundary so a rewind
+          // retaining that completion also retains the observation it owns.
+          if (row.state === "succeeded" && stepFacts.annotated) {
+            const committed = yield* attempts.get(attemptId)
+            if (Option.isNone(committed)) {
+              return yield* Effect.fail(
+                new AttemptStore.AttemptStoreError({
+                  code: "persistence_failed",
+                  method: "get",
+                  message: "Finished checkpoint outcome disappeared",
+                  cause: attemptId
+                })
+              )
+            }
+            yield* emitStepSettled(committed.value.outcome)
+          }
+          yield* Effect.forEach(records, (record) => emitLifecycle(record), { discard: true })
+          if (row.state === "succeeded" && callFacts.settles) {
+            const committed = yield* attempts.get(attemptId)
+            if (Option.isNone(committed)) {
+              return yield* Effect.die(new Error("Finished call outcome disappeared"))
+            }
+            yield* emitCallSettled(committed.value.outcome)
+          }
+          return true
+        }))
+      const fencedAtMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+      const heartbeat = yield* runs.heartbeat(deps.runId, deps.owner, fencedAtMs)
+      if (heartbeat._tag !== "Updated") {
+        return yield* Effect.interrupt
+      }
+
+      if (input.tier === "irreversible" && input.attempt > 1 && deps.idempotencyKey === undefined) {
+        return yield* Effect.fail(
+          new Action.IrreversibleRetryRequiresIdempotencyKey({
+            actionName: actionKind(input.action),
+            attempt: input.attempt
+          })
+        )
+      }
+
+      // A source glob is cacheable only after the scheduler pins and replaces
+      // it with exact measured inputs. Direct actions retain the pattern, so
+      // their key cannot prove which expansion it names and must stay local.
+      const cacheDeclaration = CacheAdmission.declaration(input)
+      const declarationMeta = {
+        tier: input.tier,
+        ...(deps.idempotencyKey === undefined ? {} : { keyed: true as const }),
+        ...(input.nondeterministic === undefined ? {} : { nondeterministic: input.nondeterministic })
+      } satisfies AttemptMeta
+
+      /**
+       * Cache-provenance producer identity (issue #124): the plain
+       * `{runId, sourceId}` identity carried no `sourceSeq`, so every re-drive of
+       * the same observation — the same replay refusal against the same
+       * recorded row, on every later dispatch of the key — appended a fresh
+       * journal row forever. A per-(key, action, recorded-provenance)
+       * identity with `sourceSeq 0` collapses the identical re-observation
+       * into a `Duplicate`, while folding the recorded row's provenance into
+       * the identity keeps a genuinely new observation — the same action
+       * against a *different* recorded row — a distinct producer that still
+       * journals.
+       *
+       * The identity names the run, the step-key digest, the decision, and the
+       * recorded row's provenance — everything the decision is ABOUT — and
+       * deliberately nothing about WHO TOOK IT. `deps.sourceId` is the host's
+       * configured `journalSource` (`EngineStore.Options.journalSource`), which
+       * a run outlives: the process that admits a row at 900 ms is routinely
+       * not the process that resumes the run at 1100 ms. Folding it in made the
+       * `ttl` verdict below re-decidable — the resuming incarnation got
+       * `Accepted` for a fresh verdict under its own identity and re-judged the
+       * age against its own clock, which is the one thing a recorded verdict
+       * exists to prevent — and made every other record here re-appendable once
+       * per incarnation rather than once per observation.
+       *
+       * The sibling `recorded` identity below still carries `deps.sourceId`, on
+       * purpose: nothing reads a decision back out of it. It exists so a
+       * convergence re-record collapses into a `Duplicate`, and a resume under
+       * another journal source costs one extra provenance row there, never a
+       * different answer.
+       */
+      const cacheSource = (
+        action: string,
+        recorded?: { readonly runId: string; readonly eventSeq: number }
+      ): JournalRecords.EventOptions => ({
+        runId: deps.runId,
+        lineageId,
+        sourceId: `cache:${keyDigest}:${action}${
+          recorded === undefined ? "" : `:${recorded.runId}:${recorded.eventSeq}`
+        }`,
+        sourceSeq: 0
+      })
+      const noteUnshareable = (unshareable: CachePublication.Unshareable) =>
+        Effect.gen(function*() {
+          const reason = yield* Schema.decodeUnknownEffect(Sha256)(`${unshareable.stage}:${unshareable.message}`).pipe(
+            Effect.orDie
+          )
+          yield* emitLifecycle(JournalRecords.cacheProvenance(cacheSource(`unpublished:${reason}`), {
+            keyDigest,
+            action: "unpublished",
+            stage: unshareable.stage,
+            message: unshareable.message
+          }))
+        })
+      /**
+       * Records the sealed completion into the shared cache with provenance,
+       * failing (by default) on a divergent first-recorded row. Used by both
+       * the fresh completion path and succeeded-attempt replay, so a crash
+       * between `attempts.finish` and `cache.put` converges on restart
+       * instead of leaving the cache permanently behind the journal.
+       */
+      const recordCache = (options: {
+        readonly result: unknown
+        readonly admission: CacheAdmission.PublishCompletion<AttemptMeta>
+        readonly createdAtMs: number
+      }) =>
+        Effect.gen(function*() {
+          // BLOBS BEFORE METADATA (issue #172). Every artifact this evidence
+          // references is made durable in the shared tier *before* the entry
+          // that references it is written, so a sibling machine can never see
+          // a hit whose outputs it cannot materialize. Bazel's REAPI ordering
+          // constraint, stated at `UploadManifest.java:630-633`; the whole
+          // protocol lives in `CachePublication`, which is a no-op when no
+          // shared tier is configured. It runs OUTSIDE the write transaction
+          // below, like every other host call.
+          //
+          // A refusal withholds the SHARED entry, never the local row and never
+          // the run: this point is reached after `attempts.finish`, so the work
+          // is already done and durably recorded, and failing here would throw
+          // a real result away because an optional accelerator was unreachable.
+          // It is journalled below instead — the same "visible, not silent"
+          // treatment an unverified read set gets (issue #106).
+          const meta = options.admission.meta
+          let unshareable = yield* CachePublication.publishArtifacts(meta.boundary)
+          // The producer identity folds a digest of the recorded content
+          // (issue #129): a constant per-key identity made a post-eviction
+          // re-record collapse into a `Duplicate` carrying the EVICTED
+          // generation's seq, so the fresh row inherited the evicted row's
+          // exact provenance and a laggard's #119 `ifRecordedBy` fence could
+          // delete the valid new row. With the content folded in, the #124
+          // convergence re-record (identical content, rebuilt from the same
+          // persisted attempt row, so the serialization is byte-stable)
+          // still collapses into a `Duplicate` whose receipt carries the
+          // original emission's canonical seq, while a re-record with
+          // different content is a distinct producer that journals fresh
+          // provenance. A re-record whose content happens to equal the
+          // evicted generation's shares its provenance by design: the two
+          // rows are indistinguishable, so a fenced evict of one is exactly
+          // a fenced evict of the other.
+          //
+          // The digest goes through `DerivedKey`, the repo's one hashing
+          // chokepoint: RFC 8785 canonical JSON, then SHA-256. Hashing
+          // `JSON.stringify` output made "byte stable" depend on key order,
+          // and the two paths do not build `meta` the same way — the fresh
+          // path spreads an object, the convergence path decodes through
+          // `Schema.decodeUnknownEffect(AttemptMeta)`, which emits keys in
+          // schema declaration order. Today the two orders coincide, so the
+          // break was latent; adding one optional field to `AttemptMeta` or
+          // `BoundaryEvidence` above an existing one, or reordering the spread
+          // at line 1405, would have made the convergence re-record compute a
+          // different `generation`, append fresh provenance on every later
+          // dispatch, and reopen the unbounded-append regression issue #124
+          // closed. Canonical JSON makes the stability structural.
+          const generation = yield* Schema.decodeUnknownEffect(DerivedKey)({
+            kind: "cache-generation",
+            meta,
+            result: options.result
+          }).pipe(Effect.orDie)
+          // The provenance record and the row it describes commit together:
+          // the row carries the record's canonical seq as its provenance, so a
+          // crash between them left either a row pointing at a sequence that
+          // does not exist or a `recorded` entry for a row nobody wrote.
+          // Losing the first-writer race is not that crash — it is a decision,
+          // and it stays journalled: the `recorded` entry says what this run
+          // tried to record and the `cache-conflict` entry below says how it
+          // resolved.
+          const recording = yield* atomically(
+            Effect.gen(function*() {
+              const receipt = yield* emitLifecycle(
+                JournalRecords.cacheProvenance({
+                  runId: deps.runId,
+                  lineageId,
+                  cacheKey: keyDigest,
+                  sourceId: `${deps.sourceId}:cache:${keyDigest}:recorded:${generation}`,
+                  sourceSeq: 0
+                }, { keyDigest, action: "recorded" })
+              )
+              const entry = {
+                keyDigest,
+                result: options.result,
+                meta,
+                createdAtMs: options.createdAtMs,
+                recordedRunId: deps.runId,
+                recordedEventSeq: receipt.seq
+              }
+              const outcome = yield* cache.put(entry)
+              // Only an inserting put creates a head, so the insertion is
+              // journalled in the same transaction. `StepCacheFold` rebuilds the
+              // heads from these entries and never from `recorded`, which a
+              // duplicate or a losing conflict also writes (issue #2053).
+              if (outcome._tag === "Inserted") {
+                const recorded = { runId: deps.runId, eventSeq: receipt.seq }
+                yield* emitLifecycle(JournalRecords.cacheProvenance(cacheSource("admitted", recorded), {
+                  keyDigest,
+                  action: "admitted",
+                  recordedRunId: recorded.runId,
+                  recordedEventSeq: recorded.eventSeq
+                }))
+              }
+              return { entry, outcome }
+            })
+          ).pipe(Effect.catchTag("@smthrs/step-cache/CacheStoreError", (error) =>
+            // Only availability of the local transaction is optional. A known
+            // deterministic Conflict below still belongs to Inconsistency.
