bugfix indexer field

internal/api/handlers_identity.go

@@ -50,6 +50,26 @@ type identityImportResponse struct {
 	UID       string `json:"uid"`
 	Curve     string `json:"curve"`
 	CreatedAt int64  `json:"createdAt"`
+	// Created distinguishes "this key was new here" (201) from "this key was
+	// already in the keystore" (200). Import is the sign-in primitive for
+	// clients: presenting the secret key is the proof of ownership, so a
+	// re-import must succeed rather than conflict. Clients use the flag only
+	// to word the confirmation ("new profile" vs "welcome back").
+	Created bool `json:"created"`
+}
+
+// identityRevealRequest asks for an existing account's private key back. The
+// key was shown once at generate; without a way to see it again, a client with
+// a real sign-out has no recovery path for a user who never saved it.
+type identityRevealRequest struct {
+	Confirm bool   `json:"confirm"`
+	UID     string `json:"uid"`
+}
+
+type identityRevealResponse struct {
+	UID           string `json:"uid"`
+	Curve         string `json:"curve"`
+	PrivateKeyHex string `json:"privateKeyHex"`
 }
 
 type identityDeleteRequest struct {
@@ -172,9 +192,17 @@ func (s *Server) handleIdentityImport(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	dir := s.config.IdentityAccountsDir()
-	if identity.ExistsByUID(dir, id.UID) {
-		writeErrorSlug(w, http.StatusConflict, "identity_exists",
-			"an account with this key already exists", false)
+	// Idempotent: importing a key that is already in the keystore is how a
+	// second device signs in to an existing account. Answer 200 with the
+	// stored account (preserving its original createdAt) instead of 409 —
+	// the caller proved ownership by presenting the key.
+	if existing, err := identity.LoadByUID(dir, id.UID); err == nil && existing != nil {
+		writeJSON(w, http.StatusOK, identityImportResponse{
+			UID:       existing.UID,
+			Curve:     existing.Curve,
+			CreatedAt: existing.CreatedAt,
+			Created:   false,
+		})
 		return
 	}
 	if err := identity.SaveAccount(dir, id); err != nil {
@@ -185,6 +213,35 @@ func (s *Server) handleIdentityImport(w http.ResponseWriter, r *http.Request) {
 		UID:       id.UID,
 		Curve:     id.Curve,
 		CreatedAt: id.CreatedAt,
+		Created:   true,
+	})
+}
+
+// handleIdentityReveal returns an existing account's private key. Deliberately
+// POST + explicit confirm (never a bare GET) so it cannot be triggered by a
+// link or an image tag, and so it reads as the deliberate act it is. The
+// endpoint is auth-gated at the perimeter like the rest of /api/identity/*.
+func (s *Server) handleIdentityReveal(w http.ResponseWriter, r *http.Request) {
+	var body identityRevealRequest
+	_ = json.NewDecoder(r.Body).Decode(&body)
+	if !body.Confirm && r.URL.Query().Get("confirm") != "true" {
+		writeErrorSlug(w, http.StatusBadRequest, "confirm_required",
+			"revealing a private key requires {\"confirm\": true} or ?confirm=true", false)
+		return
+	}
+	uid := strings.TrimSpace(body.UID)
+	if uid == "" {
+		uid = strings.TrimSpace(r.URL.Query().Get("uid"))
+	}
+	id, status, slug, msg := s.resolveAccount(uid)
+	if status != 0 {
+		writeErrorSlug(w, status, slug, msg, false)
+		return
+	}
+	writeJSON(w, http.StatusOK, identityRevealResponse{
+		UID:           id.UID,
+		Curve:         id.Curve,
+		PrivateKeyHex: id.PrivateKeyHex,
 	})
 }
 

internal/api/server.go

@@ -172,6 +172,7 @@ func (s *Server) setupRoutes() {
 	s.router.HandleFunc("/api/identity/accounts", s.handleIdentityAccounts).Methods("GET")
 	s.router.HandleFunc("/api/identity/generate", s.handleIdentityGenerate).Methods("POST")
 	s.router.HandleFunc("/api/identity/import", s.handleIdentityImport).Methods("POST")
+	s.router.HandleFunc("/api/identity/reveal", s.handleIdentityReveal).Methods("POST")
 	s.router.HandleFunc("/api/identity", s.handleIdentityDelete).Methods("DELETE")
 	s.router.HandleFunc("/api/identity/sign", s.handleIdentitySign).Methods("POST")
 	s.router.HandleFunc("/api/identity/aead-key", s.handleIdentityAEADKey).Methods("GET")
@@ -350,6 +351,21 @@ func (s *Server) Start() error {
 
 			// Republish events for all existing metadata
 			go func() {
+				// Build any missing per-record field indexes first, in ONE
+				// keyspace pass. The republish below reads every record, and
+				// an un-indexed record read costs a full-keyspace SCAN — on a
+				// real store that walk warmed at ~1.3 records/s (a ~2.6h tail,
+				// paid again on every restart). Warming first turns the whole
+				// republish into index-served reads. Idempotent, so a
+				// fully-indexed store just pays the single scan.
+				// See storage/field_index.go.
+				started := time.Now()
+				if n, err := s.storage.WarmFieldIndexes(); err != nil {
+					log.Printf("[API] Warning: field index warm failed: %v (reads fall back to SCAN)", err)
+				} else if n > 0 {
+					log.Printf("[API] Warmed %d field indexes in %s", n, time.Since(started).Round(time.Millisecond))
+				}
+
 				count, err := s.RepublishMetadata()
 				if err != nil {
 					log.Printf("[API] Warning: failed to republish metadata on startup: %v", err)

internal/api/udl.go

@@ -27,6 +27,14 @@ type UDLPutRequest struct {
 	Ts      int64            `json:"ts"`
 	Record  string           `json:"record"` // base64 CBOR, opaque + signed
 	Value   *json.RawMessage `json:"value,omitempty"`
+	// Tombstone marks this write as a deletion of (uid, cid, key). The record
+	// itself is opaque here — meta-core never decodes the CBOR, and a private-
+	// tier value and a tombstone are equally "no plaintext value" — so the
+	// writer has to say. Without it the cid is never removed from
+	// udl:idx:user:<uid>:key:<key>, which then grows for the life of the
+	// account and is returned in full on every My List / Continue Watching
+	// read. Omitted by older writers, which keeps the previous behaviour.
+	Tombstone bool `json:"tombstone,omitempty"`
 }
 
 // UDLPutResponse is the response of PUT /api/udl/record.
@@ -104,7 +112,7 @@ func (s *Server) handleUDLRecordPut(w http.ResponseWriter, r *http.Request) {
 		hasValue = true
 	}
 
-	accepted, err := s.storage.UDLUpsertIfNewer(req.Uid, req.Cid, req.Key, req.Version, req.Ts, req.Record, publicValue, hasValue)
+	accepted, err := s.storage.UDLUpsertIfNewer(req.Uid, req.Cid, req.Key, req.Version, req.Ts, req.Record, publicValue, hasValue, req.Tombstone)
 	if err != nil {
 		writeError(w, http.StatusInternalServerError, err.Error())
 		return

internal/cid/cid-rank-vectors.json

@@ -19,9 +19,10 @@
   "ladder": {
     "40": "dag-pb (IPFS block)",
     "30": "sha2-256 / sha3-256 / sha3-384 - real content digests",
+    "25": "nzb-posting - content-derived (Message-ID set digest), but not itself fetchable",
     "20": "BitTorrent infohashes - swarm interop",
     "10": "midhash256 - the record address",
-    "5": "opaque locators (nzb-release, url) - never outrank a real digest",
+    "5": "opaque locators (nzb-release, url, card) - never outrank a real digest",
     "0": "weak/unknown digests, unparseable input"
   },
   "tieBreak": "lexicographically smaller cid wins",
@@ -56,6 +57,12 @@
       "rank": 30,
       "note": "REAL meta-hash fixture"
     },
+    {
+      "name": "nzb-posting",
+      "cid": "bagbsaeraaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
+      "rank": 25,
+      "note": "codec 0x1003 + REAL sha2-256 multihash (0x12) over the normalised, sorted Message-ID set - NOT a locator, so it sits ABOVE tier 5, but it addresses no fetchable block (bitswap can never satisfy a want for it - the digest is over an id set, not over any block's bytes), so it stays BELOW a real content digest (30): once meta-share materialises the media and re-seeds it, that genuine sha2-256 cid must win the canonical election, not this one. Rank MUST match on codec BEFORE falling through to the mh switch, or this scores 30 (same trap as nzb-release/url, in reverse: those need a codec check because their mh is identity 0x00; this needs one because its mh is REAL sha2-256 and would otherwise tie with tier 30)."
+    },
     {
       "name": "btih-v2-codec-eq-mh",
       "cid": "bag3sdnzbeaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
@@ -103,6 +110,12 @@
       "rank": 5,
       "note": "codec 0x1006 + IDENTITY multihash 0x00. Same tier/reasoning as nzb-release."
     },
+    {
+      "name": "card-locator",
+      "cid": "bagdsaaanar2g2zdcor3duojvgq3ts",
+      "rank": 5,
+      "note": "codec 0x1007 + IDENTITY multihash 0x00, digest = varint(len(source)) || source || id — here (\"tmdb\",\"tv:95479\"). Same codec-slot rule as nzb-release/url: match on the CODEC, never the mh. The limit case of the family — the only CID in MetaMesh with NO bytes behind it anywhere, ever: it addresses a *work* (a series, a film), not a file, so the byte path 404s by design and bitswap never offers it (seeding is gated on sha2-256). Tier 5 is load-bearing for the planned cross-source merge: one card record may carry BOTH a TMDB and a MyAnimeList locator, and if a real card body is ever added (see card-tier-search.md §4.3) its sha2-256 must win the canonical election over both locators without any addressing change. Ranking it 0 instead would still lose to a digest, but would put the implementations back out of agreement — which is the exact failure this fixture exists to prevent."
+    },
     {
       "name": "sha1",
       "cid": "baeircff3eqvxxaqqot2lt242stkdtk3gq4c2vra",

