@smartmemory/compose 0.3.7 → 0.4.0

package/lib/fluid/smartmemory-provider.js

@@ -0,0 +1,1695 @@
+/**
+ * lib/fluid/smartmemory-provider.js — the SmartMemory fluid provider (COMP-FOH FOH-1).
+ *
+ * Implements the fluid-store seam over SmartMemory's generic MemoryItem CRUD
+ * routes, via `lib/smartmemory-client.js`. Storage only: this slice declares
+ * RECORDS, EVENTS and LINKS and nothing else. Recall, challenge, conviction,
+ * calibration and contradiction stay refused by the base class until FOH-2
+ * implements and declares them together.
+ *
+ * Storage layout (blueprint D-FOH-2):
+ *   - a record → one MemoryItem, `memory_type: fluid_<kind>`, the canonical
+ *     record JSON-encoded into `metadata.fluid_record_json`
+ *   - an event → one MemoryItem, `memory_type: fluid_event`, payload in
+ *     `metadata.fluid_event_json`
+ *
+ * WHY THE RECORD IS AN OPAQUE STRING
+ * ----------------------------------
+ * SmartMemory treats `metadata` as an open bag of graph properties it owns and
+ * rewrites. A structured record stored there does not survive the round trip:
+ * nested dicts are exploded into `parent__child` properties, empty containers
+ * are dropped, the server overwrites `created_at`, and merge-mode updates hoist
+ * old metadata to the top level as a stale shadow copy.
+ *
+ * Worse, a field CANNOT BE CLEARED: null and empty-string values are filtered
+ * out before the write and the SET simply omits them, so the previous value
+ * survives behind a 200 (filed upstream as smart-memory-core#3). Under any
+ * structured mapping, clearing a record's `priority` would silently no-op.
+ *
+ * One JSON string is inert to every one of those paths. It is never null and
+ * never empty, so it is always written; the metadata merge is a one-level
+ * spread, so one key replaces the whole record; and the backend codec
+ * marker-escapes JSON-looking strings and returns them verbatim, making the
+ * round trip byte-stable by construction rather than by hope.
+ *
+ * The cost is that record fields are not server-queryable. That costs nothing
+ * here: `/memory/list` accepts a single filter pair, the floor already filters
+ * client-side, and the seam exposes no field-query API. `handle`, `kind` and
+ * `fluid_ns` stay flat beside the blob precisely so lookup still works.
+ *
+ * WHAT THIS PROVIDER OWNS THAT THE FLOOR DOES NOT
+ * -----------------------------------------------
+ * Nothing below it enforces the seam's invariants. `memory_type` is frozen at
+ * creation server-side and the blob is opaque, so a `kind` change would diverge
+ * permanently and silently keep the old embedding policy. There is still no
+ * uniqueness CONSTRAINT on a metadata field — handle uniqueness is produced
+ * here, by allocating from a server counter and serializing every mutation
+ * behind a server lease, not enforced by the store.
+ *
+ * COORDINATION (COMP-FLUID-SEAM-GUARANTEES)
+ * -----------------------------------------
+ * Two different server primitives, for two different problems:
+ *
+ *   - `SVC-ALLOC-1` sequences allocate handles. `$inc` is atomic, so concurrent
+ *     creates get distinct numbers with no lock at all.
+ *   - `SVC-LEASE-1` leases serialize mutation, because read-modify-write on one
+ *     opaque blob cannot be made atomic by a counter.
+ *
+ * Together they are what lets `mutationScope()` answer CLUSTER instead of NONE.
+ */
+
+import { randomUUID } from 'node:crypto';
+
+import { createSmartmemoryClient } from '../smartmemory-client.js';
+import {
+  CAP,
+  CHALLENGEABLE_KINDS,
+  CONVICTION_STRATEGIES,
+  FluidAmbiguousMatch,
+  FluidConfigError,
+  FluidInvalidStrategy,
+  FluidInvalidTarget,
+  FluidKindUnsupported,
+  FluidProvider,
+  FluidRecordNotFound,
+  FluidResolutionConflict,
+  FluidResolutionIndeterminate,
+  FluidResolutionNoOp,
+  KIND,
+  MUTATION_SCOPE,
+  STORAGE_CAP,
+  normalizeRecallLimit,
+} from './provider.js';
+import {
+  HANDLE_PREFIX,
+  HANDLE_RE,
+  UNPATCHABLE,
+  assertHandle,
+  assertLink,
+  assertPatchable,
+  assertStatus,
+  eventTypeForUpdate,
+  normalizeRecord,
+} from './record-shape.js';
+import { assertValid, getFluidValidator } from './schema.js';
+
+/** Namespace markers. The analogue of the floor's `fluid_ext` presence check:
+ *  a MemoryItem without the marker is invisible to this provider, so a
+ *  workspace shared with other SmartMemory content stays safe to read. */
+const RECORD_NS = 'compose.fluid.v1';
+const EVENT_NS = 'compose.fluid.events.v1';
+
+/** Wire type prefix (blueprint D-FOH-3).
+ *
+ *  `decision` is ALREADY a registered SmartMemory memory_type with a different
+ *  meaning, so a bare kind would merge two populations in one workspace — and,
+ *  because the per-kind embedding config key is derived from `memory_type`,
+ *  would make Compose's recallability policy rewrite embedding behaviour for
+ *  native SmartMemory decisions deployment-wide. The prefix keeps the blast
+ *  radius inside Compose's own namespace. */
+const WIRE_PREFIX = 'fluid_';
+const EVENT_WIRE_TYPE = 'fluid_event';
+
+/** The route's own default page size. Every enumeration here loops; see
+ *  `_listAllItems`. */
+const PAGE_SIZE = 50;
+
+/**
+ * Kinds `recall()` may return.
+ *
+ * `cluster` stays out: it is a hand-authored grouping label, not a fuzzy-recall
+ * target. It remains stored and readable the moment you name its handle.
+ *
+ * `decision` is IN, by owner ruling 2026-08-04, reversing architecture.md §Q3's
+ * provisional INDEXED default (which flagged itself as the one item worth a
+ * second look). §Q3's stated fear was surfacing a superseded decision "as if it
+ * were live" — and the word doing the work there is *as if*. A hit carries the
+ * whole record, `status` and `killed` included, so a killed decision is
+ * distinguishable by anything that looks. See the residual limitation below.
+ *
+ * THIS LIST IS THE ENFORCEMENT, and it stays that way even once the server can
+ * be told what to embed. Filtering the OUTPUT holds regardless of how the
+ * server is configured; making it contingent on a server flag would put recall
+ * correctness at the mercy of deployment state.
+ *
+ * (Corrected 2026-08-05: the reason used to be stated as "Compose can neither
+ * set nor verify" the per-request override. Right about the outcome, wrong
+ * about the cause — core HAS the override and honours it ahead of every
+ * memory_type and config default (`_embed`, crud.py:100 and :160). No REST
+ * route plumbs it through, which is a field to expose rather than a mechanism
+ * to build. Tracked in SmartMemory as SVC-EMBED-CONTROL-1.)
+ *
+ * CORRECTED 2026-08-05, verified against live FalkorDB. This comment used to
+ * claim supersession was invisible on the superseded record — that detecting it
+ * "would mean scanning every record for links targeting this handle, which
+ * recall cannot do per hit". That was WRONG. The old node carries plain
+ * properties, readable per hit with zero extra queries:
+ *
+ *     superseded=true, superseded_by=<new_id>, superseded_at=<ts>
+ *
+ * Set by both `SmartMemory.supersede()` and `ingest_superseding()`
+ * (smart_memory.py:3531, :3546 — the latter's docstring states it outright).
+ * So a superseded decision IS distinguishable in recall today, and `decision`
+ * belongs in this set with no asterisk.
+ *
+ * The one real narrowing is different and narrower than what was claimed: both
+ * REST supersede routes CREATE the replacement (`ingest_superseding`, reached
+ * via POST /memory/{item_id}/supersede). The link-two-existing-records form,
+ * `supersede(old_id, new_id)`, is not exposed over REST — so "record B, already
+ * stored, supersedes record A" has no REST path. Tracked in SmartMemory as
+ * SVC-SUPERSEDE-LINK-1; do not build a Compose workaround for it.
+ */
+export const RECALLABLE_KINDS = Object.freeze(
+  new Set([KIND.IDEA, KIND.THREAD, KIND.QUESTION, KIND.DECISION]),
+);
+
+/**
+ * Over-fetch factor, floor and ceiling for the single recall query.
+ *
+ * One unfiltered query is issued and filtered client-side, so the fetch has to
+ * be wide enough to survive dropping every non-fluid and non-recallable hit. The
+ * floor matters more than the factor: at `limit: 1` a bare factor would fetch 4,
+ * which almost any workspace could fill with items that get filtered away.
+ *
+ * The ceiling is not decoration — `top_k` is unconstrained on the wire and the
+ * route doubles it before searching, so an unbounded value here becomes an
+ * unbounded fetch and a proportional ranking cost on the server.
+ */
+const OVERFETCH = 4;
+const MIN_FETCH = 20;
+const TOP_K_CAP = 200;
+
+/**
+ * Mutation lease (SVC-LEASE-1).
+ *
+ * ONE lease for the whole store rather than one per record, mirroring the
+ * floor's single `withDirLock` path and for the same reasons: writes here are
+ * human-scale, and a single lease also orders allocation against mutation. Two
+ * granularities would buy nothing and cost a deadlock-ordering rule.