+            noteUnshareable({ stage: "entry", message: `local cache publication failed: ${error.message}` }).pipe(
+              Effect.as(undefined)
+            )))
+          if (recording === undefined) {
+            return
+          }
+          // ENTRY LAST, AND OUTSIDE THE TRANSACTION. The local row and its
+          // provenance record are now durable together; only here does the
+          // entry become observable to other machines, which is the second half
+          // of the REAPI ordering constraint. It is deliberately not inside the
+          // transaction above: `CacheSync` speaks HTTP, and nothing that is not
+          // storage work may be held across a `DurableWriter` write — a stalled
+          // shared cache would block every other writer and roll back a row
+          // that has nothing to do with it. A `Conflict` skips publication for
+          // the reason the local tier reported it: this machine does not agree
+          // with itself about the key yet.
+          if (unshareable === undefined && recording.outcome._tag !== "Conflict") {
+            unshareable = yield* CachePublication.publishEntry(recording.entry)
+          }
+          if (unshareable !== undefined) {
+            // The refusal is journalled so a missing shared entry is
+            // explainable, never inferred from its absence. The identity folds
+            // a digest of the reason for the same reason the `recorded` record
+            // above folds one of its content (issue #129): a convergence
+            // re-record hitting the identical refusal is an exact producer
+            // retry the journal collapses into a `Duplicate`, while a
+            // *different* refusal is a genuinely new observation that journals
+            // fresh — and neither can ever be the same identity carrying
+            // different content, which is what an idempotency conflict is.
+            yield* noteUnshareable(unshareable)
+          }
+          if (recording.outcome._tag === "Conflict") {
+            const conflicting = yield* cache.get(keyDigest)
+            if (meta.nondeterministic === true) {
+              const recorded = Option.map(conflicting, (entry) => ({
+                runId: entry.recordedRunId,
+                eventSeq: entry.recordedEventSeq
+              }))
+              // Declared output nondeterminism is not a hermeticity violation,
+              // so it does not enter the Inconsistency receiver. The key folds
+              // the declaration, making first-writer-wins safe for both sides.
+              yield* emitLifecycle(
+                JournalRecords.cacheProvenance(
+                  cacheSource("conflict_first_writer", Option.getOrUndefined(recorded)),
+                  {
+                    keyDigest,
+                    action: "conflict_first_writer",
+                    recordedRunId: Option.getOrNull(Option.map(recorded, (value) => value.runId)),
+                    recordedEventSeq: Option.getOrNull(Option.map(recorded, (value) => value.eventSeq))
+                  }
+                )
+              )
+              return
+            }
+            const receiverOption = yield* Effect.serviceOption(Inconsistency.Inconsistency)
+            // Core default is STRICT: journal the conflict and fail the run,
+            // which is Skyframe's throwing `GraphInconsistencyReceiver`.
+            // Providing `Inconsistency.layerTolerant` opts out.
+            const receiver = Option.isSome(receiverOption)
+              ? receiverOption.value
+              : Inconsistency.make({ journal, verdict: "fail", owner: deps.owner })
+            const verdict = yield* receiver.note({
+              key: keyDigest,
+              existing: Option.getOrUndefined(conflicting),
+              attempted: recording.entry
+            })
+            if (verdict === "fail") {
+              return yield* Effect.fail(
+                new CacheConflictDetected({
+                  code: "cache_conflict_detected",
+                  keyDigest,
+                  recordedRunId: Option.isSome(conflicting) ? conflicting.value.recordedRunId : "unknown"
+                })
+              )
+            }
+          }
+        })
+      // Everything from the cache-hit verification to the terminal transition
+      // runs under this process's exclusive permit for the attempt key
+      // (issues #102, #103, #118): the adoption decision below is taken from
+      // a read no in-process racer can invalidate before the claim lands, a
+      // concurrent same-key dispatch waits here until the winner's terminal
+      // row is visible and replays it instead of re-executing the body, and
+      // the cache block's read-verify-materialize-evict span can never
+      // interleave with another dispatch's execution. A verified hit returns
+      // from inside the permit. The permit is keyed by (runId, stepKeyDigest)
+      // alone (issue #133): the cache row is addressed with no attempt
+      // material, so folding the attempt counter in let sanctioned keyed
+      // dispatches at skewed attempt counters (#111/#116) acquire different
+      // permits and interleave the very span the permit serializes. The step
+      // key, not the narrowed cache address, is what serializes here, because
+      // the attempt rows two scopes of one key share are what they race over.
+      return yield* admission.withPermit(`${deps.runId}|${stepKeyDigest}`)(
+        Effect.gen(function*() {
+          // Lifecycle announcements take a per-attempt producer identity
+          // (issue #91): adoption — or a replay after a crash in the
+          // finish→emit window (issue #109) — re-emits records a dead
+          // incarnation may already have announced, and lifecycle records
+          // without a `sourceSeq` allocate a fresh journal row on every
+          // emission. A dedicated `(sourceId, sourceSeq 0)` per record makes
+          // the re-emission an exact producer retry the journal collapses
+          // into a `Duplicate`.
+          const attemptSource = (record: string): JournalRecords.EventOptions => ({
+            runId: deps.runId,
+            lineageId,
+            // A sealed dispatch's result lives in the step cache, so the record
+            // carries the digest that addresses it: replay hands the projection
+            // the sealed value instead of re-deriving it, and a cache miss is
+            // simply an absent value rather than a broken fold.
+            ...(input.tier === "sealed" ? { cacheKey: keyDigest } : {}),
+            sourceId: `${deps.sourceId}:attempt:${stepKeyDigest}:${input.attempt}:${record}`,
+            sourceSeq: 0
+          })
+          /**
+           * Journal-convergence emit for the replay branches (issue #109):
+           * an identical re-emission collapses into a `Duplicate`, and an
+           * `idempotency_conflict` may mean the journal already holds this
+           * record under this producer identity as another lineage recorded
+           * it — a time-travel fork copies the parent's journal rows, so the
+           * copied record names the parent run. A cache-provenance record
+           * may also restate a fact this run already journalled with fresh
+           * measurements (an expiry re-measured after a resume). Those are
+           * the only conflicts this emit tolerates, and only after reading
+           * the occupying record and validating it: an attempt-scoped record
+           * must be this record as a retained fork ancestor emitted it — same
+           * event type, attempt coordinates and terminal state
+           * (`CopiedRecord.accept`); a provenance record must state the same
+           * fact about the same row (`ProvenanceSlot.accept`). An unrelated
+           * or contradictory record in the slot — corrupt or imported
+           * history, a producer collision — surfaces the journal's own
+           * conflict instead of a replay that reports no inconsistency.
+           */
+          const emitConverging = (record: JournalEvent.Input) =>
+            emitLifecycle(record).pipe(
+              Effect.catch((error) =>
+                error.code !== "idempotency_conflict"
+                  ? Effect.fail(error)
+                  : record.eventType === "flows.engine.cache-provenance"
+                  ? ProvenanceSlot.accept({ journal, runId: deps.runId, record, conflict: error })
+                  : CopiedRecord.accept({ journal, runs, runId: deps.runId, record, conflict: error })
+              )
+            )
+          /**
+           * Records and clears a row the declared time to live has aged out.
+           *
+           * The age judgement is journalled BEFORE the row is dropped, and
+           * with the age and the bound in the payload, so a reader can tell an
+           * expiry from a plain miss and a replay reads the recorded verdict
+           * instead of re-judging the age against a fresh clock. The eviction
+           * is what keeps `CacheStore.put`'s insert-or-nothing head from
+           * turning the re-execution's genuinely newer result into a
+           * `Conflict` — the same repair the stale-read-set branch performs
+           * (issue #99) — and it is fenced on the expired row's own provenance
+           * (issue #119) so a fresh row another process landed in the meantime
+           * survives.
+           */
+          const expireRow = (ttlMs: number, stale: CacheStore.CacheEntry) =>
+            Effect.gen(function*() {
+              const recorded = {
+                runId: stale.recordedRunId,
+                eventSeq: stale.recordedEventSeq
+              }
+              const nowMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+              yield* emitConverging(
+                JournalRecords.cacheProvenance(cacheSource("expired", recorded), {
+                  keyDigest,
+                  action: "expired",
+                  ttlMs,
+                  ageMs: nowMs - stale.createdAtMs,
+                  recordedRunId: recorded.runId,
+                  recordedEventSeq: recorded.eventSeq
+                })
+              )
+              yield* cache.evict(keyDigest, { ifRecordedBy: recorded })
+            })
+          /**
+           * Decides — once, durably — whether a row is inside the declared time
+           * to live, and answers the DECISION THIS RUN RECORDED rather than the
+           * one this clock reading would make.
+           *
+           * A time to live is the one cache input that changes answer on its
+           * own. A dispatch that served a row at 900 ms and then lost its
+           * process would, on the plain reading, re-dispatch at 1100 ms, expire
+           * the row it already served, and execute a body whose result the run
+           * had already consumed. The verdict is therefore journalled before it
+           * is acted on, under a producer identity that names the step key, the
+           * `ttl` decision, and the row's own recorded provenance
+           * (`cacheSource`), with `sourceSeq` 0. That identity belongs to the
+           * RUN, not to the incarnation driving it, so the engine that resumes
+           * the run reads the same verdict rather than taking a second one.
+           *
+           * `Accepted` records the first decision; `Duplicate` confirms it.
+           * A conflict alone does not identify the recorded verdict: TTL is
+           * not part of every action key, and copied history can carry other
+           * lineage metadata. Only an exact duplicate of the opposite verdict
+           * proves a clock change. Otherwise fail before serving or expiring
+           * the row, preserving the incompatible history for the caller.
+           *
+           * A row re-recorded under a provenance this run already expired keeps
+           * that verdict, which is the same rule read from the other side.