internal/cid/rank.go

@@ -49,12 +49,33 @@ const (
 	// these — matching on mh would see 0x00 and rank them 0.
 	CodeNzbRelease = 0x1005 // custom — self-describing Newznab release locator
 	CodeURL        = 0x1006 // custom — identity-multihash CID wrapping an http(s) URL
+	// CodeCard is the limit case of the family: the only CID with NO bytes
+	// behind it anywhere, ever. It addresses a *work* (a series, a film) via
+	// varint(len(source))‖source‖id — ("tmdb","tv:95479") — so `/api/file/{cid}`
+	// correctly 404s and bitswap never offers it. Ranking it here (rather than
+	// letting it fall to RankUnknown) is what makes the planned "one card, two
+	// CIDs" cross-source merge safe: a TMDB and a MyAnimeList locator can sit on
+	// one record, and a real digest added later still wins the election.
+	CodeCard = 0x1007 // custom — identity-multihash CID wrapping a card (work) id
+
+	// CodeNzbPosting is the INVERSE trap of the two above: its multihash is a
+	// REAL sha2-256 digest (over the normalised, sorted article Message-ID
+	// set of a Usenet posting we scanned ourselves — never a locator, no
+	// indexer host embedded), so it would silently fall into RankDigest via
+	// the mh switch unless matched on the codec FIRST. It is content-derived
+	// (so it outranks the opaque locators) but addresses no fetchable block —
+	// the digest is over an id set, not over any block's bytes, so it can
+	// never be bitswap-fetched under itself — so it must stay BELOW a real
+	// digest (once the media is materialised and re-seeded, that genuine
+	// sha2-256 cid has to win the canonical election, not this one).
+	CodeNzbPosting = 0x1003 // custom — self-scanned Usenet posting identity
 )
 
 // Rank tiers, named so the ladder reads as a ladder.
 const (
 	RankIPFS       = 40 // dag-pb — broadest network reach
 	RankDigest     = 30 // sha2-256 / sha3-* — real content digests
+	RankNzbPosting = 25 // content-derived, but the digest addresses no fetchable block
 	RankBtih       = 20 // BitTorrent infohashes — swarm interop
 	RankMidhash    = 10 // the record address — internally meaningful, externally invisible
 	RankLocator    = 5  // opaque locators — never outrank a real digest
@@ -112,11 +133,16 @@ func Decode(cidStr string) (codec uint64, mhCode uint64, err error) {
 //   - dag-pb (chunked IPFS file) wins: best retrieval over the IPFS network.
 //   - sha2-256 / sha3 digests follow: standard content hashes (a raw IPFS
 //     block of the file IS its sha2-256, so they share this tier).
+//   - nzb-posting is content-derived (a digest over a Usenet posting's article
+//     Message-ID set) so it outranks the opaque locators below, but the
+//     digest addresses no fetchable block, so it stays below a real content
+//     digest — once the media is materialised and re-seeded under a genuine
+//     sha2-256 cid, that cid must win the election, not this one.
 //   - BitTorrent infohashes (btih v2, per-file v1/v2) outrank midhash256 for
 //     swarm interop, not strength.
 //   - midhash256 is the floor among real *digests*: internally meaningful,
 //     externally invisible.
-//   - opaque locators (nzb-release, url) sit BELOW midhash. They are not
+//   - opaque locators (nzb-release, url, card) sit BELOW midhash. They are not
 //     content-derived — two URLs to the same bytes yield two different locator
 //     CIDs — so a locator must never displace a real digest as the identity a
 //     record is advertised under. Tier 5 means a locator wins the election only
@@ -139,9 +165,15 @@ func Rank(cidStr string) int {
 	}
 	// Opaque locators. Codec-only: their multihash is identity (0x00), so
 	// there is nothing to match on the mh side.
-	if codec == CodeNzbRelease || codec == CodeURL {
+	if codec == CodeNzbRelease || codec == CodeURL || codec == CodeCard {
 		return RankLocator
 	}
+	// nzb-posting: codec-only, checked BEFORE the mh switch below — its
+	// multihash is a real sha2-256 and would otherwise match CodeSha256 there
+	// and rank 30, tying with genuine content digests.
+	if codec == CodeNzbPosting {
+		return RankNzbPosting
+	}
 	switch mh {
 	case CodeSha256, CodeSha3_256, CodeSha3_384:
 		return RankDigest

internal/cid/rank_test.go

@@ -150,12 +150,31 @@ func TestRank_Tiers(t *testing.T) {
 func TestRank_StrictOrdering(t *testing.T) {
 	ipfs := Rank(buildCID(CodecDagPB, CodeSha256, 32))
 	sha := Rank(cidSha256)
+	posting := Rank(buildCID(CodeNzbPosting, CodeSha256, 32))
 	btih := Rank(buildCID(CodeBtihV2, CodeBtihV2, 32))
 	mid := Rank(buildCID(CodeMidhash256, CodeMidhash256, 32))
 	unknown := Rank(cidMd5)
-	if !(ipfs > sha && sha > btih && btih > mid && mid > unknown) {
-		t.Fatalf("rank ordering broken: ipfs=%d sha=%d btih=%d mid=%d unknown=%d",
-			ipfs, sha, btih, mid, unknown)
+	if !(ipfs > sha && sha > posting && posting > btih && btih > mid && mid > unknown) {
+		t.Fatalf("rank ordering broken: ipfs=%d sha=%d posting=%d btih=%d mid=%d unknown=%d",
+			ipfs, sha, posting, btih, mid, unknown)
+	}
+}
+
+// TestRank_NzbPostingMatchesCodecNotMultihash pins the inverse trap:
+// nzb-posting's multihash is a REAL sha2-256 (unlike nzb-release/url, whose mh
+// is identity 0x00), so Rank must match on the codec BEFORE falling into the
+// mh switch — otherwise this ties with a genuine content digest at 30 instead
+// of sitting at 25.
+func TestRank_NzbPostingMatchesCodecNotMultihash(t *testing.T) {
+	posting := buildCID(CodeNzbPosting, CodeSha256, 32)
+	if got := Rank(posting); got != RankNzbPosting {
+		t.Errorf("Rank(nzb-posting) = %d, want %d (RankNzbPosting) — codec check must precede the mh switch", got, RankNzbPosting)
+	}
+	// A real sha2-256 with the SAME digest bytes must still rank 30 — the
+	// codec, not the digest, is what distinguishes the two.
+	realDigest := buildCID(CodecRaw, CodeSha256, 32)
+	if got := Rank(realDigest); got != RankDigest {
+		t.Errorf("Rank(sha2-256) regressed to %d, want %d", got, RankDigest)
 	}
 }
 

internal/events/meta_publisher.go

@@ -9,6 +9,8 @@ import (
 	"time"
 
 	"github.com/redis/go-redis/v9"
+
+	"github.com/metazla/meta-core/internal/storage"
 )
 
 const (
@@ -126,6 +128,14 @@ func (p *MetaPublisher) publishEvent(channel, operation string) {
 	if strings.Contains(key, "__index__") || !strings.HasPrefix(key, "file:") {
 		return
 	}
+	// Skip the per-record field index. It is bookkeeping (see
+	// storage.FieldsField), not a metadata field — publishing it would make
+	// every downstream consumer that parses `file:<id>/<field>` off this
+	// stream (meta-search, meta-dup, meta-fuse, meta-sort) see a phantom
+	// field change on a key they can never read.
+	if strings.HasSuffix(key, "/"+storage.FieldsField) {
+		return
+	}
 
 	// Build event fields
 	fields := map[string]interface{}{

internal/storage/cid_resolution.go

@@ -83,6 +83,7 @@ func (c *Client) Mint(filePath string, size, mtimeNano int64) (string, error) {
 	pipe.Set(ctx, prefix+"sizeByte", strconv.FormatInt(size, 10), 0)
 	pipe.Set(ctx, prefix+"mtimeNano", strconv.FormatInt(mtimeNano, 10), 0)
 	pipe.SAdd(ctx, c.buildIndexKey(), uuid)
+	c.addFieldsToPipe(ctx, pipe, uuid, []string{"filePath", "sizeByte", "mtimeNano"})
 	if _, err := pipe.Exec(ctx); err != nil {
 		return "", fmt.Errorf("mint pipeline: %w", err)
 	}
@@ -116,7 +117,13 @@ func (c *Client) addAliasLocked(ctx context.Context, uuid, cidStr string) error
 		return fmt.Errorf("uuid and cid are required")
 	}
 
-	keysetKey := c.buildKeyPrefix(uuid) + CIDsKeyPrefix + cidStr
+	keysetField := CIDsKeyPrefix + cidStr
+	keysetKey := c.buildKeyPrefix(uuid) + keysetField
+
+	// Every branch below writes the key-set member, so index the field name
+	// once here — including the two guard branches that skip the alias itself.
+	// Missing it would leave `cids/<cid>` invisible to a record read.
+	defer c.addFieldLocked(ctx, uuid, keysetField)
 