+ *
+ * The TTL is generous because a critical section is several NETWORK calls, not
+ * a few file writes — a create is peek + allocate + event + record. It is
+ * renewed on a heartbeat rather than simply set high: a crashed holder must not
+ * strand the store for the whole TTL, and the heartbeat is what lets the TTL
+ * stay short enough for that while still covering a slow section.
+ *
+ * Server bounds (`lease.py`): 5..300s. The retry schedule is bounded and
+ * explicit — acquisition FAILS rather than waiting forever, because a caller
+ * blocked indefinitely on a lease is indistinguishable from a hung CLI.
+ */
+/**
+ * Challenge deadline (FOH-3). The client default is 3s and it spans the body
+ * read; with the LLM cascade running over up to ~10 related facts, a challenge
+ * routinely exceeds it. This is a per-call override, so CRUD/recall keep the
+ * tight default that catches a hung service.
+ */
+const CHALLENGE_TIMEOUT_MS = 30_000;
+
+// Conviction (COMP-FOH FOH-4). The resolve call is deterministic (no LLM) but
+// still does a server-side get + decay + update; 8s is generous ON PURPOSE —
+// an abort here manufactures an ambiguous outcome the provider then has to
+// reconcile, so a spurious timeout is strictly worse than a slow success.
+const RESOLVE_TIMEOUT_MS = 8_000;
+// The reconciliation poll after an ambiguous transport outcome: bounded and
+// read-only. Short, because the server mutates synchronously — if the decay is
+// going to land at all, it lands within the request's own lifetime.
+const RESOLVE_RECONCILE_ATTEMPTS = 3;
+const RESOLVE_RECONCILE_DELAY_MS = 250;
+// Contradiction linkage (COMP-FOH FOH-5). The best-effort edge write after a
+// landed decay. Safe to retry (idempotent MERGE), so a small bounded retry
+// closes over a transient network blip without ever risking a double-decay.
+const LINK_RETRY_ATTEMPTS = 3;
+const LINK_RETRY_DELAY_MS = 200;
+// Uppercase to match SmartMemory's managed contradiction path (framework.py:887)
+// so a Compose-written edge and a native-written one read back under one label.
+const CONTRADICTS_EDGE = 'CONTRADICTS';
+/** Numeric tolerance for the decay postcondition. The confidence value round
+ *  trips Python → JSON → JS; 1.0/0.5/0.0 are exact doubles but a legacy value
+ *  like 0.9 is not, and a strict `===` on a destructive-write check would
+ *  false-alarm on representation noise. */
+const CONFIDENCE_EPSILON = 1e-9;
+
+/** Python `[:200]` slices Unicode CODE POINTS; JS `String.slice` cuts UTF-16
+ *  code units. The attribution check compares our text against what the server
+ *  stored, so the truncation must match Python's or an astral character near
+ *  the boundary mis-classifies a landed decay. */
+function codePointSlice(s, n) {
+  return Array.from(String(s ?? '')).slice(0, n).join('');
+}
+
+/**
+ * Classify what a resolve attempt actually did, from the confidence-history
+ * reads bracketing it (COMP-FOH FOH-4 D4). Pure — shared by the clean-response
+ * path and the reconciliation poll so the two can never disagree.
+ *
+ * Never trusts the `/resolve` response: the server ignores `apply_decay`'s
+ * success and its `confidence` field is ambiguous. The verdict comes from
+ * SmartMemory's deterministic contract instead: one decay = challenge_count+1,
+ * confidence = max(0, old - 0.5), one attributed history event.
+ *
+ * @param {object} pre  confidence-history envelope read before the resolve
+ * @param {object} post confidence-history envelope read after
+ * @param {string} expectedFact our new_fact, truncated to 200 code points —
+ *   matched against the newest event so someone ELSE's numerically identical
+ *   decay is never mistaken for ours
+ * @returns {'landed'|'no-op'|'conflict'|'indeterminate'}
+ */
+export function classifyResolution(pre, post, expectedFact) {
+  const finite = (v) => typeof v === 'number' && Number.isFinite(v);
+  const near = (a, b) => finite(a) && finite(b) && Math.abs(a - b) < CONFIDENCE_EPSILON;
+  // The envelope must be structurally sound AND be about the same item as the
+  // pre-read — a wrong-item envelope must never classify anything.
+  if (!post || !finite(post.current_confidence)
+      || !Number.isInteger(post.challenge_count) || !Array.isArray(post.history)
+      || post.item_id !== pre.item_id) {
+    return 'indeterminate';
+  }
+
+  const delta = post.challenge_count - pre.challenge_count;
+  if (delta === 1) {
+    const ev = post.history[post.history.length - 1];
+    // Ours? The reason is fixed by the strategy and the fact text is ours.
+    const attributed = !!ev
+      && typeof ev.timestamp === 'string'
+      && ev.reason === 'manual_resolution:accept_new'
+      && ev.conflicting_fact === expectedFact;
+    if (!attributed) return 'conflict'; // someone else's decay landed; ours did not
+    const expected = Math.max(0, pre.current_confidence - 0.5);
+    const numeric = near(post.current_confidence, expected)
+      && near(ev.old_confidence, pre.current_confidence)
+      && near(ev.new_confidence, expected)
+      && ev.decay_factor === 0.5;
+    // Attributed but numerically wrong = a partial or corrupt write (the
+    // pre-fix-runtime shape: count moves, confidence does not). Not retryable.
+    return numeric ? 'landed' : 'indeterminate';
+  }
+  if (delta === 0) {
+    // A genuine no-op only if NOTHING moved — confidence AND the history
+    // CONTENT. Comparing `history_count` alone is not enough twice over: the
+    // real route reads the item and the history separately (a torn read can
+    // pair old confidence with new history), and once the history sits at its
+    // server-side cap (20) an external decay rotates entries without changing
+    // the count. Both must land in 'indeterminate', never in the one verdict
+    // that invites a retry. Same JSON round-trip on both reads, so exact
+    // comparison is right here.
+    const unchanged = post.current_confidence === pre.current_confidence
+      && JSON.stringify(post.history) === JSON.stringify(pre.history);
+    return unchanged ? 'no-op' : 'indeterminate';
+  }
+  if (delta > 1) return 'conflict'; // more than one decay interleaved
+  return 'indeterminate'; // count regression: malformed or incompatible runtime
+}
+
+const LEASE_TTL_SECONDS = 60;
+const LEASE_RENEW_MS = 20_000;
+const LEASE_BACKOFF_MS = Object.freeze([0, 50, 150, 400, 900, 1500, 2500, 4000]);
+const LEASE_KEY = 'compose.fluid.mutate';
+
+/** Handle counters, one sequence per kind. Server-side these are scoped to the
+ *  workspace, so the name carries only the namespace and the kind. Matches the
+ *  service's `SEQUENCE_NAME_PATTERN`. */
+function sequenceNameFor(kind) { return `compose.fluid.handle.${kind}`; }
+
+function nowIso() { return new Date().toISOString(); }
+
+function sleep(ms) { return new Promise((resolve) => setTimeout(resolve, ms)); }
+
+/** `memory_type` for a record kind. Never the bare kind. */
+function wireTypeFor(kind) { return `${WIRE_PREFIX}${kind}`; }
+
+export class SmartMemoryFluidProvider extends FluidProvider {
+  name() { return 'smartmemory'; }
+
+  /** Storage, plus RECALL (FOH-2), CHALLENGE (FOH-3), CONVICTION (FOH-4) and
+   *  CONTRADICTION (FOH-5). Declaring a capability without implementing it is the
+   *  one thing PROVIDER-SEAM forbids, so each line here moves with its method
+   *  (`recall()`, `challenge()`, `conviction()`/`resolveConflict()`,
+   *  `contradictions()`) — CALIBRATION stays undeclared (no subject exists) and
+   *  keeps inheriting refusal. */
+  capabilities() {
+    return new Set([
+      STORAGE_CAP.RECORDS, STORAGE_CAP.EVENTS, STORAGE_CAP.LINKS,
+      CAP.RECALL, CAP.CHALLENGE, CAP.CONVICTION, CAP.CONTRADICTION,
+    ]);
+  }
+
+  /**
+   * The floor's full kind set, not `idea` alone.
+   *
+   * The pilot workload requires it: the ideabox import creates CLUSTERS before
+   * ideas, because members reference them by handle, so an idea-only provider
+   * throws on the import's first write. Kinds are free here because storage is
+   * generic — `memory_type` is a free-form string with no typed schema behind
+   * it — so matching the floor exactly makes a provider swap lossless rather
+   * than parity-gapped.
+   *
+   * `position` and `joint` stay refused, matching the floor: the judgment layer
+   * already owns both kinds with its own store and write tools, and accepting
+   * them would give one kind two canons.
+   */
+  supportedKinds() {
+    return new Set([KIND.IDEA, KIND.DECISION, KIND.THREAD, KIND.QUESTION, KIND.CLUSTER]);
+  }
+
+  /**
+   * CLUSTER since COMP-FLUID-SEAM-GUARANTEES. This used to be a confession.
+   *
+   * It was NONE because allocation read the issued set and added one across two
+   * remote calls with nothing holding the gap, so N concurrent creates allocated
+   * the same handle. No record was LOST: this store writes by `item_id`, so all
+   * N persisted and N-1 were silently orphaned behind a handle resolving to one
+   * of them. The floor loses them outright because it keys the file by handle —
+   * do not carry the floor's loss claim across this seam (measured 2026-08-05).