+           */
+          const admitByAge = (ttlMs: number, row: CacheStore.CacheEntry) =>
+            Effect.gen(function*() {
+              const recorded = { runId: row.recordedRunId, eventSeq: row.recordedEventSeq }
+              const nowMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+              const measured: "admitted" | "expired" = nowMs - row.createdAtMs <= ttlMs ? "admitted" : "expired"
+              const decision = (verdict: "admitted" | "expired") =>
+                JournalRecords.cacheProvenance(cacheSource("ttl", recorded), {
+                  keyDigest,
+                  action: "ttl",
+                  ttlMs,
+                  verdict,
+                  recordedRunId: recorded.runId,
+                  recordedEventSeq: recorded.eventSeq
+                })
+              // A copied producer whose source fields changed must not look
+              // like a new first decision merely because emit would accept it.
+              const expectedSource = cacheSource("ttl", recorded).sourceId
+              const prior = yield* CacheAgeHistory.find(journal, deps.runId, (entry) => {
+                const payload = entry.payload as {
+                  action?: unknown
+                  keyDigest?: unknown
+                  recordedRunId?: unknown
+                  recordedEventSeq?: unknown
+                } | null
+                return entry.sourceId === expectedSource ||
+                  (payload?.action === "ttl" && payload.keyDigest === keyDigest &&
+                    payload.recordedRunId === recorded.runId && payload.recordedEventSeq === recorded.eventSeq)
+              })
+              if (prior !== undefined && (prior.sourceId !== expectedSource || prior.sourceSeq !== 0)) {
+                yield* emitConverging(JournalRecords.cacheProvenance(cacheSource("replay_failed", recorded), {
+                  keyDigest,
+                  action: "replay_failed",
+                  reason: "incompatible-age-history",
+                  recordedRunId: recorded.runId,
+                  recordedEventSeq: recorded.eventSeq
+                }))
+                return yield* Effect.fail(
+                  new Journal.JournalError({
+                    code: "idempotency_conflict",
+                    message: "incompatible recorded cache-age decision: producer identity changed",
+                    cause: prior
+                  })
+                )
+              }
+              return yield* emitLifecycle(decision(measured)).pipe(
+                Effect.as(measured),
+                Effect.catch((error) => {
+                  if (error.code !== "idempotency_conflict") {
+                    return Effect.fail(error)
+                  }
+                  const conflict = new Journal.JournalError({
+                    code: "idempotency_conflict",
+                    message:
+                      "incompatible recorded cache-age decision: restore the original TTL and history identity " +
+                      "or use a new action identity; the cache row was not served or expired",
+                    cause: error
+                  })
+                  const opposite = measured === "admitted" ? "expired" : "admitted"
+                  return emitLifecycle(decision(opposite)).pipe(
+                    Effect.catch((oppositeError) =>
+                      oppositeError.code === "idempotency_conflict"
+                        ? CacheAgeHistory.copiedVerdict({
+                          journal,
+                          runs,
+                          runId: deps.runId,
+                          decision: decision(measured),
+                          conflict
+                        }).pipe(
+                          Effect.tapError((failure) =>
+                            failure.code === "idempotency_conflict"
+                              ? emitConverging(JournalRecords.cacheProvenance(cacheSource("replay_failed", recorded), {
+                                keyDigest,
+                                action: "replay_failed",
+                                reason: "incompatible-age-history",
+                                recordedRunId: recorded.runId,
+                                recordedEventSeq: recorded.eventSeq
+                              })).pipe(Effect.asVoid)
+                              : Effect.void
+                          ),
+                          Effect.map((verdict) => ({ _tag: "Historical" as const, verdict }))
+                        )
+                        : Effect.fail(oppositeError)
+                    ),
+                    Effect.flatMap((receipt) =>
+                      receipt._tag === "Historical" ?
+                        Effect.succeed(receipt.verdict) :
+                        receipt._tag === "Duplicate"
+                        ? Effect.succeed(opposite)
+                        : Effect.fail(conflict)
+                    )
+                  )
+                })
+              )
+            })
+          /**
+           * The row this dispatch may serve, after the declared age bound.
+           *
+           * The bound is applied here rather than inside `cache.get` so the
+           * refusal is a decision this run journalled and owns, not a read
+           * policy the store re-derives from a fresh clock on every lookup.
+           * The no-TTL guard refreshes a per-executor verdict index through
+           * the journal's current tail before answering absence. Its cursor
+           * covers only fully processed pages in the same journal generation;
+           * resume or a changed history rebuilds it once, never per dispatch.
+           */
+          const admissible = (
+            ttlMs: number | undefined,
+            row: Option.Option<CacheStore.CacheEntry>
+          ): Effect.Effect<Option.Option<CacheStore.CacheEntry>, Journal.JournalError | CacheStore.CacheStoreError> =>
+            Effect.gen(function*() {
+              // Removing TTL cannot bypass a verdict this run already consumed,
+              // even if the head was subsequently evicted or replaced.
+              if (ttlMs === undefined) {
+                const recorded = yield* recordedAgeVerdict(journal, keyDigest)
+                if (recorded !== undefined) {
+                  return yield* Effect.fail(
+                    new Journal.JournalError({
+                      code: "idempotency_conflict",
+                      message:
+                        "incompatible recorded cache-age decision: ttlMs cannot be removed after a recorded verdict",
+                      cause: recorded
+                    })
+                  )
+                }
+                return row
+              }
+              if (Option.isNone(row)) {
+                return row
+              }
+              if ((yield* admitByAge(ttlMs, row.value)) === "admitted") {
+                return row
+              }
+              yield* expireRow(ttlMs, row.value)
+              return Option.none<CacheStore.CacheEntry>()
+            })
+          // A run's admitted attempt is replay authority. Shared cache rows
+          // are replaceable and cannot overwrite its result or failure.
+          const existing = yield* attempts.get(attemptId)
+          if (cacheDeclaration._tag === "Eligible") {
+            const observed = yield* cache.get(keyDigest).pipe(
+              Effect.catch((error) =>
+                noteUnshareable({ stage: "entry", message: `cache lookup failed: ${error.message}` }).pipe(
+                  Effect.as(Option.none<CacheStore.CacheEntry>())
+                )
+              )
+            )
+            const evidenceDecision = CacheAdmission.candidate(
+              Option.isSome(observed) ? decodeMeta(observed.value.meta) : undefined
+            )
+            const outputDecision = evidenceDecision._tag === "CandidateEvidence"
+              ? CacheOutputPolicy.replay(input.metadata, evidenceDecision.evidence)
+              : undefined
+            const candidate = outputDecision?._tag === "Refused" ? outputDecision : evidenceDecision
+            // A descriptor refusal cannot be turned into TTL eviction. Validate
+            // replay authority before any age decision is allowed to prune a head.
+            const cached =
+              Option.isNone(observed) || candidate._tag === "CandidateEvidence" || policy?.ttlMs === undefined
+                ? yield* admissible(policy?.ttlMs, observed)
+                : observed
+            if (Option.isSome(cached)) {
+              if (candidate._tag === "Refused") {
+                yield* emitConverging(JournalRecords.cacheProvenance(
+                  cacheSource("replay_failed", {
+                    runId: cached.value.recordedRunId,
+                    eventSeq: cached.value.recordedEventSeq
+                  }),
+                  {
+                    keyDigest,
+                    action: "replay_failed",
+                    reason: candidate.reason,
+                    recordedRunId: cached.value.recordedRunId,
+                    recordedEventSeq: cached.value.recordedEventSeq
+                  }
+                ))
+              }
+              if (candidate._tag === "CandidateEvidence" && Option.isNone(existing)) {
+                const boundary = yield* StepBoundary.StepBoundary
+                // Skyframe's dirty check, not "the declaration changed" (issue
+                // #90): the read-set digests folded into the step key are caller
+                // metadata, so reuse is justified only once the host has
+                // measured them and agreed. A stale declaration falls through to
+                // a real execution instead of replaying a pre-edit result, and
+                // the refusal is journalled so it is visible rather than silent.
+                // A boundary the host cannot enforce is likewise not a hit; the
+                // dispatch path below re-prepares and fails the attempt properly.
+                const measured = yield* boundary.prepare(cacheDeclaration.metadata).pipe(Effect.option)
+                const verified = Option.isSome(measured) && StepBoundary.readSetMatches(measured.value)
+                if (verified) {
+                  const evidence = candidate.evidence
+                  // Cache materialization is host work too. The authorized
+                  // invocation is durable before it, even without a new attempt.
+                  yield* emitCallInvoked
+                  let materialized = yield* boundary.replayOutputs(evidence).pipe(Effect.exit)
+                  if (
+                    Exit.isFailure(materialized) &&
+                    CachePublication.replayMissingArtifact(materialized.cause) !== undefined
+                  ) {
+                    // LAZY DOWNLOAD (issue #172). A shared cache row is
+                    // routinely recorded on a machine whose artifacts this one
+                    // has never seen, so "the blob is not here" is the normal
+                    // first answer, not a defect — and it is the one replay
+                    // refusal a shared artifact tier can repair. Fetch, verify,
+                    // write back, and retry the replay ONCE. Only once: a
+                    // second failure means the tier cannot serve it either, and
+                    // the fall-through below (a real execution) is strictly
+                    // better than looping. With no shared tier configured
+                    // `hydrateArtifacts` reports `false` and this costs one
+                    // cheap branch.
+                    if (yield* CachePublication.hydrateArtifacts(evidence)) {
+                      materialized = yield* boundary.replayOutputs(evidence).pipe(Effect.exit)
+                    }
+                  }
+                  if (Exit.isSuccess(materialized)) {
+                    const recorded = {
+                      runId: cached.value.recordedRunId,
+                      eventSeq: cached.value.recordedEventSeq
+                    }
+                    yield* emitConverging(JournalRecords.cacheProvenance(cacheSource("hit", recorded), {
+                      keyDigest,
+                      recordedRunId: cached.value.recordedRunId,
+                      recordedEventSeq: cached.value.recordedEventSeq
+                    }))
+                    // Counted after the provenance emit: `verified_hit` means
+                    // the cached result was served, and a journal failure on
+                    // the emit fails the dispatch before any result is
+                    // returned. A dispatch that dies mid-decision records no
+                    // decision; its exit lands in `flows_engine_dispatches`.