 	// Self-pointing alias guard. If the caller is trying to register
 	// cid:<cid> → <cid> (i.e. the root IS the cid value), that's the
@@ -233,26 +240,13 @@ func (c *Client) AddDuplicatePath(uuid, path string) (bool, error) {
 	return added > 0, nil
 }
 
-// cidsForRootLocked scans the bare-CID key-set for a root and returns its
-// members (bare CIDs, with the cids/ prefix stripped). Caller must hold c.mu.
+// cidsForRootLocked returns a root's bare-CID key-set members (with the cids/
+// prefix stripped). Caller must hold c.mu.
+//
+// Reads the field index and filters by prefix; only an un-indexed root falls
+// back to a SCAN, and then over the narrower cids/ glob. See field_index.go.
 func (c *Client) cidsForRootLocked(ctx context.Context, uuid string) ([]string, error) {
-	prefix := c.buildKeyPrefix(uuid) + CIDsKeyPrefix
-	var out []string
-	var cursor uint64
-	for {
-		keys, next, err := c.client.Scan(ctx, cursor, prefix+"*", 1000).Result()
-		if err != nil {
-			return nil, fmt.Errorf("scan cids: %w", err)
-		}
-		for _, k := range keys {
-			out = append(out, strings.TrimPrefix(k, prefix))
-		}
-		cursor = next
-		if cursor == 0 {
-			break
-		}
-	}
-	return out, nil
+	return c.fieldsWithPrefixLocked(ctx, uuid, CIDsKeyPrefix)
 }
 
 // DeleteRoot removes everything keyed off uuid: every file:<uuid>/* property

internal/storage/client.go

@@ -109,8 +109,12 @@ func (c *Client) buildIndexKey() string {
 	return c.buildKey("file:__index__")
 }
 
-// GetMetadataFlat retrieves all metadata for a file as a flat map
-// Uses flat keys: SCAN file:{hashId}/* then MGET
+// GetMetadataFlat retrieves all metadata for a file as a flat map.
+//
+// Reads the record's field-name index and MGETs those keys — two round-trips,
+// independent of keyspace size. Records written before the index existed fall
+// back to the old SCAN and are indexed on the way out. See field_index.go for
+// why the index exists and why it is safe to derive.
 func (c *Client) GetMetadataFlat(hashID string) (map[string]string, error) {
 	c.mu.RLock()
 	defer c.mu.RUnlock()
@@ -122,41 +126,7 @@ func (c *Client) GetMetadataFlat(hashID string) (map[string]string, error) {
 	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
 	defer cancel()
 
-	prefix := c.buildKeyPrefix(hashID)
-	result := make(map[string]string)
-
-	// Scan for all keys with this prefix
-	var cursor uint64
-	for {
-		keys, nextCursor, err := c.client.Scan(ctx, cursor, prefix+"*", 1000).Result()
-		if err != nil {
-			return nil, fmt.Errorf("scan failed: %w", err)
-		}
-
-		if len(keys) > 0 {
-			values, err := c.client.MGet(ctx, keys...).Result()
-			if err != nil {
-				return nil, fmt.Errorf("mget failed: %w", err)
-			}
-			for i, key := range keys {
-				field := strings.TrimPrefix(key, prefix)
-				if values[i] != nil {
-					result[field] = values[i].(string)
-				}
-			}
-		}
-
-		cursor = nextCursor
-		if cursor == 0 {
-			break
-		}
-	}
-
-	if len(result) == 0 {
-		return nil, nil
-	}
-
-	return result, nil
+	return c.readMetadataLocked(ctx, hashID)
 }
 
 // SetMetadataFlat stores metadata for a file using flat keys
@@ -179,9 +149,21 @@ func (c *Client) SetMetadataFlat(hashID string, metadata map[string]string) erro
 	prefix := c.buildKeyPrefix(hashID)
 	pipe := c.client.Pipeline()
 
+	fields := make([]string, 0, len(metadata))
 	for field, value := range metadata {
-		pipe.Set(ctx, prefix+field, value, 0)
+		// A payload carrying the reserved name (a hand-made snapshot, a buggy
+		// client) would SET a string over the index SET and break every
+		// subsequent read of this record. Drop it — it is derived state, so
+		// there is never a reason to accept it from outside.
+		if isReservedFieldName(field) {
+			continue
 		}
+		pipe.Set(ctx, prefix+field, value, 0)
+		fields = append(fields, field)
+	}
+	// Same pipeline as the SETs, so indexing the field names costs no extra
+	// round-trip. See field_index.go.
+	c.addFieldsToPipe(ctx, pipe, hashID, fields)
 
 	if _, err := pipe.Exec(ctx); err != nil {
 		return fmt.Errorf("pipeline exec failed: %w", err)
@@ -256,6 +238,12 @@ func (c *Client) SetProperty(hashID, property, value string) error {
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 
+	// See SetMetadataFlat: the reserved name addresses the derived index SET,
+	// not a metadata field.
+	if isReservedFieldName(property) {
+		return fmt.Errorf("field name %q is reserved", property)
+	}
+
 	key := c.buildKeyPrefix(hashID) + property
 
 	// Use SET to set the key
@@ -263,6 +251,8 @@ func (c *Client) SetProperty(hashID, property, value string) error {
 		return fmt.Errorf("set failed: %w", err)
 	}
 
+	c.addFieldLocked(ctx, hashID, property)
+
 	// Add to index set
 	if err := c.client.SAdd(ctx, c.buildIndexKey(), hashID).Err(); err != nil {
 		return fmt.Errorf("sadd index failed: %w", err)
@@ -387,32 +377,14 @@ func normalizeFilesRelativePath(p string) string {
 // getMetadataFlatInternal mirrors GetMetadataFlat without re-acquiring the lock.
 // Callers must already hold c.mu.
 func (c *Client) getMetadataFlatInternal(ctx context.Context, hashID string) (map[string]string, error) {
-	prefix := c.buildKeyPrefix(hashID)
-	result := make(map[string]string)
-
-	var cursor uint64
-	for {
-		keys, nextCursor, err := c.client.Scan(ctx, cursor, prefix+"*", 1000).Result()
+	result, err := c.readMetadataLocked(ctx, hashID)
 	if err != nil {
-			return nil, fmt.Errorf("scan failed: %w", err)
-		}
-		if len(keys) > 0 {
-			values, err := c.client.MGet(ctx, keys...).Result()
-			if err != nil {
-				return nil, fmt.Errorf("mget failed: %w", err)
-			}
-			for i, key := range keys {
-				if values[i] == nil {
-					continue
-				}
-				field := strings.TrimPrefix(key, prefix)
-				result[field] = values[i].(string)
-			}
-		}
-		cursor = nextCursor
-		if cursor == 0 {
-			break
+		return nil, err
 	}
+	// Unlike GetMetadataFlat, this variant hands back an empty map rather than
+	// nil for a record with no fields — GetMetadataDocument writes into it.
+	if result == nil {
+		result = make(map[string]string)
 	}
 	return result, nil
 }
@@ -449,9 +421,17 @@ func (c *Client) MergeMetadataFlat(hashID string, metadata map[string]string) (i
 	prefix := c.buildKeyPrefix(hashID)
 	pipe := c.client.Pipeline()
 
+	fields := make([]string, 0, len(metadata))
 	for field, value := range metadata {
-		pipe.Set(ctx, prefix+field, value, 0)
+		// See SetMetadataFlat: never let an inbound payload overwrite the
+		// derived index SET with a string.
+		if isReservedFieldName(field) {
+			continue
 		}
+		pipe.Set(ctx, prefix+field, value, 0)
+		fields = append(fields, field)
+	}
+	c.addFieldsToPipe(ctx, pipe, hashID, fields)
 
 	if _, err := pipe.Exec(ctx); err != nil {
 		return 0, fmt.Errorf("pipeline exec failed: %w", err)
@@ -484,7 +464,11 @@ func (c *Client) DeleteProperty(hashID, property string) error {
 	defer cancel()
 
 	key := c.buildKeyPrefix(hashID) + property
-	return c.client.Del(ctx, key).Err()
+	if err := c.client.Del(ctx, key).Err(); err != nil {
+		return err
+	}
+	c.removeFieldLocked(ctx, hashID, property)
+	return nil
 }
 
 // AddToSet adds a value to a set-type field (stored as pipe-delimited string in flat key)
@@ -530,6 +514,9 @@ func (c *Client) AddToSet(hashID, property, value string) (bool, error) {
 		return false, err
 	}
 
+	// The SET above may have created the key, so the field name is new.
+	c.addFieldLocked(ctx, hashID, property)
+
 	// Add to index set
 	if err := c.client.SAdd(ctx, c.buildIndexKey(), hashID).Err(); err != nil {
 		return true, fmt.Errorf("sadd index failed: %w", err)
@@ -582,7 +569,12 @@ func (c *Client) RemoveFromSet(hashID, property, value string) (bool, error) {
 
