perf: round 15 — grouped-listing O(N²) rebucket, exif/reseed allocs, tantivy zero-hit snippet skip

Benchmark-gated, same rule as rounds 2–14: every change ships with a
BEFORE/AFTER benchmark and an equivalence/safety gate; an AFTER that doesn't
beat its BEFORE is rolled back. The rule is encoded per harness (GATE FAIL
non-zero exit in the Rust examples, threshold expect() in vitest).

F1 — Grouped listings (trash / recent / favorites / shared-with-me)
re-bucketed the WHOLE accumulated list on every infinite-scroll page.
ResourceSectionsBuilder (new, off the reactive graph) re-buckets only the
fresh page and hands VirtualList the same rows array reference for untouched
buckets. 50×50 (2 500-item) drain: 63 750 → 2 500 bucketOf calls (25.5×),
12.5 → 1.3 ms wall (9.9×); O(N²/page) → O(N). Deep-equal to the full-rebuild
reference at every page for both a contiguous (date) and a non-contiguous
(trash-by-drive) group-by; reference-stability + fallback gated.

B1 — exif Make/Model: the display String was thrown away to allocate the
trimmed copy; display_value_trimmed trims in place (drain + truncate), 2 → 1
alloc per field (8 → 4 allocs/op, 1.26×).

B2 — content-index worker: text_extractor::supports (lowercases MIME +
extension) was called twice per file per drain batch; classify once into a
Vec<bool> and thread it through both uses. 256-file batch: 704 → 353 allocs,
34.5 → 16.7 µs (2.07×).

B3 — tantivy: skip SnippetGenerator::create on a zero-hit content search
(return Ok(vec![]) once top_docs.is_empty()); the per-hit loop was empty.
400-doc index: 1 575.6 → 1 237.2 ns (1.27×), widens with index size.

Harnesses: examples/bench_round15_micro.rs, examples/bench_round15_tantivy.rs,
frontend resourceSections.bench.test.ts; writeup in benches/ROUND15.md. Also
normalizes two round14 bench examples that were committed unformatted
(cargo fmt --all).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012o47jSrtL7xuNGTHXmtiYL

Cargo.toml

@@ -354,6 +354,24 @@ name = "bench_micro_allocs"
 path = "examples/bench_micro_allocs.rs"
 required-features = ["bench"]
 
+# Round-15 battery ────────────────────────────────────────────────────────────
+
+# Round-15 CPU/alloc micro-pack — exif Make/Model in-place trim (drop the
+# throwaway display String), content-index worker single supports() classify
+# per file (was called twice per drain batch). No Postgres.
+[[example]]
+name = "bench_round15_micro"
+path = "examples/bench_round15_micro.rs"
+required-features = ["bench"]
+
+# Round-15 tantivy zero-hit snippet skip — the SnippetGenerator::create the
+# search path builds even when the query matched no documents (pure waste on a
+# no-hit content search). Builds a RAM index; no Postgres.
+[[example]]
+name = "bench_round15_tantivy"
+path = "examples/bench_round15_tantivy.rs"
+required-features = ["bench"]
+
 # Round-14 battery ────────────────────────────────────────────────────────────
 