+                    yield* Metric.update(EngineStoreMetrics.stepCacheDecision.VerifiedHit, 1)
+                    // Served from a record: the body never ran, and the node
+                    // above this dispatch is told so rather than left to
+                    // report work nobody did.
+                    yield* Action.reportDispatch({
+                      outcome: "replayed",
+                      stepKeyDigest,
+                      attempt: input.attempt
+                    })
+                    return cached.value.result
+                  }
+                  // Evidence the host cannot re-materialize — a transient
+                  // filesystem error, or a row recorded by a foreign boundary
+                  // implementation — is not a hit and not a run failure
+                  // (issue #107): failing here while the verified row survived
+                  // repeated refuse→fail on every later run, the exact
+                  // permanent-failure loop #99 closed one branch later. The
+                  // refusal is journalled and the dispatch path below executes
+                  // for real; the row survives for hosts that can replay it.
+                  // The refusal record carries its classification (issue
+                  // #150): `corruption` for a digest mismatch at the
+                  // content-addressed blob path, `host` for everything else —
+                  // a failing disk corrupting many blobs must never journal
+                  // identically to a one-off EIO.
+                  const corruption = replayCorruption(materialized.cause)
+                  yield* emitConverging(
+                    JournalRecords.cacheProvenance(
+                      cacheSource("replay_failed", {
+                        runId: cached.value.recordedRunId,
+                        eventSeq: cached.value.recordedEventSeq
+                      }),
+                      {
+                        keyDigest,
+                        action: "replay_failed",
+                        reason: corruption === undefined ? "host" : "corruption",
+                        recordedRunId: cached.value.recordedRunId,
+                        recordedEventSeq: cached.value.recordedEventSeq
+                      }
+                    )
+                  )
+                  if (corruption !== undefined) {
+                    // Corruption is an integrity violation, not a retryable
+                    // refusal: it routes to the Inconsistency receiver like
+                    // a cache-key conflict does. The core default is STRICT
+                    // (fail); `Inconsistency.layerTolerant` (or a plugin
+                    // verdict) lets the dispatch fall back to the real
+                    // execution below, whose re-capture heals the address.
+                    const receiverOption = yield* Effect.serviceOption(Inconsistency.Inconsistency)
+                    const receiver = Option.isSome(receiverOption)
+                      ? receiverOption.value
+                      : Inconsistency.make({ journal, verdict: "fail", owner: deps.owner })
+                    const verdict = yield* receiver.noteCorruption({
+                      runId: deps.runId,
+                      keyDigest,
+                      path: corruption.path,
+                      recordedDigest: corruption.recordedDigest,
+                      measuredDigest: corruption.measuredDigest,
+                      recordedRunId: cached.value.recordedRunId,
+                      recordedEventSeq: cached.value.recordedEventSeq
+                    })
+                    // Quarantine is journal AND evict (issue #164). The
+                    // receiver's durable record preserved the evidence, but
+                    // leaving the row in place made the poison permanent for
+                    // INLINE evidence: `CacheStore.put` is insert-or-nothing,
+                    // so a tolerant re-execution never replaced the corrupt
+                    // bytes and re-detected them on every later run, while
+                    // strict mode re-failed the key forever. Evicting under
+                    // both verdicts lets the next dispatch — the tolerant
+                    // fall-through below, or the run after a strict failure —
+                    // execute and record cleanly. The evict is fenced on the
+                    // poisoned row's own provenance like the stale-read-set
+                    // branch (issue #119): a fresh row landed by a concurrent
+                    // run between this dispatch's `get` and the `evict` makes
+                    // the compare-and-swap a no-op instead of deleting valid
+                    // evidence.
+                    yield* cache.evict(keyDigest, {
+                      ifRecordedBy: {
+                        runId: cached.value.recordedRunId,
+                        eventSeq: cached.value.recordedEventSeq
+                      }
+                    })
+                    if (verdict === "fail") {
+                      return yield* Effect.fail(
+                        new CacheCorruptionDetected({
+                          code: "cache_corruption_detected",
+                          keyDigest,
+                          path: corruption.path,
+                          recordedDigest: corruption.recordedDigest,
+                          measuredDigest: corruption.measuredDigest
+                        })
+                      )
+                    }
+                  }
+                  // `replay_failed` is a fall-through decision, so it is
+                  // counted only once the refusal emit landed and the strict
+                  // corruption verdict above has NOT terminated the dispatch:
+                  // a dispatch that fails instead of re-executing records no
+                  // decision.
+                  yield* Metric.update(EngineStoreMetrics.stepCacheDecision.ReplayFailed, 1)
+                } else if (Option.isSome(measured)) {
+                  // Only a *measured* mismatch is evidence the inputs changed
+                  // (issue #110): a host that cannot measure right now — a
+                  // transient EIO/EACCES on any declared read path — says
+                  // nothing about the read set, so the hit is merely refused for
+                  // this dispatch (the path below re-prepares and surfaces the
+                  // host failure as an ordinary attempt failure) and the valid
+                  // shared row survives for every run whose host is healthy.
+                  //
+                  // The durable emit is also the fence: `emitDurable` fails with
+                  // `fence_lost` for a zombie that lost the run, surfacing as
+                  // self-interruption before the eviction below can run.
+                  yield* emitConverging(
+                    JournalRecords.cacheProvenance(
+                      cacheSource("stale_read_set", {
+                        runId: cached.value.recordedRunId,
+                        eventSeq: cached.value.recordedEventSeq
+                      }),
+                      {
+                        keyDigest,
+                        action: "stale_read_set",
+                        recordedRunId: cached.value.recordedRunId,
+                        recordedEventSeq: cached.value.recordedEventSeq
+                      }
+                    )
+                  )
+                  // Skyframe invalidation, not just refusal (issue #99): a stale
+                  // read set means the inputs changed, so the re-execution's
+                  // result is *expected* to differ — left in place, the poisoned
+                  // row turns the fresh `cache.put` into a Conflict, the strict
+                  // verdict fails the run, and nothing ever removes the row, so
+                  // every later run repeats the refuse → re-execute → conflict →
+                  // fail cycle. The refusal is journalled above; evicting here
+                  // lets the re-execution record cleanly under the same key.
+                  // The eviction is fenced on the poisoned row's own
+                  // provenance (issue #119): a fresh entry recorded by a
+                  // concurrent run between this dispatch's `get` and its
+                  // `evict` must not be deleted with the poison (issue #110).
+                  // The permit (issue #118) only closes the *in-process*
+                  // window, so the guard rides inside the DELETE rather than
+                  // in a preceding read — a foreign process landing a fresh
+                  // row simply makes the compare-and-swap a no-op.
+                  yield* cache.evict(keyDigest, {
+                    ifRecordedBy: {
+                      runId: cached.value.recordedRunId,
+                      eventSeq: cached.value.recordedEventSeq
+                    }
+                  })
+                  // The fall-through decision is counted once the refusal is
+                  // durable and the poisoned row is gone: a fenced-out zombie
+                  // self-interrupts at the emit above and records no decision.
+                  yield* Metric.update(EngineStoreMetrics.stepCacheDecision.StaleReadSet, 1)
+                } else {
+                  // The host could not measure the read set at all, so the
+                  // hit is merely refused for this dispatch; the row survives.
+                  yield* Metric.update(EngineStoreMetrics.stepCacheDecision.Unmeasurable, 1)
+                }
+              } else if (candidate._tag === "Refused") {
+                // A row whose recorded evidence cannot justify reuse — a
+                // foreign tier, an unverified capture, a recorded deviation.
+                yield* Metric.update(EngineStoreMetrics.stepCacheDecision.UnverifiableEvidence, 1)
+              }
+            } else {
+              yield* Metric.update(EngineStoreMetrics.stepCacheDecision.Miss, 1)
+            }
+          }
+
+          if (Option.isSome(existing)) {
+            const row = existing.value
+            if (row.state === "succeeded") {
+              yield* emitCallInvoked
+              const meta = decodeMeta(row.meta)
+              let corruptEvidence = false
+              const outputDecision = meta?.boundary === undefined
+                ? undefined
+                : CacheOutputPolicy.replay(input.metadata, meta.boundary)
+              const storedDecision = CacheAdmission.candidate(meta)
+              const refused = storedDecision._tag === "Refused" &&
+                  (storedDecision.reason === "contradictory-evidence" ||
+                    storedDecision.reason === "quarantined-evidence")
+                ? storedDecision
+                : outputDecision
+              if (refused?._tag === "Refused") {
+                yield* emitConverging(JournalRecords.cacheProvenance(cacheSource("replay_failed"), {
+                  keyDigest,
+                  action: "replay_failed",
+                  reason: refused.reason
+                }))
+              }
+              if (meta?.boundary !== undefined && refused?._tag !== "Refused") {
+                const boundary = yield* StepBoundary.StepBoundary
+                const materialized = yield* boundary.replayOutputs(meta.boundary).pipe(Effect.exit)
+                if (Exit.isFailure(materialized)) {
+                  // The attempt durably succeeded: its recorded outcome is
+                  // the truth, and re-materializing the workspace outputs is
+                  // best-effort — failing the dispatch here while the
+                  // terminal row survived recreated the #99 permanent
+                  // refuse→fail loop one branch earlier (issue #107). The
+                  // refusal is journalled so a missing output is
+                  // explainable rather than silent, and it carries its
+                  // classification (issue #150): corruption of the recorded
+                  // evidence routes to the Inconsistency receiver instead of
+                  // passing as an ordinary transient host refusal.
+                  const corruption = replayCorruption(materialized.cause)
+                  corruptEvidence = corruption !== undefined
+                  yield* emitConverging(
+                    JournalRecords.cacheProvenance(cacheSource("replay_failed"), {
+                      keyDigest,
+                      action: "replay_failed",
+                      reason: corruption === undefined ? "host" : "corruption"
+                    })
+                  )
+                  if (corruption !== undefined) {
+                    const receiverOption = yield* Effect.serviceOption(Inconsistency.Inconsistency)
+                    const receiver = Option.isSome(receiverOption)
+                      ? receiverOption.value
+                      : Inconsistency.make({ journal, verdict: "fail", owner: deps.owner })
+                    const verdict = yield* receiver.noteCorruption({
+                      runId: deps.runId,
+                      // The attempt row, not a cache row: this evidence is
+                      // addressed by (runId, stepKeyDigest, attempt).