 	// Save back (or delete key if empty)
 	if len(newValues) == 0 {
-		return true, c.client.Del(ctx, key).Err()
+		if err := c.client.Del(ctx, key).Err(); err != nil {
+			return true, err
+		}
+		// The key is gone, so the field name must leave the index too.
+		c.removeFieldLocked(ctx, hashID, property)
+		return true, nil
 	}
 
 	newValue := strings.Join(newValues, "|")
@@ -764,6 +756,12 @@ func (c *Client) GetSearchIndexTuples() ([]SearchTuple, error) {
 			if k == indexKey {
 				continue
 			}
+			// Per-record field indexes live under the same prefix and are SETs
+			// — MGET would error on them, and they are bookkeeping, not
+			// searchable metadata.
+			if strings.HasSuffix(k, "/"+FieldsField) {
+				continue
+			}
 			keyBatch = append(keyBatch, k)
 			if len(keyBatch) >= 1000 {
 				if err := flush(); err != nil {
@@ -1039,6 +1037,12 @@ func (c *Client) FindByValue(contains string, fields []string, limit int) ([]Fin
 		if field == "" {
 			continue
 		}
+		// The per-record field index is a SET, so an MGET over those keys
+		// would fail the whole query with WRONGTYPE. It holds field names,
+		// not values, so there is nothing here to search anyway.
+		if isReservedFieldName(field) {
+			continue
+		}
 		if len(matches) >= limit {
 			truncated = true
 			break
@@ -1207,7 +1211,17 @@ func (c *Client) SetRawValue(key, value string) error {
 	if t != "none" && t != "string" {
 		return fmt.Errorf("key has incompatible type %q (only string is editable)", t)
 	}
-	return c.client.Set(ctx, full, value, 0).Err()
+	if err := c.client.Set(ctx, full, value, 0).Err(); err != nil {
+		return err
+	}
+	// The raw editor can mint a record property out of band, so it has to keep
+	// the field index honest like every other write path. Non-record keys
+	// (`cid:*`, `udl:*`, `file:__index__`, the index set itself) split false
+	// and are left alone.
+	if hashID, field, ok := c.splitRecordKey(full); ok {
+		c.addFieldLocked(ctx, hashID, field)
+	}
+	return nil
 }
 
 // DeleteRawKey deletes the exact key. Returns true if the key existed.
@@ -1222,10 +1236,16 @@ func (c *Client) DeleteRawKey(key string) (bool, error) {
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 
-	n, err := c.client.Del(ctx, c.buildKey(key)).Result()
+	full := c.buildKey(key)
+	n, err := c.client.Del(ctx, full).Result()
 	if err != nil {
 		return false, err
 	}
+	if n > 0 {
+		if hashID, field, ok := c.splitRecordKey(full); ok {
+			c.removeFieldLocked(ctx, hashID, field)
+		}
+	}
 	return n > 0, nil
 }
 
@@ -1257,14 +1277,20 @@ func (c *Client) IndexRemoveIfEmpty(hashID string) (bool, error) {
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 
-	prefix := c.buildKeyPrefix(hashID)
-	keys, _, err := c.client.Scan(ctx, 0, prefix+"*", 1).Result()
+	nonEmpty, err := c.recordHasFieldsLocked(ctx, hashID)
 	if err != nil {
-		return false, fmt.Errorf("scan failed: %w", err)
+		return false, err
 	}
-	if len(keys) > 0 {
+	if nonEmpty {
 		return false, nil
 	}
+
+	// Nothing left but (possibly) an empty field index — drop it too, so a
+	// de-indexed root leaves no orphan key behind.
+	if err := c.client.Del(ctx, c.buildFieldsKey(hashID)).Err(); err != nil {
+		return false, fmt.Errorf("del field index: %w", err)
+	}
+
 	removed, err := c.client.SRem(ctx, c.buildIndexKey(), hashID).Result()
 	if err != nil {
 		return false, fmt.Errorf("srem failed: %w", err)