+   * `lib/dir-lock.js`
+   * could not have fixed it: it is a local mutex and this store is reachable from
+   * every machine sharing the workspace, which is the entire reason to use it.
+   *
+   * Both halves now rest on server primitives, and they are different primitives
+   * on purpose:
+   *
+   *   - ALLOCATION is not serialized at all, because it does not need to be. A
+   *     sequence `$inc` is atomic per document, so concurrent callers receive
+   *     distinct numbers with no lock and no contention. Wrapping it in the
+   *     lease would be strictly worse — slower, and no safer.
+   *   - MUTATION is serialized by a scoped lease, because read-modify-write on
+   *     one record genuinely cannot be made atomic by a counter. This is the
+   *     quieter half: a lost update leaves no trace, both writers succeed and one
+   *     edit is simply gone.
+   *
+   * Declared honestly in BOTH directions — the conformance suite asserts the
+   * failure when a provider declares NONE, so a stale declaration here starts
+   * failing rather than silently over-promising.
+   */
+  mutationScope() { return MUTATION_SCOPE.CLUSTER; }
+
+  /** A workspace on a server. Reachable from every machine holding the key. */
+  isShared() { return true; }
+
+  /**
+   * @param {string} cwd project root
+   * @param {object} config merged fluid.smartmemory + top-level smartmemory config
+   * @param {string} config.baseUrl SmartMemory endpoint
+   * @param {string} config.apiKeyEnv env var holding the API key
+   * @param {string} config.workspaceId configured workspace (owner ruling: never derived)
+   * @param {object} [config.client] injected client, for tests
+   */
+  async init(cwd, config = {}) {
+    this.cwd = cwd;
+    this.config = config;
+
+    // Fail loud, before any network call, naming the exact missing setting.
+    //
+    // All four are checked, not just the block's presence: the client
+    // interpolates baseUrl straight into the URL, so an unset endpoint would
+    // otherwise surface as a request failure against the literal string
+    // "undefined" — blaming the server for a local misconfiguration.
+    if (!config.baseUrl) {
+      throw new FluidConfigError(
+        'compose: fluid provider "smartmemory" requires smartmemory.baseUrl to be set',
+        { provider: 'smartmemory', setting: 'smartmemory.baseUrl' }
+      );
+    }
+    if (!config.apiKeyEnv) {
+      throw new FluidConfigError(
+        'compose: fluid provider "smartmemory" requires smartmemory.apiKeyEnv to be set',
+        { provider: 'smartmemory', setting: 'smartmemory.apiKeyEnv' }
+      );
+    }
+    if (!process.env[config.apiKeyEnv]) {
+      throw new FluidConfigError(
+        `compose: fluid provider "smartmemory" needs an API key in $${config.apiKeyEnv}, ` +
+        'which is unset or empty. The key needs the read:memories, write:memories AND ' +
+        'delete:memories scopes — a key missing delete fails only on deleteRecord, long ' +
+        'after setup appears to have worked.',
+        { provider: 'smartmemory', setting: config.apiKeyEnv }
+      );
+    }
+    // Owner ruling (blueprint C4): the workspace id is CONFIGURED, never
+    // derived. A locally derived project tag was never usable — the service
+    // validates the header against the principal's memberships and 403s a
+    // workspace they are not a member of.
+    if (!config.workspaceId) {
+      throw new FluidConfigError(
+        'compose: fluid provider "smartmemory" requires fluid.smartmemory.workspaceId. ' +
+        'It must be a workspace the configured API key is a member of; it is not derived ' +
+        'from the project directory.',
+        { provider: 'smartmemory', setting: 'fluid.smartmemory.workspaceId' }
+      );
+    }
+
+    this.client = config.client ?? createSmartmemoryClient({
+      baseUrl: config.baseUrl,
+      apiKeyEnv: config.apiKeyEnv,
+      workspaceId: config.workspaceId,
+      timeoutMs: config.timeoutMs,
+    });
+
+    /** The serialization mechanism this provider's `mutationScope()` claims.
+     *  Exposed under this name because the conformance suite cross-checks the
+     *  declaration against a mechanism — the floor exposes `lockPath`, a
+     *  server-backed provider exposes this. Leases are workspace-scoped
+     *  server-side, so the key needs no workspace component. */
+    this.leaseClient = this.client;
+    this.leaseKey = LEASE_KEY;
+    return this;
+  }
+
+  // -------------------------------------------------------------------------
+  // Mapping
+  // -------------------------------------------------------------------------
+
+  /**
+   * The searchable projection (blueprint D-FOH-1).
+   *
+   * Title, body and discussion text, not the title alone. Embeddings are
+   * generated from `content`, so a title-only projection would leave a later
+   * recall slice unable to reach an idea's prose while the architecture
+   * declares ideas fully recallable — a failure that would surface only after
+   * that slice shipped, as empty results.
+   *
+   * DISCLOSED LIMITATION: PATCH does not reindex, so an edited record keeps its
+   * original embedding. Nothing reads embeddings in FOH-1, so nothing is broken
+   * today; this is FOH-2's entry gate. Do NOT work around it by delete +
+   * recreate, which burns a handle the tombstone invariant forbids.
+   */
+  _renderContent(record) {
+    const parts = [record.title];
+    if (record.body) parts.push(record.body);
+    for (const entry of record.discussion ?? []) {
+      if (entry?.text) parts.push(entry.text);
+    }
+    return parts.join('\n\n');
+  }
+
+  /** Record → wire metadata. Three flat lookup fields, then the sealed blob. */
+  _toMetadata(record) {
+    return {
+      fluid_ns: RECORD_NS,
+      handle: record.handle,
+      kind: record.kind,
+      fluid_record_json: JSON.stringify(record),
+    };
+  }
+
+  /**
+   * Wire item → record, or null when the item is not one of ours / unreadable.
+   *
+   * A blob that fails to parse is treated as absent rather than thrown: these
+   * items live in a shared workspace, and one corrupt row must not make every
+   * enumeration fail. The `handle` is reported so it is findable.
+   */
+  _fromItem(item) {
+    const meta = item?.metadata ?? {};
+    if (meta.fluid_ns !== RECORD_NS) return null;
+    let parsed;
+    try {
+      parsed = JSON.parse(meta.fluid_record_json);
+    } catch {
+      process.emitWarning(
+        `fluid(smartmemory): record ${meta.handle ?? item?.item_id} has an unreadable ` +
+        'fluid_record_json blob and was skipped',
+      );
+      return null;
+    }
+    if (!parsed || typeof parsed !== 'object') return null;
+    // Normalize on every read: a record written before a field existed, or
+    // imported from the floor, reads back in today's shape without a migration.
+    return normalizeRecord(parsed);
+  }
+
+  // -------------------------------------------------------------------------
+  // Enumeration
+  // -------------------------------------------------------------------------
+
+  /**
+   * Every matching item, following pagination to the end.
+   *
+   * The route defaults to `limit=50`. A provider that silently saw only the
+   * first page would under-report the ideabox — and once the projection is
+   * generated from it, DELETE the rest from `ideabox.md`.
+   *
+   * Only one metadata filter pair is supported server-side, so the namespace
+   * check is applied client-side by the caller's mapper.
+   */
+  async _listAllItems({ metadataKey, metadataValue } = {}) {
+    const out = [];
+    let offset = 0;
+    for (;;) {
+      const page = await this.client.listItems({
+        limit: PAGE_SIZE, offset, metadataKey, metadataValue,
+      });
+      const items = page.items ?? [];
+      out.push(...items);
+      // Stop on a short page OR on reaching `total`. Either alone is fragile:
+      // a full last page with no `total` would loop forever without the first,
+      // and a `total` that lags a concurrent write would truncate without the
+      // second.
+      if (items.length < PAGE_SIZE) break;
+      if (typeof page.total === 'number' && out.length >= page.total) break;
+      offset += items.length;
+    }
+    return out;
+  }
+
+  /** Every live record item, namespace-filtered. */
+  async _allRecordItems() {
+    const items = await this._listAllItems({
+      metadataKey: 'fluid_ns', metadataValue: RECORD_NS,
+    });
+    return items.filter((it) => it?.metadata?.fluid_ns === RECORD_NS);
+  }
+
+  // -------------------------------------------------------------------------
+  // Handle resolution and duplicate repair (blueprint D-FOH-4)
+  // -------------------------------------------------------------------------
+
+  /**
+   * Every live item carrying this handle, in a deterministic total order.
+   *
+   * Ordered by the SERVER-stamped `metadata.created_at`, ties broken by
+   * `item_id`. The server stamps that field itself on every add, so this is one
+   * clock rather than each client's — and a racing writer cannot forge an
+   * earlier one. The record's own `created_at`, inside the blob, is the
+   * record's account of itself and is deliberately NOT used for ordering.
+   */
+  async _resolveItems(handle) {
+    const items = await this._listAllItems({ metadataKey: 'handle', metadataValue: handle });
+    return items
+      .filter((it) => it?.metadata?.fluid_ns === RECORD_NS && it?.metadata?.handle === handle)
+      .sort((a, b) => {
+        const ca = a.metadata?.created_at ?? '';
+        const cb = b.metadata?.created_at ?? '';
+        if (ca !== cb) return ca < cb ? -1 : 1;
+        return String(a.item_id) < String(b.item_id) ? -1 : 1;
+      });
+  }
+
+  /**
+   * Read resolution: deterministic, never writes, never throws on a duplicate.