+                      keyDigest: stepKeyDigest,
+                      path: corruption.path,
+                      recordedDigest: corruption.recordedDigest,
+                      measuredDigest: corruption.measuredDigest
+                    })
+                    // Quarantine is journal AND take a state action (issues
+                    // #164, #171). Unlike a shared cache row, this succeeded
+                    // attempt cannot be evicted: its side effects already ran,
+                    // so a miss would re-execute potentially irreversible work.
+                    // Remove only the corrupt boundary evidence and mark the
+                    // row quarantined. The durable outcome stays intact and the
+                    // next dispatch returns it without materializing the poison
+                    // or publishing the row back into the shared cache.
+                    //
+                    // The owner heartbeat and patch share one write
+                    // transaction. The patch is owner-fenced itself — it only
+                    // lands while `flows_runs` still records `deps.owner` —
+                    // and the heartbeat both refreshes the lease and reports
+                    // the loss as a run-store outcome before the patch runs.
+                    // `boundary` is dropped by omitting the key, never by
+                    // writing `undefined` over it: the attempt store admits
+                    // inert JSON, and an own property holding `undefined` is
+                    // not JSON, so the patch would be refused outright.
+                    // New quarantine writes clear both reusable proof fields;
+                    // contradictory legacy rows remain readable without repair.
+                    const { boundary: _quarantined, readSetVerified: _unverified, ...retainedMeta } = meta
+                    const quarantinedMeta: AttemptMeta = {
+                      ...retainedMeta,
+                      boundaryQuarantined: true
+                    }
+                    const quarantined = yield* atomically(Effect.gen(function*() {
+                      const quarantineAtMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+                      const fence = yield* runs.heartbeat(deps.runId, deps.owner, quarantineAtMs)
+                      if (fence._tag !== "Updated") {
+                        return false
+                      }
+                      const patched = yield* attempts.patch(
+                        attemptId,
+                        { meta: attemptPayload(quarantinedMeta) },
+                        deps.owner
+                      )
+                      return patched._tag === "Patched"
+                    }))
+                    if (!quarantined) {
+                      return yield* Effect.interrupt
+                    }
+                    if (verdict === "fail") {
+                      // The strict verdict still makes the integrity violation
+                      // visible by parking this dispatch. The row repair above
+                      // is what makes the park resumable without an out-of-band
+                      // byte repair. This is a defect, not a declared action
+                      // business error: routing it through the failure channel
+                      // would make the action's error schema replace it with
+                      // a SchemaError.
+                      return yield* Effect.die(
+                        new AttemptEvidenceQuarantined({
+                          code: "attempt_evidence_quarantined",
+                          keyDigest: stepKeyDigest,
+                          attempt: input.attempt,
+                          path: corruption.path,
+                          recordedDigest: corruption.recordedDigest,
+                          measuredDigest: corruption.measuredDigest
+                        })
+                      )
+                    }
+                  }
+                }
+              }
+              // Converge the cache with the durable completion: a crash between
+              // `attempts.finish` and `cache.put` otherwise leaves the sealed
+              // result permanently missing from the shared cache (issue #24).
+              // Reaching this branch with an eligible declaration means `cache.get`
+              // above missed or was unfit for replay.
+              // Only a row whose recorded read set was verified at prepare
+              // time may converge into the shared cache (issue #106): an
+              // unverified result was computed against content the key does
+              // not describe.
+              // A row whose boundary evidence was just measured corrupt is
+              // quarantined, never converged (issue #160): a `tolerate`
+              // verdict keeps the durable outcome as this run's truth, but
+              // publishing the known-corrupt evidence into the shared cache
+              // would hand sibling runs a poisoned hit under this run's
+              // provenance.
+              const admission = CacheAdmission.completion(
+                cacheDeclaration,
+                meta,
+                corruptEvidence || refused?._tag === "Refused"
+              )
+              if (admission._tag === "PublishCompletion") {
+                yield* recordCache({
+                  result: row.outcome,
+                  admission,
+                  createdAtMs: row.finishedAtMs ?? (yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor)))
+                })
+              }
+              // Converge the journal with the durable completion (issue
+              // #109): a crash between `attempts.finish` and the terminal
+              // emits left `attemptStarted` without `attemptFinished`
+              // forever. The per-attempt producer identity collapses the
+              // re-emission into a `Duplicate` on ordinary replays.
+              if (meta?.boundary?.deviation !== undefined) {
+                yield* emitConverging(
+                  JournalRecords.expectedSetDeviation(attemptSource("deviation"), {
+                    ...attemptId,
+                    ...meta.boundary.deviation
+                  })
+                )
+              }
+              // Replays publish the saved first observation, never a newly
+              // computed timestamp or the current history generation.
+              yield* emitStepSettled(row.outcome)
+              yield* emitConverging(
+                JournalRecords.attemptFinished(attemptSource("finished"), { ...attemptId, state: "succeeded" })
+              )
+              // Upgrade an older durable controller record without rerunning
+              // its handler; current records collapse to the original receipt.
+              yield* emitCallSettled(row.outcome)
+              // A durable attempt row answered the dispatch, so nothing ran.
+              yield* Action.reportDispatch({ outcome: "replayed", stepKeyDigest, attempt: input.attempt })
+              return row.outcome
+            }
+            if (row.state === "failed") {
+              yield* emitCallInvoked
+              // Converge the journal before rethrowing (issue #109): the
+              // violation kind survives in the row meta because the
+              // persisted cause alone cannot distinguish a boundary
+              // violation from an ordinary execution failure.
+              if (decodeMeta(row.meta)?.hardViolation === true) {
+                yield* emitConverging(
+                  JournalRecords.hardViolation(attemptSource("hard-violation"), {
+                    ...attemptId,
+                    error: rehydrateCause(row.error)
+                  })
+                )
+              }
+              yield* emitConverging(
+                JournalRecords.attemptFinished(attemptSource("finished"), { ...attemptId, state: "failed" })
+              )
+              // A durably failed attempt is replayed by rethrowing the persisted
+              // domain failure — never by readmission (issue #59). Falling
+              // through to `attempts.put` here surfaced the row as
+              // `AttemptAdmissionRejected`, whose tag can never match a
+              // policy-declared `nonRetryable` classification, so a durably
+              // failed no-retry action earned an extra real dispatch after
+              // resume. Temporal's prior art: mutable state persists the attempt
+              // failure and `ExecutionInfo.Attempt`, and its no-retry decision
+              // (`service/history/workflow/retry.go`) is re-evaluated from that
+              // persisted failure — the failure itself is durable, not just the
+              // fact that an attempt happened.
+              return yield* Effect.failCause(rehydrateCause(row.error))
+            }
+            if (row.state === "suspended") {
+              return yield* Effect.fail(
+                new AttemptSuspended({
+                  code: "attempt_suspended",
+                  runId: deps.runId,
+                  keyDigest: stepKeyDigest,
+                  attempt: input.attempt
+                })
+              )
+            }
+          }
+
+          /**
+           * A persisted `running` row read while this owner holds the run fence
+           * is crash evidence, not a live admission (issue #71): the incarnation
+           * that admitted the attempt died before finishing (SIGKILL, OOM), and
+           * the #53 stale-running sweep re-drove the run to a new owner — or the
+           * same owner after restart. The attempt never completed, so it must
+           * re-execute under its original number rather than fall through to
+           * `attempts.put` (whose `Conflict` on the differing `startedAtMs`
+           * surfaced as `AttemptAdmissionRejected`, permanently failing a
+           * no-policy run with an infrastructure tag). The row is adopted; the
+           * ordinary fenced `attempts.finish` transition below records the
+           * re-execution's outcome.
+           *
+           * Adoption requires liveness evidence (issue #86), and that evidence is
+           * the admission permit held around this whole span (issues #102, #103):
+           * a live same-key dispatch of this process would be holding the permit,
+           * so a `running` row observed here cannot belong to a live in-process
+           * fiber — it is a dead fiber of this incarnation (an in-process
+           * re-drive after an interrupt, the #71 mode the recorded-nonce guard
+           * wrongly refused) or a superseded incarnation (which provably lost the
+           * run fence this owner holds). Deciding from the recorded `admittedBy`
+           * nonce could not tell those apart, and comparing it against a stale
+           * pre-claim read left a TOCTOU window where two concurrent dispatches
+           * both saw a dead owner and both executed an irreversible body.
+           */
+          const runningRow = Option.isSome(existing) && existing.value.state === "running"
+            ? existing.value
+            : undefined
+          const runningMeta = runningRow === undefined ? undefined : decodeMeta(runningRow.meta)
+          /**
+           * ADOPTION READS THE CROSSING BACK. A `running` row is crash
+           * evidence, but "the attempt did not settle" is not "the body did
+           * not run": an effect whose terminal boundary is already durable was
+           * still re-dispatched here, because the row alone cannot say so and
+           * nothing ever read the boundary record back. That made a keyless
+           * irreversible action charge twice on a SIGKILL landing in the
+           * millisecond between the `succeeded` boundary and `attempts.finish`
+           * — the exact window the boundary record exists to close.
+           *
+           * A recorded `succeeded` crossing settles the adopted attempt from
+           * the outcome the crossing wrote, in this attempt's own number, and
+           * the body never runs again. The outcome comes from the ROW, not
+           * from the journal record: journal payloads go through redaction on
+           * write, so the record's `output` is observability and the row's
+           * `outcome` is the executable copy (see `AttemptMeta.effectCrossing`).
+           */
+          if (runningRow !== undefined && runningMeta?.effectCrossing === "succeeded") {
+            yield* emitCallInvoked
+            const finishedAtMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+            // `outcome` and `meta` are deliberately omitted: `FinishAttempt`
+            // leaves an omitted field as recorded, so the terminal transition
+            // is the crossing's own result being sealed rather than a value
+            // this incarnation invented.