internal/storage/field_index.go

@@ -0,0 +1,486 @@
+package storage
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/redis/go-redis/v9"
+)
+
+// The per-record field-name index.
+//
+// A record is stored as one flat Redis STRING per leaf
+// (`file:<uuid>/<field>`), and that shape is deliberate — it is what lets
+// keyspace notifications fire per *field*, so meta-fuse / meta-editor /
+// meta-search can react to one field changing instead of diffing a whole hash
+// (see docs/project-architecture/persistent-storage.md, and
+// dev/test/suites/flat-keys.bats which asserts records are never hashes).
+//
+// The cost of that shape was that reading one record meant asking Redis "which
+// keys start with `file:<uuid>/`?", and Redis has no prefix index — `SCAN
+// MATCH` walks the ENTIRE keyspace. On a real store (445k keys, ~36 fields per
+// record) a single record read cost ~670 SCAN round-trips and ~0.5s of Redis
+// CPU; SCAN accounted for ~97.5% of all Redis CPU time on the box. That sat on
+// the byte path for every poster meta-share served, which is how it surfaced.
+//
+// So we keep the flat keys and index their *names*:
+//
+//	file:<uuid>/__fields__   SET{ "filePath", "sizeByte", "cids/bafk…", … }
+//
+// A read becomes SMEMBERS + a chunked MGET — two round-trips instead of ~670.
+// This mirrors `file:__index__` (the set of every root) one level down, and the
+// bulk reads already worked this way: `GetTuplesForAllFiles` composes its keys
+// from the index rather than scanning per root.
+//
+// **The set is derived, never authoritative.** The flat keys remain the source
+// of truth. A record whose set is missing (anything written before this landed)
+// falls back to the old SCAN and populates the set on the way out, so existing
+// deployments heal themselves on first touch with no migration and no
+// SchemaVersion bump — which matters, because a bump is a flag day that makes
+// operators wipe Redis (see schema_version.go).
+//
+// The consistency argument that makes a derived index safe here is meta-core's
+// single-writer rule: leader election publishes an empty `redisUrl` on purpose
+// (internal/leader/election.go), so nothing outside this process can write a
+// `file:` key behind the index's back.
+
+// FieldsField is the reserved field name holding a record's field-name index.
+//
+// It is a Redis SET living at `file:<uuid>/__fields__`, i.e. inside the same
+// prefix as the record's own fields. That keeps every existing sweep correct
+// for free — `DeleteMetadata`, `DeleteRoot` and `ClearAllMetadata` all
+// SCAN+DEL `file:<uuid>/*` and therefore drop the index along with the record.
+// The flip side is that every reader which enumerates that prefix has to skip
+// it, exactly as they already skip `file:__index__`.
+const FieldsField = "__fields__"
+
+// mgetBatch caps how many keys go into one MGET. Redis limits arguments per
+// command; 1000 is the batch size `GetTuplesForAllFiles` already uses.
+const mgetBatch = 1000
+
+// buildFieldsKey constructs the key for a record's field-name index set.
+func (c *Client) buildFieldsKey(hashID string) string {
+	return c.buildKeyPrefix(hashID) + FieldsField
+}
+
+// isReservedFieldName reports whether a field name is bookkeeping rather than
+// metadata, and so must never be returned to an API caller, exported into a
+// snapshot, or folded into the search haystack.
+func isReservedFieldName(field string) bool {
+	return field == FieldsField
+}
+
+// readMetadataLocked returns a record's full flat metadata. Caller must hold
+// c.mu (either mode).
+//
+// Fast path: SMEMBERS the field index, then MGET those keys. Fallback: the
+// legacy full-keyspace SCAN, whose result is used to populate the index so the
+// next read takes the fast path.
+//
+// Returns (nil, nil) for a record with no fields — callers treat that as 404.
+func (c *Client) readMetadataLocked(ctx context.Context, hashID string) (map[string]string, error) {
+	fields, err := c.client.SMembers(ctx, c.buildFieldsKey(hashID)).Result()
+	if err != nil && err != redis.Nil {
+		return nil, fmt.Errorf("smembers fields: %w", err)
+	}
+
+	if len(fields) > 0 {
+		return c.mgetFieldsLocked(ctx, hashID, fields)
+	}
+
+	// No index yet: a record written before the index existed, or one that
+	// genuinely has no fields. Fall back to the scan, then backfill.
+	result, scanned, err := c.scanMetadataLocked(ctx, hashID)
+	if err != nil {
+		return nil, err
+	}
+	if len(scanned) > 0 {
+		// Best-effort: a failed backfill costs the next read a scan, it does
+		// not make the read wrong. Never fail the caller's read for it.
+		if err := c.client.SAdd(ctx, c.buildFieldsKey(hashID), toAny(scanned)...).Err(); err != nil {
+			logFieldIndexBackfillFailure(hashID, err)
+		}
+	}
+	if len(result) == 0 {
+		return nil, nil
+	}
+	return result, nil
+}
+
+// mgetFieldsLocked reads the named fields of a record in chunked MGETs.
+//
+// A field named in the index whose key is gone reads back as nil and is simply
+// skipped: a *stale* index entry is harmless. (A *missing* one would silently
+// drop data, which is why every write path maintains the set.)
+func (c *Client) mgetFieldsLocked(ctx context.Context, hashID string, fields []string) (map[string]string, error) {
+	prefix := c.buildKeyPrefix(hashID)
+
+	// Keep the reserved name out of the key list AND out of the result — it is
+	// a SET, so MGET would error on it in real Redis rather than return nil.
+	wanted := make([]string, 0, len(fields))
+	for _, f := range fields {
+		if !isReservedFieldName(f) {
+			wanted = append(wanted, f)
+		}
+	}
+	if len(wanted) == 0 {
+		return nil, nil
+	}
+
+	result := make(map[string]string, len(wanted))
+	for start := 0; start < len(wanted); start += mgetBatch {
+		end := start + mgetBatch
+		if end > len(wanted) {
+			end = len(wanted)
+		}
+		chunk := wanted[start:end]
+		keys := make([]string, len(chunk))
+		for i, f := range chunk {
+			keys[i] = prefix + f
+		}
+		values, err := c.client.MGet(ctx, keys...).Result()
+		if err != nil {
+			return nil, fmt.Errorf("mget failed: %w", err)
+		}
+		for i, v := range values {
+			if s, ok := v.(string); ok {
+				result[chunk[i]] = s
+			}
+		}
+	}
+
+	if len(result) == 0 {
+		return nil, nil
+	}
+	return result, nil
+}
+
+// scanMetadataLocked is the pre-index read: SCAN the record's key prefix, then
+// MGET. Retained as the fallback for un-indexed records and as the source for
+// the backfill. Returns the metadata plus the field names it saw (which
+// includes fields whose value came back nil, so the index stays complete).
+func (c *Client) scanMetadataLocked(ctx context.Context, hashID string) (map[string]string, []string, error) {
+	prefix := c.buildKeyPrefix(hashID)
+	result := make(map[string]string)
+	var fields []string
+
+	var cursor uint64
+	for {
+		keys, nextCursor, err := c.client.Scan(ctx, cursor, prefix+"*", 1000).Result()
+		if err != nil {
+			return nil, nil, fmt.Errorf("scan failed: %w", err)
+		}
+
+		if len(keys) > 0 {
+			wanted := make([]string, 0, len(keys))
+			for _, key := range keys {
+				field := strings.TrimPrefix(key, prefix)
+				if isReservedFieldName(field) {
+					continue
+				}
+				fields = append(fields, field)
+				wanted = append(wanted, key)
+			}
+			if len(wanted) > 0 {
+				values, err := c.client.MGet(ctx, wanted...).Result()
+				if err != nil {
+					return nil, nil, fmt.Errorf("mget failed: %w", err)
+				}
+				for i, key := range wanted {
+					if s, ok := values[i].(string); ok {
+						result[strings.TrimPrefix(key, prefix)] = s
+					}
+				}
+			}
+		}
+
+		cursor = nextCursor
+		if cursor == 0 {
+			break
+		}
+	}
+
+	return result, fields, nil
+}
+
+// addFieldsToPipe records field names in the index as part of an existing
+// pipeline, so a multi-field write stays one round-trip.
+func (c *Client) addFieldsToPipe(ctx context.Context, pipe redis.Pipeliner, hashID string, fields []string) {
+	names := filterReserved(fields)
+	if len(names) == 0 {
+		return
+	}
+	pipe.SAdd(ctx, c.buildFieldsKey(hashID), toAny(names)...)
+}
+
+// addFieldLocked records a single field name in the index. Caller must hold
+// c.mu. Best-effort by the same argument as the backfill: the flat key is the
+// source of truth, and a lost index entry is repaired by the next scan
+// fallback — so it must not fail the write it accompanies.
+func (c *Client) addFieldLocked(ctx context.Context, hashID, field string) {
+	if isReservedFieldName(field) {
+		return
+	}
+	if err := c.client.SAdd(ctx, c.buildFieldsKey(hashID), field).Err(); err != nil {
+		logFieldIndexWriteFailure(hashID, field, err)
+	}
+}
+
+// removeFieldLocked drops a field name from the index. Caller must hold c.mu.
+//
+// A stale entry left behind by a failure is harmless (the MGET reads nil and
+// skips it), so this is best-effort too.
+func (c *Client) removeFieldLocked(ctx context.Context, hashID, field string) {
+	if isReservedFieldName(field) {
+		return
+	}
+	if err := c.client.SRem(ctx, c.buildFieldsKey(hashID), field).Err(); err != nil {
+		logFieldIndexWriteFailure(hashID, field, err)
+	}
+}
+
+// fieldsWithPrefixLocked returns the record's indexed field names under a
+// field-name prefix (e.g. "cids/"), falling back to a SCAN when the record has
+// no index yet. Caller must hold c.mu.
+func (c *Client) fieldsWithPrefixLocked(ctx context.Context, hashID, fieldPrefix string) ([]string, error) {
+	fields, err := c.client.SMembers(ctx, c.buildFieldsKey(hashID)).Result()
+	if err != nil && err != redis.Nil {
+		return nil, fmt.Errorf("smembers fields: %w", err)
+	}
+	if len(fields) == 0 {
+		// Un-indexed record: scan the narrower `<prefix><fieldPrefix>*` glob
+		// rather than the whole record, and do not backfill from it — a
+		// partial view must never be mistaken for a complete index.
+		return c.scanFieldsWithPrefixLocked(ctx, hashID, fieldPrefix)
+	}
+
+	var out []string
+	for _, f := range fields {
+		if isReservedFieldName(f) {
+			continue
+		}
+		if strings.HasPrefix(f, fieldPrefix) {
+			out = append(out, strings.TrimPrefix(f, fieldPrefix))
+		}
+	}
+	// SMEMBERS order is arbitrary and SCAN order was too, so the CID array in
+	// GetMetadataDocument used to shuffle between calls. Sorting costs nothing
+	// at these sizes and makes the response byte-stable, which matters for
+	// anything diffing or caching it.
+	sort.Strings(out)
+	return out, nil
+}
+
+func (c *Client) scanFieldsWithPrefixLocked(ctx context.Context, hashID, fieldPrefix string) ([]string, error) {
+	prefix := c.buildKeyPrefix(hashID) + fieldPrefix
+	var out []string
+	var cursor uint64
+	for {
+		keys, next, err := c.client.Scan(ctx, cursor, prefix+"*", 1000).Result()
+		if err != nil {
+			return nil, fmt.Errorf("scan fields: %w", err)
+		}
+		for _, k := range keys {
+			out = append(out, strings.TrimPrefix(k, prefix))
+		}
+		cursor = next
+		if cursor == 0 {
+			break
+		}
+	}
+	sort.Strings(out)
+	return out, nil
+}
+
+// WarmFieldIndexes builds every missing field index in a SINGLE keyspace pass.
+//
+// Without it the indexes still appear — each record's first read falls back to
+// a scan and backfills — but that transition is O(records x keyspace) and it
+// runs again from scratch after any wipe. Measured on the real store (12,345
+// records / 445k keys) the lazy path warmed at ~1.3 records/s: a ~2.6 HOUR tail
+// during which meta-core stays pinned exactly as it was before the fix. Worse,
+// `RepublishMetadata` walks every record at startup, so that tail is paid on
+// every restart.
+//
+// One bulk `file:*` scan grouped by record collapses it to a few seconds. This
+// is the same shape `GetSearchIndexTuples` already uses for the bulk read, and
+// for the same reason.
+//
+// Idempotent, and safe to run against a partially-indexed store: SADD of a
+// member that is already there is a no-op. Returns the number of records it
+// wrote an index for.
+func (c *Client) WarmFieldIndexes() (int, error) {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+
+	if c.client == nil {
+		return 0, fmt.Errorf("not connected")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
+	defer cancel()
+
+	filePrefix := c.buildKey("file:") // {filePrefix}{hashID}/{field}
+	indexKey := c.buildIndexKey()
+
+	byRecord := make(map[string][]string)
+	var cursor uint64
+	for {
+		keys, next, err := c.client.Scan(ctx, cursor, filePrefix+"*", 1000).Result()
+		if err != nil {
+			return 0, fmt.Errorf("scan: %w", err)
+		}
+		for _, k := range keys {
+			if k == indexKey {
+				continue
+			}
+			// hashID has no '/'; field may contain them.
+			rest := strings.TrimPrefix(k, filePrefix)
+			slash := strings.IndexByte(rest, '/')
+			if slash < 0 {
+				continue
+			}
+			hashID, field := rest[:slash], rest[slash+1:]
+			if isReservedFieldName(field) {
+				// Already indexed — remember the record so we don't treat it
+				// as missing, but there is nothing to add for this key.
+				if _, seen := byRecord[hashID]; !seen {
+					byRecord[hashID] = nil
+				}
+				continue
+			}
+			byRecord[hashID] = append(byRecord[hashID], field)
+		}
+		cursor = next
+		if cursor == 0 {
+			break
+		}
+	}
+
+	written := 0
+	pipe := c.client.Pipeline()
+	pending := 0
+	for hashID, fields := range byRecord {
+		if len(fields) == 0 {
+			continue
+		}
+		pipe.SAdd(ctx, c.buildFieldsKey(hashID), toAny(fields)...)
+		written++
+		pending++
+		// Bound the pipeline so a large store doesn't buffer every command.
+		if pending >= 500 {
+			if _, err := pipe.Exec(ctx); err != nil {
+				return written, fmt.Errorf("sadd pipeline: %w", err)
+			}
+			pipe = c.client.Pipeline()
+			pending = 0
+		}
+	}
+	if pending > 0 {
+		if _, err := pipe.Exec(ctx); err != nil {
+			return written, fmt.Errorf("sadd pipeline: %w", err)
+		}
+	}
+	return written, nil
+}
+
+// recordHasFieldsLocked reports whether a record still has at least one
+// property key. Caller must hold c.mu.
+//
+// `__fields__` must not count — it lives under the record's own prefix, so a
+// naive "does any `file:<id>/*` key exist?" probe would answer yes for a record
+// whose every property has been deleted, and the root could never be
+// de-indexed.
+func (c *Client) recordHasFieldsLocked(ctx context.Context, hashID string) (bool, error) {
+	n, err := c.client.SCard(ctx, c.buildFieldsKey(hashID)).Result()
+	if err != nil && err != redis.Nil {
+		return false, fmt.Errorf("scard fields: %w", err)
+	}
+	if n > 0 {
+		return true, nil
+	}
+
+	// Either genuinely empty, or an un-indexed record from before the index
+	// existed — the scan is the only way to tell those apart.
+	prefix := c.buildKeyPrefix(hashID)
+	var cursor uint64
+	for {
+		keys, next, err := c.client.Scan(ctx, cursor, prefix+"*", 1000).Result()
+		if err != nil {
+			return false, fmt.Errorf("scan failed: %w", err)
+		}
+		for _, k := range keys {
+			if !isReservedFieldName(strings.TrimPrefix(k, prefix)) {
+				return true, nil
+			}
+		}
+		cursor = next
+		if cursor == 0 {
+			return false, nil
+		}
+	}
+}
+
+// splitRecordKey decomposes a full Redis key into (hashID, field) when it is a
+// record property key. Reports ok=false for anything else — the index set
+// itself, `file:__index__`, and any key outside the `file:` namespace.
+//
+// This is what lets the raw KV escape hatch (SetRawValue / DeleteRawKey) keep
+// the index honest without its callers knowing the layout.
+func (c *Client) splitRecordKey(key string) (hashID, field string, ok bool) {
+	rest := key
+	if c.prefix != "" {
+		var found bool
+		rest, found = strings.CutPrefix(rest, c.prefix)
+		if !found {
+			return "", "", false
+		}
+	}
+	rest, found := strings.CutPrefix(rest, "file:")
+	if !found {
+		return "", "", false
+	}
+	hashID, field, found = strings.Cut(rest, "/")
+	if !found || hashID == "" || field == "" {
+		// `file:__index__` has no slash and lands here.
+		return "", "", false
+	}
+	if isReservedFieldName(field) {
+		return "", "", false
+	}
+	return hashID, field, true
+}
+
+// The index is derived and self-healing, so a write failure degrades
+// performance (the next read falls back to a scan and re-populates) rather
+// than correctness. Log it rather than failing the caller's write — but log it
+// loudly, because a *persistent* failure means every read is scanning again.
+func logFieldIndexWriteFailure(hashID, field string, err error) {
+	log.Printf("[Storage] field index update failed for %s/%s: %v (reads fall back to SCAN)", hashID, field, err)
+}
+
+func logFieldIndexBackfillFailure(hashID string, err error) {
+	log.Printf("[Storage] field index backfill failed for %s: %v (reads keep falling back to SCAN)", hashID, err)
+}
+
+func filterReserved(fields []string) []string {
+	out := make([]string, 0, len(fields))
+	for _, f := range fields {
+		if !isReservedFieldName(f) {
+			out = append(out, f)
+		}
+	}
+	return out
+}
+
+func toAny(ss []string) []interface{} {
+	out := make([]interface{}, len(ss))
+	for i, s := range ss {
+		out[i] = s
+	}
+	return out
+}