 # Round-14 query-shape pack — lightbox face-box narrow projection (drop the

benches/ROUND15.md

@@ -0,0 +1,166 @@
+# Round 15 — grouped-listing O(N²) rebucket, exif/reseed allocations, tantivy zero-hit snippet skip
+
+Benchmark-gated, same rule as ROUND2–14: every change ships with a
+BEFORE/AFTER benchmark and an equivalence/safety gate; an AFTER that doesn't
+beat its BEFORE is rolled back (never applied). The roll-back rule is encoded
+directly into each harness as a `GATE FAIL … rollback` non-zero exit (Rust) or
+a threshold `expect()` (frontend), so a regression fails CI rather than
+shipping.
+
+This round lands the ROUND14-deferred **flagship** — the grouped-listing
+`sections` rebuild that was the last O(N²)-per-page accumulation left in the
+SvelteKit listing surfaces — plus three backend items pulled from the same
+deferred list: two allocation cuts on the photo-ingest / content-reseed worker
+paths, and a wasted `SnippetGenerator` build removed from the zero-hit content
+search path.
+
+Measured on 4 cores / 15 GiB, local PostgreSQL not needed for any Round-15 arm
+(all no-Postgres: release profile for the Rust examples; Node 22 / vitest for
+the frontend). Reproduce any row with the command in its section.
+
+## Summary
+
+| # | change | key metric | before → after |
+|--:|---|---|---|
+| F1 | Grouped listings (trash / recent / favorites / shared-with-me) re-bucketed the WHOLE accumulated list on every infinite-scroll page; `ResourceSectionsBuilder` re-buckets only the fresh page and reuses each untouched bucket's array reference | 50×50 (2 500-item) drain, month buckets | **63 750 → 2 500 `bucketOf` calls (25.5×)** · **12.5 → 1.3 ms wall (9.9×)** · O(N²/page) → O(N) |
+| B1 | exif `Make`/`Model` — `display_value().to_string().trim_matches('"').trim().to_string()` throws the display `String` away to allocate the trimmed copy; the in-place `drain`+`truncate` helper keeps one allocation | 4 sample values/op | **150.4 → 119.4 ns (1.26×)** · **8 → 4 allocs/op** (2 → 1 per field) |
+| B2 | Content-index worker called `text_extractor::supports` (lowercases MIME + extension) TWICE per file per drain batch; classify once into a `Vec<bool>` and thread it through both uses | 256-file reseed batch | **34.5 → 16.7 µs (2.07×)** · **704 → 353 allocs/op** |
+| B3 | Zero-hit content search still built a `SnippetGenerator` (query-compile + term weighting) though the per-hit loop was empty; return `Ok(vec![])` as soon as `top_docs.is_empty()` | no-hit query, 400-doc index | **1 575.6 → 1 237.2 ns (1.27×)** · **21 → 19 allocs/op** (widens with index size) |
+
+## [F1] Grouped listings — incremental swimlane builder
+
+```
+cd frontend && npx vitest run src/lib/utils/resourceSections.bench.test.ts
+```
+
+Every grouped listing page (`/trash`, `/recent`, `/favorites`,
+`/shared-with-me`) loads its rows via infinite scroll (`raw = [...raw,
+...page]`), and `ResourceList`'s `sections` `$derived.by` re-bucketed the
+**whole accumulated list** on every page: Σ ≈ O(N²/page) `bucketOf` + `ctxOf`
+calls across a drain, and a brand-new rows array for *every* bucket each page
+(so `VirtualList` re-diffed every swimlane every page). This was the ROUND14
+"flagship follow-up" — the same O(N²)-per-page class ROUND6 fixed for the files
+listing, ROUND14 §F2 for favorites' `favoriteIds`, and `PhotoTimeline` for the
+photos grid.
+
+`ResourceSectionsBuilder` (extracted to `$lib/utils/resourceSections`, off the
+Svelte reactive graph so it's unit/benchmark-testable) exploits the append
+invariant: a grouped listing is server-sorted by the active group's `orderBy`,
+so a fresh page only ever extends existing buckets or appends new ones. It
+detects the append (prefix-identity on the boundary object), re-buckets only
+the fresh page, and hands back the **same array reference** for every untouched
+bucket — the property `VirtualList` (which diffs its `items` prop by reference)
+relies on to skip re-rendering it — while emitting a fresh array only for
+buckets the page actually grew. Any non-append (group-by switch, deletion,
+dotfile-filter toggle) falls back to a full rebuild, so the output is always
+deep-equal to the pure `buildResourceSections` reference.
+
+Correctness does **not** depend on bucket contiguity: the one non-monotonic
+group-by in the set — trash grouped **by drive** but ordered by name, so a page
+sprays items across every already-emitted drive bucket — stays byte-for-byte
+equal to the full rebuild (it just refreshes more buckets per page). Header
+labels are recomputed every sync (never cached): a group-by's `labelOf` can
+resolve asynchronously (owner / sharer names arrive after the rows), and a
+cached label would freeze the header at its fallback.
+
+50×50 (2 500-item) month-bucketed drain: **63 750 → 2 500 `bucketOf` calls
+(25.5× fewer), 12.5 → 1.3 ms wall (9.9×)**. Gates: (1) equivalence — the
+incremental output is deep-equal to `buildResourceSections` at *every* page for
+both a contiguous (date) and a non-contiguous (drive) group-by; (2) reference
+stability — untouched buckets keep their exact array reference across an append
+while a grown bucket gets a fresh one; (3) correct fallback on group-by switch,
+deletion and the flat pass-through; (4) perf — `bucketOf` work is exactly O(N)
+across the drain and wall drops ≥3×.
+
+## [B1]–[B2] exif / content-reseed allocation cuts
+
+```