+            const finished = yield* settleAttempt(
+              { ...attemptId, state: "succeeded", finishedAtMs },
+              [JournalRecords.attemptFinished(attemptSource("finished"), { ...attemptId, state: "succeeded" })]
+            )
+            if (!finished) {
+              return yield* Effect.interrupt
+            }
+            return runningRow.outcome
+          }
+          /**
+           * AN UNRESOLVED CROSSING IS NOT SILENTLY RE-DISPATCHED. `intended`
+           * without a terminal crossing means the body was dispatched and
+           * nothing durable says how it ended: the charge may have gone
+           * through, and no record can settle the question. An idempotency key
+           * is the one thing that makes the repeat safe, so with a key the
+           * attempt re-executes exactly as before, and without one this
+           * refuses under the same error a numbered retry refuses under. The
+           * run stops with a typed, durable "we cannot tell whether this
+           * happened" for an operator, which is the honest outcome; charging a
+           * second time to keep the run moving is not.
+           *
+           * Only tier 3 refuses. A compensable body is re-executed on a
+           * restored pre-image (issue #87), which is what makes its repeat
+           * safe by construction.
+           */
+          if (
+            input.tier === "irreversible" &&
+            runningMeta?.effectCrossing === "intended" &&
+            deps.idempotencyKey === undefined
+          ) {
+            return yield* Effect.fail(
+              new Action.IrreversibleRetryRequiresIdempotencyKey({
+                actionName: actionKind(input.action),
+                attempt: input.attempt
+              })
+            )
+          }
+          // THE REPLAY-ONLY GATE. Every durable record that could serve this
+          // dispatch has been consulted above; what remains is admitting an
+          // attempt and running the body. A replay-only engine stops here,
+          // before the attempt row, the snapshot, the boundary and the body.
+          if (deps.replayOnly !== undefined) {
+            replay.gated = true
+            const refused = {
+              runId: deps.runId,
+              stepKeyDigest,
+              attempt: input.attempt,
+              action: actionKind(input.action)
+            }
+            yield* deps.replayOnly.observe({
+              ...refused,
+              tier: input.tier,
+              outcome: runningRow === undefined ? "would-execute" : "resumes"
+            })
+            return yield* Effect.die(new ReplayOnly.WouldExecute(refused))
+          }
+          const adopted = runningRow !== undefined
+          // The admission row and its announcement commit as one unit: an
+          // `attemptStarted` never describes a row that rolled back, and an
+          // admitted attempt is never invisible to the journal.
+          yield* atomically(Effect.gen(function*() {
+            if (!adopted) {
+              const now = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+              const initialMeta: AttemptMeta = { ...declarationMeta, admittedBy: deps.owner }
+              const put = yield* attempts.put(
+                { ...attemptId, state: "running", startedAtMs: now, meta: attemptPayload(initialMeta) },
+                deps.owner
+              )
+              if (put._tag === "FenceLost" || put._tag === "RunNotFound") {
+                return yield* Effect.interrupt
+              }
+              if (put._tag !== "Inserted") {
+                return yield* Effect.fail(
+                  new AttemptAdmissionRejected({
+                    code: "attempt_admission_rejected",
+                    keyDigest: stepKeyDigest,
+                    outcome: put._tag
+                  })
+                )
+              }
+            } else {
+              // The claim is fenced at the moment it lands (issue #102): re-verify
+              // run ownership immediately before re-homing the row, so a process
+              // that lost the fence while waiting on the permit parks instead of
+              // patching a run it no longer owns. The patch below carries the
+              // owner fence itself; the heartbeat additionally refreshes the
+              // lease, and the permit excludes in-process racers.
+              const claimAtMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+              const claimFence = yield* runs.heartbeat(deps.runId, deps.owner, claimAtMs)
+              if (claimFence._tag !== "Updated") {
+                return yield* Effect.interrupt
+              }
+              // Re-home the adopted row to the current incarnation; the patch
+              // keeps the dead incarnation's other meta (tier, pre-image
+              // snapshot) intact. A vanished row or a lost fence means the
+              // durable state moved under us — surface it as self-interruption
+              // like the other fence losses.
+              const rehomed = yield* attempts.patch(attemptId, {
+                meta: attemptPayload(
+                  { ...runningMeta, ...declarationMeta, admittedBy: deps.owner } satisfies AttemptMeta
+                )
+              }, deps.owner)
+              if (rehomed._tag !== "Patched") {
+                return yield* Effect.interrupt
+              }
+            }
+            yield* emitLifecycle(
+              JournalRecords.attemptStarted(attemptSource("started"), { ...attemptId, tier: input.tier })
+            )
+            // Shares attempt admission's transaction. Refusal rolls admission
+            // back before a handler, snapshot, or sandbox can execute.
+            yield* emitCallInvoked
+          }))
+
+          /** The pre-image as the attempt row carries it: the jj operation only when reported. */
+          const snapshotFields = (id: string, operationId: string | undefined) =>
+            operationId === undefined
+              ? { meta: { snapshotId: id }, announced: { snapshotId: id } }
+              : {
+                meta: { snapshotId: id, snapshotOperationId: operationId },
+                announced: { snapshotId: id, operationId }
+              }
+          const announceSnapshot = (fields: ReturnType<typeof snapshotFields>) =>
+            emitLifecycle(
+              JournalRecords.snapshotIdentified(attemptSource("snapshot"), { ...attemptId, ...fields.announced })
+            )
+          let snapshotId: string | undefined
+          let snapshotMeta: Pick<AttemptMeta, "snapshotId" | "snapshotOperationId" | "boundarySnapshot"> = {}
+          if (input.tier !== "compensable") {
+            /**
+             * THE TIER-2 ANCHOR FOR AN ORDINARY FRAME.
+             *
+             * Time travel (`packages/smithers/flows/time-travel/docs/README.md`) requires the jj pointer
+             * current when a seq was journaled to be recorded at the frame,
+             * because replay cannot derive it. Only compensable work took a
+             * fresh snapshot, so every other frame had no anchor at all and a
+             * rewind to it restored the workspace to whatever the nearest
+             * *compensable* attempt happened to leave behind.
+             *
+             * The anchor is `carried`: it asserts "the same pointer as the
+             * previous anchor in this lineage" rather than naming one. That is
+             * the cheap half of the obligation — no jj call, no host round
+             * trip, one journal row — and the snapshot projector resolves it by
+             * copying the last change id forward. A lineage that has taken no
+             * snapshot yet carries nothing forward, which is honest: there is
+             * no pointer to restore, and a rewind reports none rather than
+             * inventing one.
+             */
+            yield* emitLifecycle(
+              JournalRecords.snapshotIdentified(attemptSource("snapshot"), { ...attemptId, carried: true })
+            )
+          }
+          if (input.tier === "compensable") {
+            const jj = deps.engineJj === undefined ? yield* Jj.Jj : deps.engineJj
+            /**
+             * THE ENGINE'S BOUNDARY HANDLE (issue #1805). The engine restores
+             * the retry sequence's original handle inside `input.snapshot` and
+             * hands back this attempt's pre-image. It runs only here, past the
+             * journal and cache lookups, so a replayed attempt never touches
+             * the boundary. An adopted attempt keeps the handle its dead
+             * incarnation persisted: that one predates any partial mutation.
+             * Every meta write below carries it, so it reaches the row before
+             * the body can run.
+             */
+            const boundaryHandle = input.snapshot === undefined ? undefined : yield* input.snapshot
+            const boundarySnapshot = runningMeta?.boundarySnapshot ?? (
+              input.snapshot === undefined
+                ? undefined
+                : boundaryHandle === undefined
+                ? {}
+                : { handle: boundaryHandle }
+            )
+            const boundaryMeta = boundarySnapshot === undefined ? {} : { boundarySnapshot }
+            if (adopted && runningMeta?.snapshotId !== undefined) {
+              // The dead incarnation persisted this attempt's own pre-image
+              // before mutating the workspace (issue #87): restore it so the
+              // re-execution runs on the clean tree, and keep it as the attempt's
+              // compensation baseline instead of snapshotting the dirty state.
+              yield* jj.restore(runningMeta.snapshotId)
+              snapshotId = runningMeta.snapshotId
+              const fields = snapshotFields(snapshotId, runningMeta.snapshotOperationId)
+              snapshotMeta = { ...fields.meta, ...boundaryMeta }
+              // Re-announcing a pre-image the row already records durably:
+              // there is no state write to pair this announcement with.
+              yield* announceSnapshot(fields)
+            } else {
+              if (input.attempt > 1) {
+                const previous = yield* attempts.get({ ...attemptId, attempt: input.attempt - 1 })
+                if (
+                  Option.isSome(previous) &&
+                  decodeMeta(previous.value.meta)?.snapshotId !== undefined
+                ) {
+                  yield* jj.restore(decodeMeta(previous.value.meta)!.snapshotId!)
+                }
+              }
+              const snapshot = yield* jj.snapshot(`smithers action ${stepKeyDigest} attempt ${input.attempt}`)
+              // The immutable commit id: a retry or an adoption restores this
+              // pointer, and a change id would follow an agent's rewrite of the
+              // pre-image change (`jj squash`) and restore the step's edits.
+              snapshotId = snapshot.commitId
+              const fields = snapshotFields(snapshotId, snapshot.operationId)
+              snapshotMeta = { ...fields.meta, ...boundaryMeta }
+              // Persist the pre-image into the running row before announcing it
+              // (issue #87): a SIGKILL mid-attempt must not lose the only
+              // reference to the clean tree, or adoption re-executes on top of
+              // the dead incarnation's partial mutations. The announcement
+              // shares the patch's transaction, so the journal never names a
+              // pre-image the row does not carry. The Jj snapshot itself stays
+              // outside: a host call must never run inside a write transaction.
+              yield* atomically(
+                attempts.patch(attemptId, {
+                  meta: attemptPayload(
+                    { ...declarationMeta, admittedBy: deps.owner, ...snapshotMeta } satisfies AttemptMeta
+                  )
+                }, deps.owner).pipe(Effect.andThen(announceSnapshot(fields)))
+              )
+            }
+          }
+
+          /**
+           * Settles an attempt its boundary refused and rethrows the refusal.