+   *
+   * SmartMemory has no uniqueness constraint on a metadata field, so a duplicate
+   * is still REPRESENTABLE even though the counter and the lease mean this
+   * provider no longer produces one: a workspace written by a pre-SVC-ALLOC-1
+   * Compose can already hold a pair, and nothing retroactively repairs history.
+   * Throwing here — the earlier design — makes that state PERMANENTLY
+   * unreadable, which is strictly worse than the ambiguity it was meant to flag.
+   * Earliest wins, loudly.
+   */
+  async _resolveOne(handle) {
+    const items = await this._resolveItems(handle);
+    if (items.length === 0) return null;
+    if (items.length > 1) {
+      process.emitWarning(
+        `fluid(smartmemory): handle ${handle} resolves to ${items.length} live records; ` +
+        'using the earliest. The duplicate is repaired on the next write to this handle.',
+      );
+    }
+    return items[0];
+  }
+
+  /**
+   * Write resolution: repairs a duplicate before proceeding.
+   *
+   * Keeps the earliest and reassigns every later item a freshly allocated
+   * handle, appending a `reassigned` event so the trail is auditable. Nothing
+   * is discarded and no handle is ever reissued — which makes this strictly
+   * better than the local floor, where two racing creates collide on one path
+   * and one idea is silently lost.
+   */
+  async _resolveForWrite(handle) {
+    const items = await this._resolveItems(handle);
+    if (items.length === 0) return null;
+    if (items.length === 1) return items[0];
+
+    const [keep, ...duplicates] = items;
+    for (const dup of duplicates) {
+      const record = this._fromItem(dup);
+      if (!record) continue;
+      const fresh = await this._nextHandle(record.kind);
+      const moved = { ...record, handle: fresh, updated_at: nowIso() };
+      await this.client.updateItem(dup.item_id, {
+        content: this._renderContent(moved),
+        metadata: this._toMetadata(moved),
+      });
+      await this.appendEvent({
+        handle: fresh,
+        type: 'created',
+        at: nowIso(),
+        detail: { kind: record.kind, reassigned_from: handle },
+      });
+      process.emitWarning(
+        `fluid(smartmemory): duplicate handle ${handle} repaired — the later record ` +
+        `was reassigned ${fresh}`,
+      );
+    }
+    return keep;
+  }
+
+  // -------------------------------------------------------------------------
+  // Handle allocation
+  // -------------------------------------------------------------------------
+
+  /** Every handle ever issued — live records UNION the append-only event log.
+   *  Derived from live records alone it would miss a deleted record's handle,
+   *  and handles are quoted in docs, commits and conversation.
+   *
+   *  This is the expensive read in the file: two full paginated enumerations.
+   *  It used to run on EVERY create; since COMP-FLUID-SEAM-GUARANTEES it runs only on the
+   *  explicit-handle path (the import) and once per kind to seed the counter.
+   *  Keep it that way — anything that puts it back in the automatic path undoes
+   *  the change. */
+  async _issuedHandles() {
+    const issued = new Set();
+    for (const item of await this._allRecordItems()) {
+      const h = item?.metadata?.handle;
+      if (h) issued.add(h);
+    }
+    for (const event of await this.readEvents()) {
+      if (event?.handle) issued.add(event.handle);
+    }
+    return issued;
+  }
+
+  /** The highest number already issued under `prefix`. Only ever read to SEED
+   *  the counter — never to allocate from. See `_sequenceFloor`. */
+  async _maxIssued(prefix) {
+    let max = 0;
+    for (const handle of await this._issuedHandles()) {
+      const m = HANDLE_RE.exec(handle);
+      if (m && m[1] === prefix) max = Math.max(max, Number(m[2]));
+    }
+    return max;
+  }
+
+  /**
+   * The `floor` to send with an allocation, or `undefined` when none is needed.
+   *
+   * `floor` is a high-water mark applied with `$max` BEFORE the `$inc`, so it can
+   * only ever raise the counter — idempotent, and safe to race.
+   *
+   * It is computed from a full scan EXACTLY ONCE per kind per workspace: on the
+   * first touch, when the counter does not exist yet. That case is not
+   * hypothetical — a workspace written by a pre-sequence Compose already holds
+   * handles the counter has never seen, and a counter starting at 1 would reissue
+   * every one of them. Once seeded, the counter is authoritative and the scan
+   * never runs again, which is the whole point of the change: allocation drops
+   * from two full paginated enumerations (every record UNION every event) to one
+   * cheap GET plus one POST.
+   *
+   * Racing two seeders is harmless: both scan the same store, both send the same
+   * `$max`, and the `$inc` still hands them distinct numbers.
+   *
+   * @param {string} kind
+   * @param {number} atLeast raise the counter to at least this, for a handle
+   *   supplied by the caller rather than allocated
+   */
+  async _sequenceFloor(kind, atLeast = 0) {
+    const seeded = await this.client.peekSequence(sequenceNameFor(kind)) !== null;
+    if (seeded) return atLeast > 0 ? atLeast : undefined;
+    return Math.max(await this._maxIssued(HANDLE_PREFIX[kind]), atLeast);
+  }
+
+  /**
+   * The next handle for `kind`, from the server's monotonic counter.
+   *
+   * Deliberately NOT wrapped in the mutation lease. `$inc` is atomic per
+   * document, so concurrent callers receive distinct numbers with no lock;
+   * holding the lease across it would serialize allocations that never needed
+   * serializing. The counter is also the reason the old
+   * `if (issued.has(candidate)) throw` guard is gone: it defended against a
+   * derivation that could repeat itself, and a counter cannot.
+   *
+   * Numbers allocated but not used are LOST. Gaps are guaranteed by the
+   * primitive and cost nothing here — a handle is a citation, not a count.
+   */
+  async _nextHandle(kind) {
+    const floor = await this._sequenceFloor(kind);
+    const { value } = await this.client.allocateSequence(sequenceNameFor(kind), { floor });
+    return `${HANDLE_PREFIX[kind]}-${value}`;
+  }
+
+  /**
+   * Raise the counter past a handle the CALLER supplied, so a later allocation
+   * cannot hand it out a second time.
+   *
+   * Passing `n - 1` rather than `n` is what keeps a sequential import gapless:
+   * `$max(n-1)` then `$inc` leaves the counter at exactly `n`, so the next
+   * automatic handle is `n + 1`. The number this call consumes is discarded —
+   * there is no way to raise the floor without consuming one, and one wasted
+   * number per explicit handle is a great deal cheaper than a counter that lags
+   * a live handle.
+   */
+  async _burnHandle(kind, handle) {
+    const m = HANDLE_RE.exec(handle);
+    if (!m) return;
+    const floor = await this._sequenceFloor(kind, Math.max(Number(m[2]) - 1, 0));
+    await this.client.allocateSequence(sequenceNameFor(kind), { floor });
+  }
+
+  async _handleWasIssued(handle) {
+    return (await this._issuedHandles()).has(handle);
+  }
+
+  /**
+   * Was this handle issued, never made live, and never deleted?
+   *
+   * The seam's `reclaimAborted` predicate, mirroring the floor's
+   * `_isAbortedAllocation` deliberately — the narrowness IS the safety. Only a
+   * handle stranded between its tombstone and its record qualifies:
+   *
+   *   - a handle whose record EXISTS is in use, and
+   *   - a handle with a `deleted` event is retired, and retired handles stay
+   *     retired, because reissuing one repoints every existing citation at a
+   *     different record.
+   *
+   * So the only thing this can hand back is a handle that names nothing and
+   * never will — which is exactly the state an interrupted import leaves and
+   * nothing else produces.
+   */
+  async _isAbortedAllocation(handle) {
+    if (await this.getRecord(handle)) return false;
+    for (const event of await this.readEvents()) {
+      if (event?.handle === handle && event.type === 'deleted') return false;
+    }
+    return true;
+  }
+
+  // -------------------------------------------------------------------------
+  // Mutation lease (SVC-LEASE-1) — what makes mutationScope() CLUSTER
+  // -------------------------------------------------------------------------
+
+  /**
+   * Run `fn` while holding the store's single mutation lease.
+   *
+   * Every public mutating method is a thin wrapper over a `*Locked` body for the
+   * same reason the floor's are: **this is NOT reentrant.** A locked body that
+   * called another public method would ask the server for a lease it already
+   * holds, be told `lock_held` by its own other half, and fail after the full
+   * backoff. Compose by calling the inner form.
+   *
+   * `appendEvent` deliberately stays unlocked: it is an append-only write that
+   * every locked body performs, and locking it would deadlock all of them.
+   *
+   * Failure is loud and total. A caller that believes it is serialized and is
+   * not loses records without an error — the exact failure this whole mechanism
+   * exists to make impossible — so exhausting the backoff throws rather than
+   * proceeding unserialized.
+   */
+  async _withLease(op, fn) {
+    let lease = null;
+    for (const wait of LEASE_BACKOFF_MS) {
+      if (wait) await sleep(wait);
+      lease = await this.client.acquireLock(this.leaseKey, { ttlSeconds: LEASE_TTL_SECONDS });
+      if (lease) break;
+    }
+    if (!lease) {
+      throw new Error(
+        `fluid(smartmemory): ${op} could not acquire the mutation lease ` +
+        `"${this.leaseKey}" after ${LEASE_BACKOFF_MS.length} attempts — another writer ` +
+        'is holding it. Nothing was written.',
+      );
+    }
+
+    // Heartbeat rather than a long TTL: a crashed holder must not strand every
+    // other writer for the full lease, and renewing is what lets the TTL stay
+    // short enough for that while still covering a slow section.