+cargo run --release --features bench --example bench_round15_micro
+```
+
+Counting-allocator micro-bench; each section is BEFORE (the shipped-before
+shape) vs AFTER (the shipped function / shape) with a byte-identity gate.
+
+- **[B1] exif `Make`/`Model` single-allocation trim.** `ExifService::extract`
+  read the camera make + model as
+  `field.display_value().to_string().trim_matches('"').trim().to_string()` —
+  the first `to_string()` materializes the display value (unavoidable), then
+  `.trim_matches('"').trim().to_string()` allocates a **second** `String` for
+  the trimmed copy and drops the first. The new `display_value_trimmed` applies
+  the same two-stage trim in place on the already-owned buffer (`drain` drops
+  the stripped prefix, `truncate` the suffix — both reuse the allocation), so a
+  quoted `"Canon"` costs one allocation instead of two. Per ingested photo (the
+  Make + Model fields). 4 sample values/op: **8 → 4 allocs/op (2 → 1 per
+  field), 1.26× wall**. Gate: byte-identical to the old chain across quoted /
+  padded / clean shapes.
+- **[B2] Content-index worker single `supports()` classify.** `supports`
+  lowercases the MIME (and, on a generic MIME, the extension) — 1–2 allocations
+  — and the drain loop called it **twice per file**: once in the
+  `wanted_hashes` filter, once again in the per-file records loop. The worker
+  now classifies each file once into a `Vec<bool>` and threads the flag through
+  both. On a full reseed that is one redundant classify (and its allocations)
+  removed for *every file in the library*. 256-file batch: **704 → 353
+  allocs/op, 34.5 → 16.7 µs (2.07×)**. Gate: the `(wanted, supported)` tallies
+  are identical before/after.
+
+## [B3] Tantivy zero-hit snippet skip
+
+```
+cargo run --release --features bench --example bench_round15_tantivy
+```
+
+`TantivyContentIndex::search_blocking` built the `SnippetGenerator` from the
+query right after the `TopDocs` search — but `SnippetGenerator::create`
+compiles the query against the index (collects query terms, looks up each
+term's document frequency, builds the weighting), and when the query matched
+**no** documents that generator is never used: the per-hit loop is empty. A
+content search for a term that isn't in any indexed document (a common miss)
+paid that build for nothing on the request path.
+
+The fix returns `Ok(Vec::new())` the moment `top_docs.is_empty()`, before the
+create. The bench reproduces the exact skipped operation on a RAM index built
+with the public tantivy API (same crate + version): BEFORE = search + create,
+AFTER = search + the `is_empty()` early return; the delta is the wasted create.
+No-hit query against a 400-document index: **1 575.6 → 1 237.2 ns (1.27×), 21 →
+19 allocs/op** — the create adds ~338 ns + 2 allocs on top of the search on
+*every* zero-hit content query, and its per-term `doc_freq` lookups grow with
+the index (the RAM bench's 400 docs understate the production term dictionary).
+Gates: the miss query genuinely returns zero hits, and a control arm confirms a
+term that *does* hit still yields a snippet fragment (the skip only ever
+triggers on a true zero-hit query).
+
+## Not shipped — carried forward from the ROUND14 deferred list
+
+Still queued, unchanged in scope (each wants its own decision, Postgres
+fixture, or bigger refactor):
+
+- **Query-shape (needs Postgres):** `music_storage_adapter::list_public_playlists`
+  1 + N `COUNT(*)` fold (opt-in public-gallery path); contact REST listings
+  (`search_contacts`, `get_contacts_by_address_book_paginated`,
+  `get_contacts_in_group`) over-fetch the multi-KB `vcard` TEXT though every
+  caller maps to a `ContactDto` with no `vcard` field (wants a *lite* row
+  mapper, since the non-paginated sibling is shared with the CardDAV stream).
+- **Frontend:** `shared/+page.svelte` rebuilds the full `lanes` tree per page
+  and on every grant edit; the same incremental-builder pattern F1 uses is the
+  follow-up. (F1 removed the `ResourceList.sections` half of the ROUND14
+  "flagship" bullet; the `lanes` half remains.)
+- **CPU/alloc (background):** REST calendar-event edit re-`format!`s the whole
+  `ical_data` body once per changed property; `dedup_service` hash-`String`
+  re-allocations; `exif_service` still double-allocates the GPS-ref display in
+  `parse_gps_coord` (single-alloc, low-frequency — folded into B1's helper is
+  possible but the ref is compared to `"S"`/`"W"` as a borrow, so it never
+  needed the second alloc the Make/Model path did).
+- **Storage I/O (cached-remote class):** `CachedBlobBackend` per-write
+  `create_dir_all` + inline eviction `remove_file` on the reactor thread;
+  `encrypted_blob_backend` 64 KiB vs 256 KiB plaintext frames.
+
+## Environment / methodology
+
+- `cargo run --release --features bench --example bench_round15_micro`
+  — counting allocator, no Postgres (`BENCH_ITERS`, `BENCH_BATCH`).
+- `cargo run --release --features bench --example bench_round15_tantivy`
+  — builds a RAM tantivy index, no Postgres (`BENCH_ITERS`, `BENCH_DOCS`).
+- `cd frontend && npx vitest run src/lib/utils/resourceSections.bench.test.ts`.
+- Roll-back rule encoded per harness: the Rust examples `std::process::exit(1)`
+  with `GATE FAIL … rollback` if an AFTER arm fails to beat its BEFORE; the
+  vitest gate `expect()`s the O(N) call count and the ≥3× wall.