internal/storage/field_index_test.go

@@ -0,0 +1,464 @@
+package storage
+
+import (
+	"context"
+	"reflect"
+	"sort"
+	"testing"
+)
+
+// Tests for the per-record field-name index (field_index.go).
+//
+// The index is what makes a record read O(fields) instead of O(keyspace). It is
+// derived state, so the properties worth pinning are: (1) a read served from
+// the index returns exactly what a scan would have, (2) a record with no index
+// still reads correctly and gets one, and (3) every write path keeps it in
+// sync — because a *missing* entry silently drops a field from every
+// subsequent read, which is the one failure mode that would be invisible.
+//
+// Every case runs with and without a namespace prefix, matching clear_test.go.
+
+func forEachPrefix(t *testing.T, fn func(t *testing.T, c *Client)) {
+	t.Helper()
+	for _, prefix := range []string{"", "mm:"} {
+		name := "no-prefix"
+		if prefix != "" {
+			name = "prefix-" + prefix
+		}
+		t.Run(name, func(t *testing.T) {
+			c, _ := newTestClient(t, prefix)
+			fn(t, c)
+		})
+	}
+}
+
+func fieldNames(t *testing.T, c *Client, hashID string) []string {
+	t.Helper()
+	got, err := c.client.SMembers(context.Background(), c.buildFieldsKey(hashID)).Result()
+	if err != nil {
+		t.Fatalf("smembers fields: %v", err)
+	}
+	sort.Strings(got)
+	return got
+}
+
+func mustGetMetadata(t *testing.T, c *Client, hashID string) map[string]string {
+	t.Helper()
+	got, err := c.GetMetadataFlat(hashID)
+	if err != nil {
+		t.Fatalf("GetMetadataFlat: %v", err)
+	}
+	return got
+}
+
+// SetMetadataFlat must index every field it writes, and the read must come back
+// through the index rather than a scan.
+func TestFieldIndex_SetMetadataFlatIndexesFields(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		want := map[string]string{
+			"filePath":                      "/files/watch/a.mkv",
+			"sizeByte":                      "123",
+			"cids/bafkreiaaa":               "true",
+			"titles/jpn":                    "ナルト",
+			"categories/TV/UHD":             "true",
+			"info/files/0/path":             "a.mkv",
+			"provenance":                    `{"source":"gateway"}`,
+			"plot/eng":                      "a plot with / slashes",
+			"stream/3":                      "eng",
+			"genres/action":                 "true",
+			"originaltitle":                 "Naruto",
+			"deeply/nested/field/name/here": "x",
+		}
+		if err := c.SetMetadataFlat("root1", want); err != nil {
+			t.Fatalf("SetMetadataFlat: %v", err)
+		}
+
+		if got := mustGetMetadata(t, c, "root1"); !reflect.DeepEqual(got, want) {
+			t.Errorf("GetMetadataFlat mismatch\n got %v\nwant %v", got, want)
+		}
+
+		// The index must name every field — including the ones containing '/',
+		// which is where a naive "split on the first slash" would go wrong.
+		wantNames := make([]string, 0, len(want))
+		for k := range want {
+			wantNames = append(wantNames, k)
+		}
+		sort.Strings(wantNames)
+		if got := fieldNames(t, c, "root1"); !reflect.DeepEqual(got, wantNames) {
+			t.Errorf("field index mismatch\n got %v\nwant %v", got, wantNames)
+		}
+	})
+}
+
+// A record written before the index existed has flat keys but no set. The read
+// must still be correct, and must leave an index behind so the next read takes
+// the fast path. This is the migration path for the ~12k records already on
+// disk: no flag day, no SchemaVersion bump.
+func TestFieldIndex_LazyBackfillForUnindexedRecord(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		ctx := context.Background()
+		prefix := c.buildKeyPrefix("legacy")
+		want := map[string]string{
+			"filePath":        "/files/watch/old.mkv",
+			"sizeByte":        "42",
+			"cids/bafkreiold": "true",
+		}
+		// Write the flat keys directly, bypassing every index-maintaining path
+		// — exactly the shape a pre-upgrade store is in.
+		for f, v := range want {
+			if err := c.client.Set(ctx, prefix+f, v, 0).Err(); err != nil {
+				t.Fatalf("seed %s: %v", f, err)
+			}
+		}
+		if names := fieldNames(t, c, "legacy"); len(names) != 0 {
+			t.Fatalf("precondition: expected no index, got %v", names)
+		}
+
+		if got := mustGetMetadata(t, c, "legacy"); !reflect.DeepEqual(got, want) {
+			t.Errorf("scan-fallback read mismatch\n got %v\nwant %v", got, want)
+		}
+
+		wantNames := []string{"cids/bafkreiold", "filePath", "sizeByte"}
+		if got := fieldNames(t, c, "legacy"); !reflect.DeepEqual(got, wantNames) {
+			t.Errorf("backfill mismatch\n got %v\nwant %v", got, wantNames)
+		}
+
+		// Second read, now served from the index, must agree with the first.
+		if got := mustGetMetadata(t, c, "legacy"); !reflect.DeepEqual(got, want) {
+			t.Errorf("indexed read disagrees with scan read\n got %v\nwant %v", got, want)
+		}
+	})
+}
+
+// Each single-field write path must add to the index, and each delete path must
+// remove from it. A field written but not indexed vanishes from every later
+// read — silently — so this is the property that matters most.
+func TestFieldIndex_SingleFieldWritePathsStayInSync(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		if err := c.SetProperty("root2", "title", "Naruto"); err != nil {
+			t.Fatalf("SetProperty: %v", err)
+		}
+		if _, err := c.AddToSet("root2", "tags", "anime"); err != nil {
+			t.Fatalf("AddToSet: %v", err)
+		}
+		if _, err := c.MergeMetadataFlat("root2", map[string]string{"year": "2002"}); err != nil {
+			t.Fatalf("MergeMetadataFlat: %v", err)
+		}
+
+		want := map[string]string{"title": "Naruto", "tags": "anime", "year": "2002"}
+		if got := mustGetMetadata(t, c, "root2"); !reflect.DeepEqual(got, want) {
+			t.Errorf("after writes\n got %v\nwant %v", got, want)
+		}
+
+		// DeleteProperty drops the key and the index entry.
+		if err := c.DeleteProperty("root2", "title"); err != nil {
+			t.Fatalf("DeleteProperty: %v", err)
+		}
+		// RemoveFromSet deletes the key once its last member is gone.
+		if _, err := c.RemoveFromSet("root2", "tags", "anime"); err != nil {
+			t.Fatalf("RemoveFromSet: %v", err)
+		}
+
+		want = map[string]string{"year": "2002"}
+		if got := mustGetMetadata(t, c, "root2"); !reflect.DeepEqual(got, want) {
+			t.Errorf("after deletes\n got %v\nwant %v", got, want)
+		}
+		if got := fieldNames(t, c, "root2"); !reflect.DeepEqual(got, []string{"year"}) {
+			t.Errorf("index after deletes = %v, want [year]", got)
+		}
+	})
+}
+
+// The raw KV editor can mint or drop a record property out of band. It is the
+// one write path whose caller does not know the record layout, so the storage
+// layer has to keep the index honest for it.
+func TestFieldIndex_RawKeyEditorStaysInSync(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		if err := c.SetRawValue("file:root3/hand-written", "v"); err != nil {
+			t.Fatalf("SetRawValue: %v", err)
+		}
+		if got := fieldNames(t, c, "root3"); !reflect.DeepEqual(got, []string{"hand-written"}) {
+			t.Errorf("index after SetRawValue = %v", got)
+		}
+		if got := mustGetMetadata(t, c, "root3"); !reflect.DeepEqual(got, map[string]string{"hand-written": "v"}) {
+			t.Errorf("read after SetRawValue = %v", got)
+		}
+
+		if _, err := c.DeleteRawKey("file:root3/hand-written"); err != nil {
+			t.Fatalf("DeleteRawKey: %v", err)
+		}
+		if got := fieldNames(t, c, "root3"); len(got) != 0 {
+			t.Errorf("index after DeleteRawKey = %v, want empty", got)
+		}
+
+		// Keys outside the record namespace must be left alone entirely.
+		if err := c.SetRawValue("cid:bafkreiaaa", "root3"); err != nil {
+			t.Fatalf("SetRawValue non-record: %v", err)
+		}
+		if _, _, ok := c.splitRecordKey(c.buildKey("cid:bafkreiaaa")); ok {
+			t.Error("cid: key was treated as a record property")
+		}
+		if _, _, ok := c.splitRecordKey(c.buildIndexKey()); ok {
+			t.Error("file:__index__ was treated as a record property")
+		}
+	})
+}
+
+// Mint and AddAlias write fields through their own pipelines, so they need the
+// index too. AddAlias in particular is how `cids/<cid>` members appear — miss
+// it and a record's CID key-set becomes invisible to every read.
+func TestFieldIndex_MintAndAliasIndexTheirFields(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		uuid, err := c.Mint("/files/watch/x.mkv", 10, 20)
+		if err != nil {
+			t.Fatalf("Mint: %v", err)
+		}
+		if got := fieldNames(t, c, uuid); !reflect.DeepEqual(got, []string{"filePath", "mtimeNano", "sizeByte"}) {
+			t.Errorf("index after Mint = %v", got)
+		}
+
+		if err := c.AddAlias(uuid, "bafkreiaaa"); err != nil {