+    const heartbeat = setInterval(() => {
+      this.client.renewLock(this.leaseKey, lease.token, { ttlSeconds: LEASE_TTL_SECONDS })
+        .then((renewed) => {
+          if (renewed === null) {
+            process.emitWarning(
+              `fluid(smartmemory): the mutation lease expired during ${op} and was taken by ` +
+              'another writer; this write is no longer serialized',
+            );
+          }
+        })
+        .catch((err) => {
+          process.emitWarning(`fluid(smartmemory): lease renewal failed during ${op}: ${err.message}`);
+        });
+    }, LEASE_RENEW_MS);
+    // Never hold the event loop open for a renewal nobody is waiting on.
+    heartbeat.unref?.();
+
+    try {
+      return await fn();
+    } finally {
+      clearInterval(heartbeat);
+      try {
+        await this.client.releaseLock(this.leaseKey, lease.token);
+      } catch (err) {
+        // Releasing is an optimisation, not a correctness requirement — the TTL
+        // reclaims it either way. Warn rather than mask the caller's own error.
+        process.emitWarning(
+          `fluid(smartmemory): releasing the mutation lease after ${op} failed ` +
+          `(${err.message}); it expires within ${LEASE_TTL_SECONDS}s`,
+        );
+      }
+    }
+  }
+
+  // -------------------------------------------------------------------------
+  // Records
+  // -------------------------------------------------------------------------
+
+  /** Absence is `null`, matching the floor. Only mutating paths throw. A
+   *  malformed handle is a miss, not a crash — this is the lookup a caller
+   *  makes with untrusted input. */
+  async getRecord(handle) {
+    if (!HANDLE_RE.test(handle ?? '')) return null;
+    const item = await this._resolveOne(handle);
+    return item ? this._fromItem(item) : null;
+  }
+
+  async listRecords(filter = {}) {
+    const items = await this._allRecordItems();
+    let records = items.map((it) => this._fromItem(it)).filter(Boolean);
+    if (filter.kind) records = records.filter((r) => r.kind === filter.kind);
+    if (filter.status) records = records.filter((r) => r.status === filter.status);
+    if (filter.cluster !== undefined) records = records.filter((r) => r.cluster === filter.cluster);
+    // Same stable order as the floor, so a projection regenerated after a
+    // provider swap does not churn.
+    return records.sort((a, b) => {
+      const ca = a.cluster_order ?? Number.MAX_SAFE_INTEGER;
+      const cb = b.cluster_order ?? Number.MAX_SAFE_INTEGER;
+      if (ca !== cb) return ca - cb;
+      return (Number(HANDLE_RE.exec(a.handle)?.[2] ?? 0)) - (Number(HANDLE_RE.exec(b.handle)?.[2] ?? 0));
+    });
+  }
+
+  /**
+   * Genuinely atomic since COMP-FLUID-SEAM-GUARANTEES: the lookup and the create happen
+   * inside ONE hold of the mutation lease, so a second writer racing the same
+   * title waits, then finds the record the first one made instead of creating a
+   * twin. Overridden rather than inherited because the base implementation
+   * composes on the PUBLIC `createRecord`, which would try to re-acquire.
+   */
+  async findOrCreateRecord({ kind, title }, input = {}) {
+    return this._withLease('findOrCreateRecord', async () => {
+      const matches = (await this.listRecords({ kind }))
+        .filter((r) => r.title.toLowerCase() === String(title).toLowerCase());
+      if (matches.length > 1) {
+        throw new FluidAmbiguousMatch(kind, title, matches.map((m) => m.handle));
+      }
+      if (matches.length === 1) return { record: matches[0], created: false };
+      return { record: await this._createRecordLocked({ ...input, kind, title }), created: true };
+    });
+  }
+
+  async createRecord(input) {
+    return this._withLease('createRecord', () => this._createRecordLocked(input));
+  }
+
+  async _createRecordLocked(input) {
+    const kind = input.kind ?? KIND.IDEA;
+    this.requireKind(kind);
+    if (!input.title) throw new Error('fluid: createRecord requires a title');
+
+    let handle;
+    if (input.handle === undefined) {
+      handle = await this._nextHandle(kind);
+    } else {
+      handle = assertHandle(input.handle, kind);
+      // Checked against every handle EVER issued, not just the live ones. A
+      // retired handle is still spoken for: reissuing it would repoint existing
+      // citations at a different record.
+      //
+      // `reclaimAborted` is honoured here with the floor's exact semantics
+      // (COMP-FLUID-SEAM-GUARANTEES). It was ignored until now, and the guard
+      // below threw unconditionally — which made the one-time import
+      // NON-RESTARTABLE against this provider. The window it has to survive is
+      // far wider here than on the floor: creation burns the handle in one
+      // network call and writes the record in the next, so any blip between them
+      // stranded that handle permanently and `ensureIdeaboxMigrated`'s resume
+      // path then failed forever, since `importIdeabox` passes `reclaimAborted`
+      // and this provider dropped it.
+      //
+      // Needs no server primitive — it is a question about this provider's own
+      // event log and record set, both already readable, which is why it was
+      // fixed before SVC-LEASE-1 landed rather than waiting on it. The counter
+      // does not subsume this check and never will: it knows the numbers it has
+      // handed out, not which of them were retired.
+      if (await this._handleWasIssued(handle)
+        && !(input.reclaimAborted && await this._isAbortedAllocation(handle))) {
+        throw new Error(
+          `fluid: handle ${handle} has already been issued and cannot be reused ` +
+          `(handles are external citations; retired ones stay retired)`
+        );
+      }
+      // A caller-supplied handle bypasses the counter, so the counter has to be
+      // told about it — otherwise a later automatic allocation could hand out
+      // the same number. Done BEFORE anything is written, so a failure here
+      // leaves no record whose handle the counter does not know.
+      await this._burnHandle(kind, handle);
+    }
+
+    for (const link of input.links ?? []) assertLink(link);
+    const status = assertStatus(input.status ?? 'new');
+    const now = nowIso();
+    const record = normalizeRecord({
+      handle,
+      kind,
+      title: input.title,
+      body: input.body ?? '',
+      status,
+      status_label: input.status_label ?? null,
+      priority: input.priority ?? null,
+      effort: input.effort ?? null,
+      impact: input.impact ?? null,
+      cluster: input.cluster ?? null,
+      cluster_order: input.cluster_order ?? null,
+      tags: input.tags ?? [],
+      source: input.source ?? null,
+      links: input.links ?? [],
+      killed: input.killed ?? null,
+      discussion: input.discussion ?? [],
+      provenance: {
+        origin: input.provenance?.origin ?? 'cli:ideabox',
+        recorded_at: input.provenance?.recorded_at ?? now,
+        author: input.provenance?.author ?? null,
+      },
+      // Provider-assigned and provider-scoped, per the contract: `id` changes
+      // when a record moves providers, while `handle` survives the swap.
+      // Deliberately NOT SmartMemory's item_id, which does not exist until the
+      // write returns — the contract requires `id` on the object being written.
+      id: randomUUID(),
+      created_at: now,
+      updated_at: now,
+    });
+    assertValid('record', record, 'record');
+
+    // The tombstone is written BEFORE the record exists. Persisting first and
+    // appending after means a failed append leaves a discoverable record whose
+    // handle was never burned — delete it and the handle is reissued, defeating
+    // the invariant. Burning first can at worst waste a handle, and a wasted
+    // handle is free while a reissued one is unrecoverable.
+    await this.appendEvent({
+      handle,
+      type: input.provenance?.origin === 'import:ideabox' ? 'imported' : 'created',
+      at: now,
+      detail: { kind },
+    });
+
+    await this.client.createItem({
+      content: this._renderContent(record),
+      memoryType: wireTypeFor(kind),
+      metadata: this._toMetadata(record),
+    });
+
+    return record;
+  }
+
+  async updateRecord(handle, patch) {
+    return this._withLease('updateRecord', () => this._updateRecordLocked(handle, patch));
+  }
+
+  async _updateRecordLocked(handle, patch) {
+    const item = await this._resolveForWrite(handle);
+    if (!item) throw new FluidRecordNotFound(handle, this.name());
+
+    assertPatchable(patch, this.name());
+
+    const current = this._fromItem(item);
+    if (!current) throw new FluidRecordNotFound(handle, this.name());
+
+    const merged = {
+      ...current,
+      ...patch,
+      ...Object.fromEntries(UNPATCHABLE.map((f) => [f, current[f]])),
+      updated_at: nowIso(),
+    };
+
+    // Validate the RAW merge, then normalize — order is load-bearing.
+    // Normalizing first would hand the schema an already-sanitized object and
+    // the schema would approve what the sanitizer had quietly repaired.
+    assertStatus(merged.status);
+    assertValid('record', merged, 'record');
+    for (const link of merged.links) assertLink(link);
+
+    const next = normalizeRecord(merged);
+    const eventType = eventTypeForUpdate(current, next, patch);
+
+    // One metadata key carrying the whole record. The server's merge is a
+    // one-level spread, so this replaces the record wholesale while leaving
+    // every server-owned metadata key untouched. `properties` is deliberately
+    // not used: it bypasses the merge and is the mass-assignment surface the
+    // server's protected fields exist to guard.