+           * Preparation, sandbox open, and settlement can each refuse, and all
+           * three are the same durable transition: a `failed` row marked
+           * `hardViolation`, committed with the hard-violation record and the
+           * failed finish. A lost fence self-interrupts instead.
+           *
+           * A boundary exists only for sealed work, while snapshots are
+           * created only for compensable work. The two capabilities are
+           * disjoint, so a refused attempt never carries a snapshot.
+           */
+          const failBoundaryAttempt = (cause: Cause.Cause<unknown>) =>
+            Effect.gen(function*() {
+              const finishedAtMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+              const finished = yield* settleAttempt({
+                ...attemptId,
+                state: "failed",
+                finishedAtMs,
+                error: attemptPayload(persistCause(cause)),
+                meta: { ...declarationMeta, hardViolation: true }
+              }, [
+                JournalRecords.hardViolation(attemptSource("hard-violation"), { ...attemptId, error: cause }),
+                JournalRecords.attemptFinished(attemptSource("finished"), { ...attemptId, state: "failed" })
+              ])
+              if (!finished) {
+                return yield* Effect.interrupt
+              }
+              return yield* Effect.failCause(cause)
+            })
+          const boundary = input.tier === "sealed" && input.metadata !== undefined
+            ? yield* StepBoundary.StepBoundary
+            : undefined
+          const preparedResult = boundary === undefined || input.metadata === undefined
+            ? undefined
+            : yield* boundary.prepare(input.metadata).pipe(Effect.exit)
+          if (preparedResult !== undefined && Exit.isFailure(preparedResult)) {
+            return yield* failBoundaryAttempt(preparedResult.cause)
+          }
+          const prepared = preparedResult === undefined ? undefined : preparedResult.value
+
+          /**
+           * THE ISOLATED EXECUTION (this lane). A sealed action carrying a
+           * boundary descriptor runs inside a workspace transaction when one
+           * is composed: the body observes only its declared read set, its
+           * writes become a diff bundle, and the host is untouched until
+           * copy-back. That is what makes whole-tree write verification
+           * structural rather than inferred — and therefore what lets a
+           * production-composed result enter the shared cache at all.
+           *
+           * The service is optional. Without it the body runs directly
+           * against the host, exactly as before, and its evidence keeps the
+           * honest omission that withholds it from the shared cache.
+           */
+          const stepSandbox = boundary === undefined || input.metadata === undefined
+            ? Option.none<StepSandbox.Service>()
+            : yield* Effect.serviceOption(StepSandbox.StepSandbox)
+          const opened = Option.isSome(stepSandbox) ? yield* stepSandbox.value.open.pipe(Effect.exit) : undefined
+          if (opened !== undefined && Exit.isFailure(opened)) {
+            // A host that cannot isolate (`layerNoop`, a refusing forest) is a
+            // typed refusal, not a crash: settle the attempt, or the row stays
+            // "running" and reads as an abandoned attempt to the reclaim
+            // machinery.
+            return yield* failBoundaryAttempt(opened.cause)
+          }
+          const sandbox = boundary === undefined || input.metadata === undefined
+            ? undefined
+            : opened !== undefined
+            ? opened.value
+            : Option.getOrUndefined(yield* Effect.serviceOption(WorkspaceSandbox.WorkspaceSandbox))
+          const isolated = sandbox === undefined || input.metadata === undefined
+            ? undefined
+            : yield* SandboxedExecution.execute({
+              sandbox,
+              descriptor: input.metadata,
+              workflow: execute(input, stepKeyDigest)
+            }).pipe(Effect.exit)
+          // The settlement is the isolated execution's whole story; the
+          // attempt's outcome is only its `result`, so the ordinary failure
+          // handling below is unchanged by which path produced it.
+          const settlement = isolated !== undefined && Exit.isSuccess(isolated) ? isolated.value : undefined
+          /**
+           * THE EFFECT BOUNDARY. An irreversible dispatch can change the world
+           * outside this journal, and a compensable one mutates the workspace
+           * a rewind must restore — both are wrapped: `intended` commits
+           * before the body starts, and the terminal record commits after it
+           * settles — `succeeded` with the recorded result, `unknown` for a
+           * failure, defect, or interruption whose external outcome nobody can
+           * testify to.
+           *
+           * The compensable record is what makes the tier-2 restore REAL: a
+           * rewind classifies the doomed suffix by its boundary rows, so a
+           * compensable action that recorded only its pre-image snapshot left
+           * nothing for the rewind to restore against — the suffix archived
+           * "completed" while the tree kept the discarded future's bytes. The
+           * record names the attempt's anchored `changeId` so the evidence and
+           * the pointer travel together.
+           *
+           * The settlement is uninterruptible: cancellation must not strand an
+           * effect that has already crossed without at least attempting to say
+           * so. Sealed work is deliberately outside this — a sealed result is
+           * cache evidence, and replay answers it from the recorded cache
+           * entry rather than an operator decision.
+           */
+          const effect = input.tier === "irreversible" || input.tier === "compensable"
+            ? {
+              id: `${deps.runId}:${stepKeyDigest}:${input.attempt}`,
+              kind: actionKind(input.action),
+              tier: input.tier,
+              runId: deps.runId,
+              lineageId,
+              sourceId: deps.sourceId,
+              attempt: input.attempt,
+              ...(deps.idempotencyKey === undefined ? {} : { idempotencyKey: deps.idempotencyKey }),
+              ...(snapshotId === undefined ? {} : { changeId: snapshotId })
+            } satisfies EffectRecords.Descriptor
+            : undefined
+          /**
+           * Whether a failing exit ends this dispatch in a durable PARK rather
+           * than a settlement (N-08). A body that reaches a wait point marks
+           * its instance suspended and interrupts, so an interrupt-only exit
+           * under a suspended instance is a run that parked. Quarantine from
+           * a nested action also parks its wrapper: resume must re-enter it
+           * to consume the inner action's preserved outcome. Both the effect
+           * boundary and the attempt row read it: a park closes neither.
+           */
+          const parked = (cause: Cause.Cause<unknown>) =>
+            Effect.map(
+              Effect.serviceOption(FlowRuntime.FlowInstance),
+              (instance) =>
+                evidenceQuarantined(cause) !== undefined ||
+                (Option.getOrUndefined(instance)?.suspended === true && Cause.hasInterruptsOnly(cause))
+            )
+          /**
+           * The row's meta as this dispatch last wrote it. The crossing writes
+           * below patch `meta` wholesale, so everything the attempt already
+           * carries — the adopted incarnation's fields, the tier declaration,
+           * the compensable pre-image — has to be carried across with it.
+           */
+          const dispatchMeta = {
+            ...runningMeta,
+            ...declarationMeta,
+            admittedBy: deps.owner,
+            ...snapshotMeta
+          } satisfies AttemptMeta
+          /**
+           * Commits a crossing into the attempt row, with the boundary record
+           * that describes it when there is one, as ONE transaction.
+           *
+           * The row is what recovery reads (see `AttemptMeta.effectCrossing`),
+           * so a crash must never leave the two disagreeing: a record without
+           * the row would re-dispatch a crossed effect, and a row without the
+           * record would settle a run from evidence the audit cannot show. A
+           * refused patch is a lost fence, surfaced as self-interruption like
+           * every other fence outcome here, and reported from OUTSIDE the
+           * transaction so the transaction's own failures stay storage
+           * failures.
+           */
+          const recordCrossing = (
+            crossing: NonNullable<AttemptMeta["effectCrossing"]>,
+            patch: AttemptStore.AttemptPatch,
+            record?: JournalEvent.Input
+          ) =>
+            Effect.flatMap(
+              atomically(Effect.gen(function*() {
+                if (record !== undefined) {
+                  yield* emitLifecycle(record)
+                }
+                const patched = yield* attempts.patch(
+                  attemptId,
+                  {
+                    ...patch,
+                    meta: attemptPayload({ ...dispatchMeta, effectCrossing: crossing } satisfies AttemptMeta)
+                  },
+                  deps.owner
+                )
+                return patched._tag === "Patched"
+              })),
+              (recorded) => recorded ? Effect.void : Effect.interrupt
+            )
+          const dispatch = effect === undefined
+            ? execute(input, stepKeyDigest)
+            : Effect.uninterruptibleMask((restore) =>
+              Effect.gen(function*() {
+                yield* recordCrossing("intended", {}, EffectRecords.boundary(effect, "intended"))
+                const exit = yield* Effect.exit(restore(execute(input, stepKeyDigest)))
+                yield* Exit.isSuccess(exit)
+                  ? recordCrossing(
+                    "succeeded",
+                    { outcome: attemptPayload(exit.value) },
+                    EffectRecords.boundary(effect, "succeeded", exit.value)
+                  )
+                  // A park leaves the boundary OPEN. The terminal record says
+                  // the effect finished and nobody can testify to how; a parked
+                  // dispatch has not finished, and the next drive re-enters the
+                  // SAME attempt. Writing it here made a routine park and
+                  // resume read `intended, unknown, intended, succeeded` in the
+                  // journal a rewind classifies a doomed suffix from.
+                  //
+                  // The park is still recorded in the ROW, with no record to
+                  // pair it with: a park and a SIGKILL mid-body leave the same
+                  // `running` row and the same open boundary, and only this
+                  // says which one happened. Without it the resumed wait looked
+                  // like an unresolved crossing and the refusal above made every
+                  // keyless irreversible wait unresumable.