+			t.Fatalf("AddAlias: %v", err)
+		}
+		got := mustGetMetadata(t, c, uuid)
+		if got["cids/bafkreiaaa"] != "true" {
+			t.Errorf("alias key-set member missing from read: %v", got)
+		}
+
+		// And the CID enumeration, which now reads the index instead of
+		// scanning, must see it.
+		cids, err := c.cidsForRootLocked(context.Background(), uuid)
+		if err != nil {
+			t.Fatalf("cidsForRootLocked: %v", err)
+		}
+		if !reflect.DeepEqual(cids, []string{"bafkreiaaa"}) {
+			t.Errorf("cidsForRoot = %v, want [bafkreiaaa]", cids)
+		}
+	})
+}
+
+// The index lives under the record's own prefix, so the "is this record empty?"
+// probe would otherwise always find it and the root could never be de-indexed.
+func TestFieldIndex_IndexRemoveIfEmptyIgnoresTheIndexItself(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		if err := c.SetProperty("root4", "title", "x"); err != nil {
+			t.Fatalf("SetProperty: %v", err)
+		}
+
+		removed, err := c.IndexRemoveIfEmpty("root4")
+		if err != nil {
+			t.Fatalf("IndexRemoveIfEmpty: %v", err)
+		}
+		if removed {
+			t.Error("de-indexed a record that still has a field")
+		}
+
+		if err := c.DeleteProperty("root4", "title"); err != nil {
+			t.Fatalf("DeleteProperty: %v", err)
+		}
+		removed, err = c.IndexRemoveIfEmpty("root4")
+		if err != nil {
+			t.Fatalf("IndexRemoveIfEmpty: %v", err)
+		}
+		if !removed {
+			t.Fatal("record with no fields left was not de-indexed — the field index masked emptiness")
+		}
+		// ...and the now-empty index key must not be left orphaned.
+		n, err := c.client.Exists(context.Background(), c.buildFieldsKey("root4")).Result()
+		if err != nil {
+			t.Fatalf("exists: %v", err)
+		}
+		if n != 0 {
+			t.Error("orphaned field index left behind after de-indexing")
+		}
+	})
+}
+
+// The same probe must still work for a record that predates the index.
+func TestFieldIndex_IndexRemoveIfEmptyHandlesUnindexedRecord(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		ctx := context.Background()
+		if err := c.client.Set(ctx, c.buildKeyPrefix("legacy2")+"title", "x", 0).Err(); err != nil {
+			t.Fatalf("seed: %v", err)
+		}
+		if err := c.IndexAdd("legacy2"); err != nil {
+			t.Fatalf("IndexAdd: %v", err)
+		}
+		removed, err := c.IndexRemoveIfEmpty("legacy2")
+		if err != nil {
+			t.Fatalf("IndexRemoveIfEmpty: %v", err)
+		}
+		if removed {
+			t.Error("de-indexed an un-indexed record that still has a field")
+		}
+	})
+}
+
+// The reserved name is bookkeeping. It must never reach an API caller as a
+// metadata field, and an inbound payload must never be able to overwrite the
+// SET with a string — that would break every subsequent read of the record.
+func TestFieldIndex_ReservedNameIsNeverMetadata(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		err := c.SetMetadataFlat("root5", map[string]string{
+			"title":     "x",
+			FieldsField: "malicious",
+		})
+		if err != nil {
+			t.Fatalf("SetMetadataFlat: %v", err)
+		}
+
+		got := mustGetMetadata(t, c, "root5")
+		if _, present := got[FieldsField]; present {
+			t.Errorf("%s surfaced as a metadata field: %v", FieldsField, got)
+		}
+		if !reflect.DeepEqual(got, map[string]string{"title": "x"}) {
+			t.Errorf("read = %v, want just title", got)
+		}
+
+		// The index must still be a SET — i.e. not clobbered by the payload.
+		if got := fieldNames(t, c, "root5"); !reflect.DeepEqual(got, []string{"title"}) {
+			t.Errorf("index = %v, want [title]", got)
+		}
+
+		if err := c.SetProperty("root5", FieldsField, "malicious"); err == nil {
+			t.Error("SetProperty accepted the reserved field name")
+		}
+	})
+}
+
+// Deleting a record must take its index with it, or a recreated root would
+// inherit phantom field names.
+func TestFieldIndex_DeleteMetadataRemovesTheIndex(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		if err := c.SetMetadataFlat("root6", map[string]string{"title": "x", "year": "2002"}); err != nil {
+			t.Fatalf("SetMetadataFlat: %v", err)
+		}
+		if _, err := c.DeleteMetadata("root6"); err != nil {
+			t.Fatalf("DeleteMetadata: %v", err)
+		}
+		if got := fieldNames(t, c, "root6"); len(got) != 0 {
+			t.Errorf("index survived DeleteMetadata: %v", got)
+		}
+		got, err := c.GetMetadataFlat("root6")
+		if err != nil {
+			t.Fatalf("GetMetadataFlat: %v", err)
+		}
+		if got != nil {
+			t.Errorf("deleted record still reads back: %v", got)
+		}
+	})
+}
+
+// A stale index entry (name present, key gone) is the benign direction: the
+// MGET reads nil and the field is skipped rather than erroring.
+func TestFieldIndex_StaleEntryIsHarmless(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		ctx := context.Background()
+		if err := c.SetMetadataFlat("root7", map[string]string{"title": "x"}); err != nil {
+			t.Fatalf("SetMetadataFlat: %v", err)
+		}
+		// Name a field that has no key behind it.
+		if err := c.client.SAdd(ctx, c.buildFieldsKey("root7"), "ghost").Err(); err != nil {
+			t.Fatalf("sadd ghost: %v", err)
+		}
+
+		got := mustGetMetadata(t, c, "root7")
+		if !reflect.DeepEqual(got, map[string]string{"title": "x"}) {
+			t.Errorf("read = %v, want just title", got)
+		}
+	})
+}
+
+// WarmFieldIndexes is what keeps an upgrade from spending hours reading
+// un-indexed records one full-keyspace scan at a time. It must index every
+// record in one pass, be idempotent, and leave already-indexed records alone.
+func TestFieldIndex_WarmBuildsEveryMissingIndex(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		ctx := context.Background()
+
+		// Two legacy records seeded straight into Redis (no index), plus one
+		// written through the normal path (already indexed).
+		legacy := map[string]map[string]string{
+			"old1": {"filePath": "/a.mkv", "cids/bafkreiaaa": "true"},
+			"old2": {"filePath": "/b.mkv", "titles/jpn": "ナルト", "deep/a/b": "x"},
+		}
+		for root, fields := range legacy {
+			for f, v := range fields {
+				if err := c.client.Set(ctx, c.buildKeyPrefix(root)+f, v, 0).Err(); err != nil {
+					t.Fatalf("seed: %v", err)
+				}
+			}
+			if err := c.IndexAdd(root); err != nil {
+				t.Fatalf("IndexAdd: %v", err)
+			}
+		}
+		if err := c.SetMetadataFlat("fresh", map[string]string{"title": "x"}); err != nil {
+			t.Fatalf("SetMetadataFlat: %v", err)
+		}
+
+		n, err := c.WarmFieldIndexes()
+		if err != nil {
+			t.Fatalf("WarmFieldIndexes: %v", err)
+		}
+		if n < 2 {
+			t.Errorf("warmed %d records, want at least the 2 legacy ones", n)
+		}
+
+		if got := fieldNames(t, c, "old1"); !reflect.DeepEqual(got, []string{"cids/bafkreiaaa", "filePath"}) {
+			t.Errorf("old1 index = %v", got)
+		}
+		if got := fieldNames(t, c, "old2"); !reflect.DeepEqual(got, []string{"deep/a/b", "filePath", "titles/jpn"}) {
+			t.Errorf("old2 index = %v", got)
+		}
+		if got := fieldNames(t, c, "fresh"); !reflect.DeepEqual(got, []string{"title"}) {
+			t.Errorf("fresh index disturbed = %v", got)
+		}
+
+		// Reads must now agree with what the records actually hold.
+		for root, want := range legacy {
+			if got := mustGetMetadata(t, c, root); !reflect.DeepEqual(got, want) {
+				t.Errorf("%s read = %v, want %v", root, got, want)
+			}
+		}
+
+		// Idempotent: a second pass changes nothing.
+		if _, err := c.WarmFieldIndexes(); err != nil {
+			t.Fatalf("second WarmFieldIndexes: %v", err)
+		}
+		if got := fieldNames(t, c, "old2"); !reflect.DeepEqual(got, []string{"deep/a/b", "filePath", "titles/jpn"}) {
+			t.Errorf("old2 index after second warm = %v", got)
+		}
+	})
+}
+
+// The bulk search read enumerates the whole file: keyspace and MGETs what it
+// finds. The index keys are SETs living in that same namespace, so it has to
+// skip them or the MGET fails with WRONGTYPE and search dies wholesale.
+func TestFieldIndex_SearchIndexTuplesSkipTheIndexKeys(t *testing.T) {
+	forEachPrefix(t, func(t *testing.T, c *Client) {
+		if err := c.SetMetadataFlat("root8", map[string]string{
+			"title":    "Naruto Shippuden",
+			"fileName": "naruto.mkv",
+		}); err != nil {
+			t.Fatalf("SetMetadataFlat: %v", err)
+		}
+
+		tuples, err := c.GetSearchIndexTuples()
+		if err != nil {
+			t.Fatalf("GetSearchIndexTuples: %v", err)
+		}
+		var found bool
+		for _, tu := range tuples {
+			if tu.HashID != "root8" {
+				continue
+			}
+			found = true
+			if _, present := tu.Fields[FieldsField]; present {
+				t.Errorf("%s leaked into the search tuple: %v", FieldsField, tu.Fields)
+			}
+		}
+		if !found {
+			t.Fatal("record missing from search tuples")
+		}
+	})
+}