+    await this.client.updateItem(item.item_id, {
+      content: this._renderContent(next),
+      metadata: this._toMetadata(next),
+    });
+
+    await this.appendEvent({
+      handle,
+      type: eventType,
+      at: nowIso(),
+      detail: { fields: Object.keys(patch) },
+    });
+
+    return next;
+  }
+
+  /** The only way discussion grows. Append-only in the API, not merely by
+   *  convention in the contract — a deliberation trail that can be rewritten is
+   *  not evidence. */
+  async appendDiscussion(handle, entry) {
+    return this._withLease('appendDiscussion', () => this._appendDiscussionLocked(handle, entry));
+  }
+
+  async _appendDiscussionLocked(handle, entry) {
+    const item = await this._resolveForWrite(handle);
+    if (!item) throw new FluidRecordNotFound(handle, this.name());
+    if (!entry?.text) throw new Error('fluid: a discussion entry requires text');
+
+    const record = this._fromItem(item);
+    if (!record) throw new FluidRecordNotFound(handle, this.name());
+
+    record.discussion.push({
+      at: entry.at ?? nowIso(),
+      text: entry.text,
+      author: entry.author ?? null,
+    });
+    record.updated_at = nowIso();
+    assertValid('record', record, 'record');
+
+    await this.client.updateItem(item.item_id, {
+      content: this._renderContent(record),
+      metadata: this._toMetadata(record),
+    });
+    await this.appendEvent({ handle, type: 'discussed', at: nowIso(), detail: {} });
+    return record;
+  }
+
+  /**
+   * Hard delete. NOT the lifecycle path — killing an idea is
+   * `updateRecord(handle, {status: 'killed'})`, which keeps the record and its
+   * reasoning.
+   *
+   * The discussion is carried into the append-only log before the record goes,
+   * so "append-only" is not true of every path except the one that erases it.
+   */
+  async deleteRecord(handle) {
+    return this._withLease('deleteRecord', () => this._deleteRecordLocked(handle));
+  }
+
+  async _deleteRecordLocked(handle) {
+    const item = await this._resolveForWrite(handle);
+    if (!item) throw new FluidRecordNotFound(handle, this.name());
+    const record = this._fromItem(item);
+
+    await this.appendEvent({
+      handle,
+      type: 'deleted',
+      at: nowIso(),
+      detail: {
+        kind: record?.kind,
+        title: record?.title,
+        discussion: record?.discussion ?? [],
+      },
+    });
+
+    await this.client.deleteItem(item.item_id);
+    return { ok: true };
+  }
+
+  // -------------------------------------------------------------------------
+  // Links
+  // -------------------------------------------------------------------------
+
+  async addLink(handle, link) {
+    return this._withLease('addLink', () => this._addLinkLocked(handle, link));
+  }
+
+  async _addLinkLocked(handle, link) {
+    assertLink(link);
+    const item = await this._resolveForWrite(handle);
+    if (!item) throw new FluidRecordNotFound(handle, this.name());
+    const record = this._fromItem(item);
+    if (!record) throw new FluidRecordNotFound(handle, this.name());
+
+    const exists = record.links.some((l) => l.type === link.type && l.target === link.target);
+    // Idempotent, and deliberately emits no event on a repeat: a `linked` event
+    // per repeat would inflate the lifecycle history a later semantic layer
+    // reads as signal.
+    if (!exists) {
+      record.links.push({ ...link });
+      record.updated_at = nowIso();
+      await this.client.updateItem(item.item_id, {
+        content: this._renderContent(record),
+        metadata: this._toMetadata(record),
+      });
+      await this.appendEvent({ handle, type: 'linked', at: nowIso(), detail: { ...link } });
+    }
+    return record;
+  }
+
+  async removeLink(handle, link) {
+    return this._withLease('removeLink', () => this._removeLinkLocked(handle, link));
+  }
+
+  async _removeLinkLocked(handle, link) {
+    const item = await this._resolveForWrite(handle);
+    if (!item) throw new FluidRecordNotFound(handle, this.name());
+    const record = this._fromItem(item);
+    if (!record) throw new FluidRecordNotFound(handle, this.name());
+
+    const before = record.links.length;
+    record.links = record.links.filter((l) => !(l.type === link.type && l.target === link.target));
+    if (record.links.length !== before) {
+      record.updated_at = nowIso();
+      await this.client.updateItem(item.item_id, {
+        content: this._renderContent(record),
+        metadata: this._toMetadata(record),
+      });
+    }
+    return record;
+  }
+
+  // -------------------------------------------------------------------------
+  // Recall (CAP.RECALL)
+  // -------------------------------------------------------------------------
+
+  /**
+   * Semantic recall over fluid records.
+   *
+   * One unfiltered, over-fetched query, then filtered client-side to this
+   * provider's namespace and to `RECALLABLE_KINDS`. Filtering the OUTPUT rather
+   * than scoping the REQUEST is deliberate: `memory_type` on the wire takes a
+   * single value, so covering three recallable kinds would need three queries,
+   * and scores from separate searches are not comparable — merging them would
+   * fabricate a ranking.
+   *
+   * DISCLOSED BOUND: because filtering happens after ranking, recall can return
+   * fewer than `limit` results even when more matching records exist, if enough
+   * higher-scoring non-recallable items fill the over-fetch. The alternative is
+   * unbounded fetching. Asserted in the tests so a future change is deliberate.
+   *
+   * DISCLOSED BOUND: an edited record keeps its original embedding, because
+   * PATCH does not reindex and no per-item reindex exists over HTTP
+   * (smart-memory-core#4). The record returned here is CURRENT — search hydrates
+   * from the live graph node — so what lags is why a record matched and where it
+   * ranked, not what you receive. The lexical channels still see current text.
+   *
+   * @param {string} query
+   * @param {{limit?: number}} [opts]
+   * @returns {Promise<import('./provider.js').RecallHit[]>}
+   */
+  async recall(query, opts = {}) {
+    this.require(CAP.RECALL);
+
+    const limit = normalizeRecallLimit(opts?.limit);
+    const topK = Math.min(Math.max(limit * OVERFETCH, MIN_FETCH), TOP_K_CAP);
+
+    const raw = await this.client.searchItems(query, { topK });
+
+    const hits = [];
+    const seen = new Map();
+    for (const result of raw?.results ?? []) {
+      const meta = result?.metadata ?? {};
+      // Namespace first: a shared workspace holds items that are not ours at
+      // all, and they must never reach a caller.
+      if (meta.fluid_ns !== RECORD_NS) continue;
+
+      const record = this._fromItem(result);
+      // `_fromItem` warns and returns null on an unreadable blob. One corrupt
+      // row must degrade to "that record is not in these results", never to a
+      // failed recall.
+      if (!record) continue;
+
+      // THE ENFORCEMENT. Holds with the deployment's embedding config unset,
+      // which is exactly the case the config dial cannot defend.
+      if (!RECALLABLE_KINDS.has(record.kind)) continue;
+
+      const score = typeof result?.score === 'number' ? result.score : null;
+      const createdAt = meta.created_at ?? '';
+      const itemId = String(result?.item_id ?? '');
+
+      // Collapse duplicate handles using D-FOH-4's full order — earliest
+      // server-stamped created_at, ties broken by item_id. Dropping the second
+      // key would leave same-instant duplicates resolving nondeterministically,
+      // and disagreeing with getRecord(). Both keys are already on the hit, so
+      // this costs no extra request.
+      const existing = seen.get(record.handle);
+      if (existing) {
+        const isEarlier = createdAt !== existing.createdAt
+          ? createdAt < existing.createdAt
+          : itemId < existing.itemId;
+        if (isEarlier) {
+          existing.hit.record = record;
+          existing.createdAt = createdAt;
+          existing.itemId = itemId;
+        }
+        // The surviving hit keeps its ORIGINAL rank and score: the duplicate is
+        // a storage artifact, and letting it reshuffle the ranking would leak
+        // that artifact into the answer.
+        continue;
+      }
+
+      const hit = { handle: record.handle, score, record };
+      seen.set(record.handle, { hit, createdAt, itemId });
+      hits.push(hit);
+    }
+
+    // Server order is the ranking. Filtering removes entries; it never reorders
+    // the survivors, and nothing here recomputes a score.
+    return hits.slice(0, limit);
+  }
+
+  // -------------------------------------------------------------------------
+  // Challenge (CAP.CHALLENGE)
+  // -------------------------------------------------------------------------
+
+  /**
+   * Same-kind contradiction detection (FOH-3).
+   *
+   * A record is challenged against OTHERS OF ITS OWN KIND, because the service's
+   * `memory_type` is an exact filter with no wildcard (`search.py:123`): one call
+   * covers one kind, and we send the record's exact wire type. Only `decision`
+   * and `idea` are challengeable — the endpoint runs its cascade directly with no
+   * `should_challenge` gate, so a non-assertional kind is refused, not fed to it.
+   *
+   * Two filters compose. The SERVER filters candidates to the requested type;
+   * THIS method then drops anything that is not one of ours (namespace) or is the
+   * challenged record itself, exactly as recall does — a shared workspace holds
+   * items that must never reach a caller. Because that second filter changes the
+   * conflict set, `hasConflicts` and `confidence` are recomputed from what
+   * survives, never the service's pre-filter aggregates (which would let a result
+   * claim conflicts it then shows none of).