+                  : Effect.flatMap(
+                    parked(exit.cause),
+                    (isPark) =>
+                      isPark
+                        ? recordCrossing("parked", {})
+                        : emitLifecycle(EffectRecords.boundary(effect, "unknown")).pipe(
+                          Effect.catchCause((cause) =>
+                            Effect.logWarning("engine-store: failed to persist unknown effect boundary", cause).pipe(
+                              Effect.annotateLogs({ runId: deps.runId, step: input.key })
+                            )
+                          )
+                        )
+                  )
+                return yield* exit
+              })
+            )
+          const outcome = isolated === undefined
+            ? yield* dispatch.pipe(Effect.exit)
+            : Exit.map(isolated, (settled) => settled.result)
+          if (Exit.isFailure(outcome)) {
+            /**
+             * A DURABLE PARK IS NOT A SETTLEMENT (N-08). A body that reaches a
+             * wait point ends its fiber by self-interruption — `Flow.suspend`
+             * marks the instance and interrupts — so an interrupt-only exit
+             * from a dispatch whose instance is suspended describes a run that
+             * parked, not an attempt that failed. Settling it below wrote a
+             * `failed` row carrying the interrupt cause, and the replay branch
+             * above then rethrew that cause on every later drive: the run died
+             * with "All fibers interrupted without error" instead of resuming.
+             * The memory engine kept no attempt row and so never saw it, which
+             * is why `HumanTask.timeoutMs`, `Action.raceAll`, and every durable
+             * wait raced against a timer worked in memory and not on SQLite.
+             *
+             * The row therefore stays `running`, which is exactly the shape the
+             * adoption path above re-enters: the next drive re-executes the
+             * body under the SAME attempt number, so the park costs nothing
+             * against the retry budget and the raced deferreds re-register
+             * against their persisted completions. `running` is also the only
+             * in-progress state the attempt store admits, so a later settlement
+             * of this attempt is an ordinary fenced `finish`.
+             */
+            if (yield* parked(outcome.cause)) {
+              return yield* Effect.failCause(outcome.cause)
+            }
+            const finishedAtMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+            // A boundary violation raised while the body ran is classified
+            // like one raised at settle time (issue #109): the row records it
+            // so a post-crash replay can re-emit the violation record, which
+            // the persisted cause alone cannot distinguish.
+            const violation = declarationViolated(outcome.cause)
+            const finished = yield* settleAttempt({
+              ...attemptId,
+              state: "failed",
+              finishedAtMs,
+              error: attemptPayload(persistCause(outcome.cause)),
+              meta: attemptPayload(
+                {
+                  ...declarationMeta,
+                  ...(violation ? { hardViolation: true as const } : {}),
+                  ...snapshotMeta
+                } satisfies AttemptMeta
+              )
+            }, [
+              ...(violation
+                ? [
+                  JournalRecords.hardViolation(attemptSource("hard-violation"), {
+                    ...attemptId,
+                    error: outcome.cause
+                  })
+                ]
+                : []),
+              JournalRecords.attemptFinished(attemptSource("finished"), { ...attemptId, state: "failed" })
+            ])
+            if (!finished) {
+              return yield* Effect.interrupt
+            }
+            return yield* Effect.failCause(outcome.cause)
+          }
+          if (settlement !== undefined) {
+            // Forensics requires both halves as journal facts, never inferred
+            // from an absence: what the
+            // transaction proposed, and that it reached the host.
+            yield* emitConverging(
+              JournalRecords.diffBundleCaptured(attemptSource("diff-bundle"), {
+                ...attemptId,
+                bundleIdentity: settlement.bundleIdentity,
+                changedPaths: settlement.files.map((change) => change.path),
+                deviations: settlement.deviations
+              })
+            )
+            // THE DISPATCH STAGE. Queued effects fire here — after copy-back
+            // settled, outside the transaction, deduplicated by idempotency
+            // key so a body that queued the same key twice, and a bundle that
+            // rebased before it landed, both send exactly once.
+            const dispatcher = Option.getOrUndefined(
+              yield* Effect.serviceOption(WorkspaceSandbox.EffectDispatcher)
+            )
+            const deduped = new Set<string>()
+            for (const queued of settlement.effects) {
+              if (deduped.has(queued.idempotencyKey)) continue
+              deduped.add(queued.idempotencyKey)
+              if (dispatcher !== undefined) yield* dispatcher.dispatch(queued)
+            }
+            yield* emitConverging(
+              JournalRecords.copyBackSettled(attemptSource("copy-back"), {
+                ...attemptId,
+                bundleIdentity: settlement.bundleIdentity,
+                rebases: settlement.rebases,
+                queued: [...deduped],
+                dispatched: dispatcher === undefined ? [] : [...deduped]
+              })
+            )
+          }
+
+          const settled = prepared === undefined || boundary === undefined
+            ? undefined
+            : yield* boundary.settle(prepared).pipe(Effect.exit)
+          yield* settled === undefined ? Effect.void : Metric.update(
+            EngineStoreMetrics.boundarySettlement[
+              Exit.isSuccess(settled)
+                ? settled.value.deviation === undefined ? "Clean" : "Deviation"
+                : settlementViolated(settled.cause)
+                ? "Violation"
+                : "Refused"
+            ],
+            1
+          )
+          if (settled !== undefined && Exit.isFailure(settled)) {
+            return yield* failBoundaryAttempt(settled.cause)
+          }
+          const settledEvidence = settled === undefined ? undefined : settled.value
+          /**
+           * THE RETIRED LIMITATION. `StepBoundary`'s filesystem layer omits
+           * `wholeTreeWritesVerified` because it can only re-measure paths it
+           * was told about, and only `layerTest` ever set it — so under the
+           * production composition nothing could enter the shared cache.
+           *
+           * An isolated execution answers the question the boundary could not:
+           * the transaction *is* the tree, so a write outside the declared set
+           * is a map comparison rather than an inference. The claim is made
+           * here, by the code that knows the body ran in isolation, and only
+           * when the whole-tree diff found no deviation — the same whole-tree
+           * view also supplies a deviation the boundary's declared-read scan
+           * would have missed entirely.
+           */
+          const evidence = settledEvidence === undefined || settlement === undefined
+            ? settledEvidence
+            : {
+              ...settledEvidence,
+              ...(settlement.deviations.length === 0
+                ? {
+                  wholeTreeWritesVerified: true as const,
+                  // `StepSandbox` is the injectable façade; the
+                  // `WorkspaceSandbox` service is the same isolated
+                  // transaction backend and proves the same read property.
+                  hermeticReadsVerified: true as const
+                }
+                : {
+                  deviation: {
+                    _tag: "ExpectedSetDeviation" as const,
+                    paths: settlement.deviations,
+                    diffIdentity: settlement.bundleIdentity
+                  }
+                })
+            }
+          const finishedAtMs = yield* Clock.currentTimeMillis.pipe(Effect.map(Math.floor))
+          // The declared read set is the key input; the prepare-time
+          // measurement is the evidence it described reality when the body
+          // ran (issue #106). A mismatch means the result was computed from
+          // different inputs than the key claims — the attempt itself is
+          // fine, but the completion must never enter the shared cache.
+          const readSetVerified = prepared !== undefined && StepBoundary.readSetMatches(prepared)
+          const meta: AttemptMeta = {
+            ...declarationMeta,
+            ...snapshotMeta,
+            ...(evidence === undefined ? {} : { boundary: evidence }),
+            ...(readSetVerified ? { readSetVerified: true as const } : {})
+          }
+          const finished = yield* settleAttempt(
+            {
+              ...attemptId,
+              state: "succeeded",
+              finishedAtMs,
+              outcome: attemptPayload(outcome.value),
+              meta: attemptPayload(meta)
+            },
+            [
+              ...(evidence?.deviation === undefined ? [] : [
+                JournalRecords.expectedSetDeviation(attemptSource("deviation"), {
+                  ...attemptId,
+                  ...evidence.deviation
+                })
+              ]),
+              JournalRecords.attemptFinished(attemptSource("finished"), { ...attemptId, state: "succeeded" })
+            ]
+          )
+          if (!finished) return yield* Effect.interrupt
+
+          const admission = CacheAdmission.completion(cacheDeclaration, meta, false)
+          if (admission._tag === "PublishCompletion") {
+            yield* recordCache({ result: outcome.value, admission, createdAtMs: finishedAtMs })
+          } else if (admission.reason === "unverified-recorded-reads") {
+            // Visible, not silent (issue #106): the run continues on its
+            // own result, but the stale declaration is journalled so the
+            // missing cache entry is explainable.
+            yield* emitConverging(
+              JournalRecords.cacheProvenance(cacheSource("unverified_read_set"), {
+                keyDigest,
+                action: "unverified_read_set"
+              })
+            )
+          }
+          return outcome.value
+        })
+      )
+    }).pipe(
+      // The operation's own span is annotated above once the digest exists;
+      // this ambient context gives every child store, boundary, and sandbox
+      // span the dispatch identity as it opens.
+      //
+      // `stepKey`, never `key`: `@smthrs/journal`'s `Redaction.isSensitiveKey`
+      // treats a standalone trailing `key` word as a credential name and
+      // replaces the value, and `RedactedLogger` applies that rule to every
+      // log annotation under a real host. The span uses the same name so a
+      // trace and a log line name the dispatch identically.
+      Effect.annotateSpans({
+        runId: deps.runId,
+        stepKey: input.key,
+        attempt: input.attempt,
+        tier: input.tier
+      }),
+      Effect.annotateLogs({
+        runId: deps.runId,
+        stepKey: input.key,
+        attempt: input.attempt,
+        tier: input.tier
+      }),
+      EngineStoreMetrics.observe({
+        timer: EngineStoreMetrics.dispatchDuration,
+        counter: EngineStoreMetrics.dispatch
+      }),
+      // A dispatch that settled without reaching the gate was served from a
+      // durable record: a recorded success, or a recorded failure it rethrew.
+      // An interruption or a defect settled nothing and reports nothing.
+      Effect.onExit((exit) => {
+        const replayOnly = deps.replayOnly
+        const stepKeyDigest = replay.stepKeyDigest
+        if (replayOnly === undefined || replay.gated || stepKeyDigest === undefined) return Effect.void
+        if (Exit.isFailure(exit) && (Cause.hasInterrupts(exit.cause) || Cause.hasDies(exit.cause))) return Effect.void
+        return replayOnly.observe({
+          runId: deps.runId,
+          stepKeyDigest,
+          attempt: input.attempt,
+          action: actionKind(input.action),
+          tier: input.tier,
+          outcome: "replayed"
+        })
+      })
+    )
+  })
+}