internal/storage/udl.go

@@ -72,7 +72,15 @@ func udlIdxCidUsers(cid string) string      { return fmt.Sprintf("udl:idx:cid:%s
 //
 //	KEYS: 1=cell 2=idx:user:key 3=idx:user:cid 4=idx:cid:key 5=idx:cid:users
 //	ARGV: 1=version 2=ts 3=record 4=value 5=hasValue("1"/"0")
-//	      6=cid 7=key 8=uid
+//	      6=cid 7=key 8=uid 9=tombstone("1"/"0")
+//
+// A tombstone still writes the cell — that is the CRDT state, and dropping it
+// would let an older write resurrect the value — but it SREMs the cid from the
+// user+key index. That index is what `active_for_user_key` returns, so leaving
+// deleted cids in it means every My List / Continue Watching read carries every
+// title ever un-liked or finished, forever. The other three indexes are left
+// alone: they answer "who has an opinion about this cid", where a tombstone is
+// still an answer.
 var udlUpsertScript = redis.NewScript(`
 local stored = redis.call('HGET', KEYS[1], 'version')
 if stored and tonumber(stored) >= tonumber(ARGV[1]) then
@@ -84,7 +92,11 @@ if ARGV[5] == '1' then
 else
   redis.call('HDEL', KEYS[1], 'value')
 end
-redis.call('SADD', KEYS[2], ARGV[6])
+if ARGV[9] == '1' then
+  redis.call('SREM', KEYS[2], ARGV[6])
+else
+  redis.call('SADD', KEYS[2], ARGV[6])
+end
 redis.call('SADD', KEYS[3], ARGV[7])
 redis.call('SADD', KEYS[4], ARGV[8])
 redis.call('SADD', KEYS[5], ARGV[8])
@@ -94,8 +106,10 @@ return 1
 // UDLUpsertIfNewer stores a record cell iff version strictly exceeds the stored
 // one. publicValue/hasValue carry the plaintext JSON scalar for public keys
 // (used by aggregation); pass hasValue=false for private-tier keys and
-// tombstones. Returns accepted=false when the write is stale.
-func (c *Client) UDLUpsertIfNewer(uid, cid, key string, version, ts int64, record, publicValue string, hasValue bool) (bool, error) {
+// tombstones. Set tombstone=true when the record deletes the cell, which also
+// prunes the cid from the user+key index. Returns accepted=false when the write
+// is stale.
+func (c *Client) UDLUpsertIfNewer(uid, cid, key string, version, ts int64, record, publicValue string, hasValue, tombstone bool) (bool, error) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 
@@ -117,7 +131,11 @@ func (c *Client) UDLUpsertIfNewer(uid, cid, key string, version, ts int64, recor
 	if hasValue {
 		hv = "1"
 	}
-	res, err := udlUpsertScript.Run(ctx, c.client, keys, version, ts, record, publicValue, hv, cid, key, uid).Result()
+	tomb := "0"
+	if tombstone {
+		tomb = "1"
+	}
+	res, err := udlUpsertScript.Run(ctx, c.client, keys, version, ts, record, publicValue, hv, cid, key, uid, tomb).Result()
 	if err != nil {
 		return false, fmt.Errorf("udl upsert failed: %w", err)
 	}

internal/storage/udl_test.go

@@ -0,0 +1,79 @@
+package storage
+
+import (
+	"testing"
+)
+
+// A tombstone must leave the cell in place but drop the cid from the user+key
+// index.
+//
+// The cell has to stay: it *is* the CRDT state, and deleting it would let an
+// older write resurrect the value the tombstone retracted. The index entry has
+// to go: `active_for_user_key` returns every member of that set, so a set that
+// only ever grows means each My List / Continue Watching read carries every
+// title the user has ever un-liked or finished — with its full record blob —
+// forever. The reader filters them out again, so the cost is invisible until
+// the account is old.
+func TestUDLUpsert_TombstonePrunesUserKeyIndexButKeepsCell(t *testing.T) {
+	c, _ := newTestClient(t, "")
+
+	const (
+		uid = "zAlice"
+		cid = "bagcsaaalocator"
+		key = "seek"
+	)
+
+	// A live position: indexed, readable.
+	if ok, err := c.UDLUpsertIfNewer(uid, cid, key, 1, 100, "rec-v1", "", false, false); err != nil || !ok {
+		t.Fatalf("live upsert: ok=%v err=%v", ok, err)
+	}
+	entries, err := c.UDLListUserKey(uid, key)
+	if err != nil {
+		t.Fatalf("active_for_user_key: %v", err)
+	}
+	if len(entries) != 1 || entries[0].Cid != cid {
+		t.Fatalf("live position should be indexed, got %+v", entries)
+	}
+
+	// Finishing it tombstones the seek.
+	if ok, err := c.UDLUpsertIfNewer(uid, cid, key, 2, 200, "rec-v2-tombstone", "", false, true); err != nil || !ok {
+		t.Fatalf("tombstone upsert: ok=%v err=%v", ok, err)
+	}
+
+	entries, err = c.UDLListUserKey(uid, key)
+	if err != nil {
+		t.Fatalf("active_for_user_key after tombstone: %v", err)
+	}
+	if len(entries) != 0 {
+		t.Fatalf("tombstoned cid must leave the user+key index, got %+v", entries)
+	}
+
+	// ...but the cell itself survives, at the tombstone's version.
+	cell, exists, err := c.UDLGetCell(uid, cid, key)
+	if err != nil {
+		t.Fatalf("get cell: %v", err)
+	}
+	if !exists {
+		t.Fatal("tombstone must not delete the cell — an older write could then resurrect the value")
+	}
+	if cell.Version != 2 || cell.Record != "rec-v2-tombstone" {
+		t.Fatalf("cell should hold the tombstone record at v2, got %+v", cell)
+	}
+
+	// And the high-water mark still rejects the superseded write.
+	if ok, err := c.UDLUpsertIfNewer(uid, cid, key, 1, 300, "rec-v1-replay", "", false, false); err != nil || ok {
+		t.Fatalf("stale replay must be rejected: ok=%v err=%v", ok, err)
+	}
+
+	// Re-watching re-establishes the cell and puts it back in the index.
+	if ok, err := c.UDLUpsertIfNewer(uid, cid, key, 3, 400, "rec-v3", "", false, false); err != nil || !ok {
+		t.Fatalf("re-establish upsert: ok=%v err=%v", ok, err)
+	}
+	entries, err = c.UDLListUserKey(uid, key)
+	if err != nil {
+		t.Fatalf("active_for_user_key after re-establish: %v", err)
+	}
+	if len(entries) != 1 || entries[0].Cid != cid {
+		t.Fatalf("re-established cid must return to the index, got %+v", entries)
+	}
+}