+   *
+   * @param {string} handle
+   * @param {{useLlm?: boolean, timeoutMs?: number}} [opts]
+   * @returns {Promise<import('./provider.js').ChallengeResult>}
+   */
+  async challenge(handle, opts = {}) {
+    this.require(CAP.CHALLENGE);
+
+    const record = await this.getRecord(handle);
+    if (!record) throw new FluidRecordNotFound(handle, this.name());
+    if (!CHALLENGEABLE_KINDS.has(record.kind)) {
+      throw new FluidKindUnsupported(record.kind, this.name(), [...CHALLENGEABLE_KINDS]);
+    }
+
+    const assertion = this._renderContent(record);
+    const raw = await this.client.challenge(assertion, {
+      memoryType: wireTypeFor(record.kind), // exact fluid_<kind> — the crux
+      useLlm: opts.useLlm ?? true,
+      timeoutMs: opts.timeoutMs ?? CHALLENGE_TIMEOUT_MS,
+    });
+
+    const conflicts = [];
+    for (const c of raw?.conflicts ?? []) {
+      const item = await this.client.getItem(String(c?.existing_item_id ?? ''));
+      const meta = item?.metadata ?? {};
+      if (meta.fluid_ns !== RECORD_NS) continue;         // not ours → drop (D1a)
+      const conflictHandle = meta.handle;
+      if (!conflictHandle || conflictHandle === record.handle) continue; // self → drop (D1b)
+      conflicts.push({
+        handle: conflictHandle,
+        existingText: c.existing_fact ?? '',
+        conflictType: c.conflict_type ?? '',
+        confidence: typeof c.confidence === 'number' ? c.confidence : 0,
+        explanation: c.explanation ?? '',
+        suggestedResolution: c.suggested_resolution ?? '',
+      });
+    }
+    conflicts.sort((a, b) => b.confidence - a.confidence); // best-first
+
+    // D1c: aggregates from the RETAINED set. The service's confidence formula,
+    // recomputed over what the caller actually receives (challenger.py:256-262).
+    const hasConflicts = conflicts.length > 0;
+    const confidence = hasConflicts
+      ? Math.max(0, 1 - (conflicts.reduce((s, c) => s + c.confidence, 0) / conflicts.length) * 0.5)
+      : 1.0;
+
+    return { assertion, hasConflicts, confidence, conflicts };
+  }
+
+  // -------------------------------------------------------------------------
+  // Conviction (CAP.CONVICTION) — COMP-FOH FOH-4
+  // -------------------------------------------------------------------------
+
+  /** One `/confidence-history` envelope → the seam's ConvictionResult. Used by
+   *  the read AND by a landed resolution, so both return the same shape from
+   *  the same source. `challenged`/`lastChallengedAt` are DERIVED — the
+   *  envelope is the read path's single call, no `getItem` behind it. */
+  _toConvictionResult(handle, raw) {
+    const history = raw.history.map((e) => ({
+      timestamp: e.timestamp,
+      oldConfidence: e.old_confidence,
+      newConfidence: e.new_confidence,
+      decayFactor: e.decay_factor,
+      reason: e.reason,
+      ...(e.conflicting_fact !== undefined ? { conflictingFact: e.conflicting_fact } : {}),
+    }));
+    return {
+      handle,
+      confidence: raw.current_confidence,
+      challenged: raw.challenge_count > 0,
+      challengeCount: raw.challenge_count,
+      lastChallengedAt: history.length ? history[history.length - 1].timestamp : null,
+      history,
+    };
+  }
+
+  /**
+   * Belief-strength read. A never-resolved record honestly returns
+   * `{confidence: 1.0, challenged: false, history: []}` — nothing has moved it.
+   *
+   * @param {string} handle
+   * @returns {Promise<import('./provider.js').ConvictionResult>}
+   */
+  async conviction(handle) {
+    this.require(CAP.CONVICTION);
+    if (!HANDLE_RE.test(handle ?? '')) throw new FluidRecordNotFound(handle, this.name());
+    const item = await this._resolveOne(handle);
+    if (!item) throw new FluidRecordNotFound(handle, this.name());
+    let raw;
+    try {
+      raw = await this.client.confidenceHistory(String(item.item_id));
+    } catch (err) {
+      // The item vanished between resolution and read (a race with a delete).
+      if (err?.status === 404) throw new FluidRecordNotFound(handle, this.name());
+      throw err;
+    }
+    return this._toConvictionResult(handle, raw);
+  }
+
+  /**
+   * Gated resolution: decay `targetHandle`'s confidence because `sourceHandle`
+   * supersedes it. See the base class JSDoc for the caller-facing contract
+   * (explicit strategy, near-irreversible, only NoOp retryable, trust
+   * boundary); what lives HERE is the safety machinery:
+   *
+   *   - Every authorization invariant is enforced at this seam — the server's
+   *     `/resolve` does no re-detection and would decay whatever id it is
+   *     handed. Handles resolve via the READ-ONLY `_resolveOne` (the write
+   *     resolver repairs duplicates as a side effect, which must not run
+   *     before authorization).
+   *   - The contradicting fact is derived from the SOURCE record's own stored
+   *     content, exactly as `challenge()` derives its assertion. No caller
+   *     text reaches the wire.
+   *   - The whole sequence runs inside the workspace lease, which serializes
+   *     Compose-side mutations. It does NOT serialize direct SmartMemory
+   *     callers — which is why the postcondition has a 'conflict' verdict.
+   *   - The `/resolve` response is never trusted. The outcome is classified
+   *     from confidence-history reads bracketing the call
+   *     ({@link classifyResolution}); an ambiguous transport failure (abort,
+   *     malformed 2xx, gateway 5xx — none of which prove the server's
+   *     synchronous mutation stopped) goes through a bounded read-only
+   *     reconciliation poll and, failing a positive match, surfaces as
+   *     Indeterminate rather than inviting a double-decaying retry.
+   *
+   * @param {string} sourceHandle
+   * @param {string} targetHandle
+   * @param {{strategy: string, timeoutMs?: number}} [opts]
+   * @returns {Promise<import('./provider.js').ConvictionResult>}
+   */
+  async resolveConflict(sourceHandle, targetHandle, opts = {}) {
+    this.require(CAP.CONVICTION);
+    const strategy = opts.strategy;
+    if (!CONVICTION_STRATEGIES.includes(strategy)) {
+      throw new FluidInvalidStrategy(strategy, [...CONVICTION_STRATEGIES]);
+    }
+
+    // The lease serializes the DECAY. The best-effort contradiction link (FOH-5)
+    // is deliberately done AFTER the lease releases: the edge write is idempotent
+    // (MERGE by identity) and touches no confidence, so holding the mutation lease
+    // across its network retries would only serialize unrelated callers for no
+    // safety gain. The closure therefore returns {result, link}: `link` is set on
+    // EITHER landed exit (clean or reconciled), null on every throw.
+    const { result, link } = await this._withLease('resolveConflict', async () => {
+      const sourceItem = await this._resolveOne(sourceHandle);
+      if (!sourceItem) throw new FluidRecordNotFound(sourceHandle, this.name());
+      const targetItem = await this._resolveOne(targetHandle);
+      if (!targetItem) throw new FluidRecordNotFound(targetHandle, this.name());
+      // _fromItem is null for a corrupt blob even on a namespaced item — refuse
+      // typed, before touching `.kind`.
+      const sourceRec = this._fromItem(sourceItem);
+      if (!sourceRec) throw new FluidRecordNotFound(sourceHandle, this.name());
+      const targetRec = this._fromItem(targetItem);
+      if (!targetRec) throw new FluidRecordNotFound(targetHandle, this.name());
+      if (!CHALLENGEABLE_KINDS.has(sourceRec.kind)) {
+        throw new FluidKindUnsupported(sourceRec.kind, this.name(), [...CHALLENGEABLE_KINDS]);
+      }
+      if (targetRec.kind !== sourceRec.kind) {
+        throw new FluidKindUnsupported(targetRec.kind, this.name(), [sourceRec.kind]);
+      }
+      if (String(sourceItem.item_id) === String(targetItem.item_id)) {
+        throw new FluidInvalidTarget(targetHandle, 'self-target');
+      }
+
+      const newFact = this._renderContent(sourceRec);
+      const expectedFact = codePointSlice(newFact, 200);
+      const targetId = String(targetItem.item_id);
+      let pre;
+      try {
+        pre = await this.client.confidenceHistory(targetId);
+      } catch (err) {
+        // The target vanished between resolution and the pre-read (a direct
+        // caller's delete) — the same race conviction() maps. Nothing mutated.
+        if (err?.status === 404) throw new FluidRecordNotFound(targetHandle, this.name());
+        throw err;
+      }
+
+      try {
+        await this.client.resolveConflict({
+          existingItemId: targetId,
+          newFact,
+          strategy,
+          timeoutMs: opts.timeoutMs ?? RESOLVE_TIMEOUT_MS,
+        });
+      } catch (err) {
+        const status = typeof err?.status === 'number' ? err.status : 0;
+        const ambiguous = status === 0 || err?.kind === 'malformed-response' || status >= 500;
+        if (!ambiguous) {
+          // A real 4xx: rejected before the mutation. Nothing to reconcile.
+          if (status === 404) throw new FluidRecordNotFound(targetHandle, this.name());
+          throw err;
+        }
+        // Returns only on a landed decay (else throws) — so a value here means
+        // the decay landed and the contradiction link is owed.
+        return {
+          result: await this._reconcileResolution(targetHandle, targetId, pre, expectedFact),
+          link: { sourceId: String(sourceItem.item_id), targetId },
+        };
+      }
+
+      // Clean response: the handler finished, so one immediate re-read is
+      // authoritative. If even that read fails, fall back to the poll — a
+      // reconciliation-read failure must not surface as its own error.
+      let post;
+      try {
+        post = await this.client.confidenceHistory(targetId);
+      } catch {
+        // Returns only on a landed decay (else throws) — so a value here means
+        // the decay landed and the contradiction link is owed.
+        return {
+          result: await this._reconcileResolution(targetHandle, targetId, pre, expectedFact),
+          link: { sourceId: String(sourceItem.item_id), targetId },
+        };
+      }
+      const verdict = classifyResolution(pre, post, expectedFact);
+      if (verdict === 'landed') {
+        return {
+          result: this._toConvictionResult(targetHandle, post),
+          link: { sourceId: String(sourceItem.item_id), targetId },
+        };
+      }
+      if (verdict === 'no-op') throw new FluidResolutionNoOp(targetHandle);
+      if (verdict === 'conflict') throw new FluidResolutionConflict(targetHandle);
+      throw new FluidResolutionIndeterminate(targetHandle);
+    });
+
+    // Post-success epilogue (FOH-5), OUTSIDE the lease. The decay is already
+    // durable and authoritative; the CONTRADICTS edge is a best-effort enhancement
+    // so `contradictions()` can resolve this decay back to its source. A failed
+    // link never fails the resolution — it degrades `contradictions()` to a lower
+    // bound, which its contract already permits.
+    if (link) await this._linkContradiction(link.sourceId, link.targetId);
+    return result;
+  }
+
+  /**
+   * Best-effort durable link: source CONTRADICTS target, written after the decay
+   * landed. Bounded retry, because the edge write is safe to repeat (idempotent
+   * MERGE by identity, no confidence touched) — unlike the resolve it follows.
+   * On persistent failure it WARNS and returns; it never throws, because the
+   * resolution it belongs to already succeeded and the caller must not be handed
+   * a failure for a durability detail.
+   */
+  async _linkContradiction(sourceId, targetId) {
+    for (let attempt = 0; attempt < LINK_RETRY_ATTEMPTS; attempt += 1) {
+      try {
+        await this.client.addEdge({
+          sourceId,
+          targetId,
+          relationType: CONTRADICTS_EDGE,
+          properties: { origin: 'fluid:resolveConflict' },
+        });
+        return;
+      } catch (err) {
+        if (attempt + 1 < LINK_RETRY_ATTEMPTS) {
+          await sleep(LINK_RETRY_DELAY_MS);
+          continue;
+        }
+        process.emitWarning(
+          `fluid(smartmemory): decayed ${targetId} but could not link its contradiction from `
+          + `${sourceId} after ${LINK_RETRY_ATTEMPTS} attempts (${err?.message ?? err}); `
+          + 'contradictions() will under-report this pair until re-linked',
+        );
+      }
+    }
+  }
+
+  /**
+   * Bounded read-only poll after an ambiguous outcome. 'no-op' is deliberately
+   * NOT trusted here: an aborted request's mutation may still be in flight
+   * server-side, so "unchanged" only means "not yet" — the round-3 race. Only
+   * a positive 'landed' (ours, verified) or 'conflict' (someone else moved it)
+   * ends the poll early; anything less resolves to Indeterminate.
+   */
+  async _reconcileResolution(targetHandle, targetId, pre, expectedFact) {
+    for (let attempt = 0; attempt < RESOLVE_RECONCILE_ATTEMPTS; attempt += 1) {
+      await sleep(RESOLVE_RECONCILE_DELAY_MS);
+      let post;
+      try {
+        post = await this.client.confidenceHistory(targetId);
+      } catch {
+        continue; // a failed read is a failed attempt, never its own error
+      }
+      const verdict = classifyResolution(pre, post, expectedFact);
+      if (verdict === 'landed') return this._toConvictionResult(targetHandle, post);
+      if (verdict === 'conflict') throw new FluidResolutionConflict(targetHandle);
+    }
+    throw new FluidResolutionIndeterminate(targetHandle);
+  }
+
+  // -------------------------------------------------------------------------
+  // Contradiction (CAP.CONTRADICTION) — COMP-FOH FOH-5
+  // -------------------------------------------------------------------------
+
+  /**
+   * The records that contradict `handle`. See the base-class JSDoc for the
+   * caller-facing contract (best-effort lower bound, canonicalized, handle is
+   * authority); the read machinery lives here:
+   *
+   *   - Direction: a resolution writes `source CONTRADICTS target`, so the
+   *     records contradicting `handle` are its INCOMING CONTRADICTS neighbours.
+   *   - The target-deletion race (handle resolves, then vanishes before the
+   *     neighbours read) maps to FluidRecordNotFound, exactly as {@link conviction}.
+   *   - Each hit is canonicalized: a neighbour is emitted only when it IS the
+   *     canonical item for its own handle, so `hit.record` can never disagree
+   *     with a later `getRecord(hit.handle)`. Later duplicates are skipped.
+   *   - Per-neighbour failure is typed: a 404 is a deletion race (skip via
+   *     `getItem`'s null), a corrupt/non-fluid item is skipped (`_fromItem`
+   *     already warned), and ANY other fetch failure PROPAGATES — a silent
+   *     partial is worse than an error for a read that is already a lower bound.
+   *
+   * @param {string} handle
+   * @returns {Promise<import('./provider.js').ContradictionHit[]>}
+   */
+  async contradictions(handle) {
+    this.require(CAP.CONTRADICTION);
+    if (!HANDLE_RE.test(handle ?? '')) throw new FluidRecordNotFound(handle, this.name());
+    const item = await this._resolveOne(handle);
+    if (!item) throw new FluidRecordNotFound(handle, this.name());
+
+    let neighbours;
+    try {
+      neighbours = await this.client.neighbors(String(item.item_id));
+    } catch (err) {
+      // Target deleted between resolve and read — the race conviction() maps.
+      if (err?.status === 404) throw new FluidRecordNotFound(handle, this.name());
+      throw err;
+    }
+
+    const hits = [];
+    const seen = new Set();
+    for (const n of neighbours) {
+      if (n?.link_type !== CONTRADICTS_EDGE || n?.direction !== 'incoming') continue;
+      const sourceId = String(n.item_id);
+
+      // `getItem` returns null on 404 (deletion race → skip) and THROWS on any
+      // other failure, which is exactly the "fail rather than under-report"
+      // semantics — so it is deliberately not caught here.
+      const raw = await this.client.getItem(sourceId);
+      if (raw === null) continue;
+
+      const record = this._fromItem(raw);
+      if (!record) continue; // corrupt / non-fluid: _fromItem already warned
+
+      // `_fromItem` guards a bad-JSON blob but NOT a parseable-but-schema-invalid
+      // one (e.g. `{}` or a handle with no kind). Such a blob would otherwise be
+      // emitted as a malformed hit, or — with no handle — crash `_resolveOne`'s
+      // paired metadata filter below. A shared workspace can hold one, so skip it
+      // exactly like a corrupt row rather than letting it break the whole read.
+      if (!getFluidValidator().validate('record', record).valid) {
+        process.emitWarning(
+          `fluid(smartmemory): CONTRADICTS neighbour ${sourceId} holds a schema-invalid `
+          + 'record blob; skipped',
+        );
+        continue;
+      }
+
+      // Canonical-handle rule: emit only when this neighbour is the item
+      // `getRecord(record.handle)` resolves to (the earliest). Otherwise the
+      // returned record would disagree with the authority the handle names.
+      const canonical = await this._resolveOne(record.handle);
+      if (!canonical || String(canonical.item_id) !== sourceId) continue;
+      if (seen.has(record.handle)) continue;
+      seen.add(record.handle);
+
+      hits.push({ handle: record.handle, kind: record.kind, record });
+    }
+    return hits;
+  }
+
+  // -------------------------------------------------------------------------
+  // Lifecycle events
+  // -------------------------------------------------------------------------
+
+  /**
+   * Events are SEPARATE items on purpose.
+   *
+   * Stored inside a record's metadata they would be destroyed with the record,
+   * and handle retirement is precisely what must outlive deletion. They are
+   * written and never updated or deleted.
+   */
+  async appendEvent(event) {
+    const full = { at: nowIso(), ...event };
+    // The log is append-only, so a malformed entry is permanent. Validate
+    // before it lands rather than on a later read with no caller to blame.
+    assertValid('lifecycle_event', full, 'lifecycle event');
+    await this.client.createItem({
+      content: `${full.type} ${full.handle}`,
+      memoryType: EVENT_WIRE_TYPE,
+      metadata: {
+        fluid_ns: EVENT_NS,
+        handle: full.handle,
+        fluid_event_json: JSON.stringify(full),
+      },
+    });
+    return full;
+  }
+
+  async readEvents(handle) {
+    const items = await this._listAllItems({
+      metadataKey: 'fluid_ns', metadataValue: EVENT_NS,
+    });
+    const events = [];
+    for (const item of items) {
+      const meta = item?.metadata ?? {};
+      if (meta.fluid_ns !== EVENT_NS) continue;
+      if (handle && meta.handle !== handle) continue;
+      try {
+        events.push(JSON.parse(meta.fluid_event_json));
+      } catch {
+        process.emitWarning(
+          `fluid(smartmemory): event item ${item?.item_id} has an unreadable payload and was skipped`,
+        );
+      }
+    }
+    // Chronological, matching the floor's append-only file order. Handle
+    // allocation reads this, so a stable order keeps allocation reproducible.
+    return events.sort((a, b) => String(a?.at ?? '').localeCompare(String(b?.at ?? '')));
+  }
+}