examples/bench_round14_micro.rs

@@ -64,7 +64,10 @@ unsafe impl GlobalAlloc for CountingAlloc {
 static GLOBAL: CountingAlloc = CountingAlloc;
 
 fn env_or<T: std::str::FromStr>(key: &str, default: T) -> T {
-    env::var(key).ok().and_then(|v| v.parse().ok()).unwrap_or(default)
+    env::var(key)
+        .ok()
+        .and_then(|v| v.parse().ok())
+        .unwrap_or(default)
 }
 
 struct Measured {
@@ -80,11 +83,17 @@ fn measure<F: FnMut()>(iters: usize, mut f: F) -> Measured {
     }
     let wall = t.elapsed().as_nanos() as f64 / iters as f64;
     let allocs = (ALLOC_CALLS.load(Ordering::Relaxed) - a0) as f64 / iters as f64;
-    Measured { wall_ns_per_op: wall, allocs_per_op: allocs }
+    Measured {
+        wall_ns_per_op: wall,
+        allocs_per_op: allocs,
+    }
 }
 
 fn print_row(label: &str, m: &Measured) {
-    println!("| {:<40} | {:>12.1} | {:>10.2} |", label, m.wall_ns_per_op, m.allocs_per_op);
+    println!(
+        "| {:<40} | {:>12.1} | {:>10.2} |",
+        label, m.wall_ns_per_op, m.allocs_per_op
+    );
 }
 
 fn header_footer(name: &str, before: &Measured, after: &Measured) {
@@ -107,12 +116,15 @@ fn section_cookie() {
 
     let iters: usize = env_or("BENCH_ITERS", 200_000);
     let name = "oxicloud_access";
-    let jwt = "eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIwMTIzNDU2Nzg5YWJjZGVmIn0.c2lnbmF0dXJlLXBsYWNlaG9sZGVy";
+    let jwt =
+        "eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIwMTIzNDU2Nzg5YWJjZGVmIn0.c2lnbmF0dXJlLXBsYWNlaG9sZGVy";
     let mut headers = HeaderMap::new();
     headers.insert(
         header::COOKIE,
-        HeaderValue::from_str(&format!("{name}={jwt}; oxicloud_csrf=3f2504e0-4f89-41d3-9a0c-0305e82c3301"))
-            .unwrap(),
+        HeaderValue::from_str(&format!(
+            "{name}={jwt}; oxicloud_csrf=3f2504e0-4f89-41d3-9a0c-0305e82c3301"
+        ))
+        .unwrap(),
     );
 
     // Gate: byte-identical value.
@@ -221,7 +233,10 @@ fn section_relevance() {
             "relevance differs for ({name:?}, {q:?})"
         );
     }
-    println!("# [A2] gate: ASCII fast path matches Unicode lowercase across {} cases — OK", corpus.len());
+    println!(
+        "# [A2] gate: ASCII fast path matches Unicode lowercase across {} cases — OK",
+        corpus.len()
+    );
 
     let m_before = measure(iters, || {
         for (name, q) in corpus {
@@ -234,7 +249,10 @@ fn section_relevance() {
         }
     });
 
-    println!("\n## [A2] compute_relevance over a {}-row result page (per search / keystroke)", corpus.len());
+    println!(
+        "\n## [A2] compute_relevance over a {}-row result page (per search / keystroke)",
+        corpus.len()
+    );
     header_footer("relevance whole corpus", &m_before, &m_after);
     if m_after.wall_ns_per_op >= m_before.wall_ns_per_op {
         eprintln!("GATE FAIL [A2]: ASCII fast path not faster — rollback");
@@ -252,7 +270,11 @@ fn section_sub_parse() {
     let pre_parsed = Uuid::parse_str(&sub).unwrap();
 
     // Gate: the pre-parsed uuid equals a fresh parse.
-    assert_eq!(Uuid::parse_str(&sub).unwrap(), pre_parsed, "uuid parse differs");
+    assert_eq!(
+        Uuid::parse_str(&sub).unwrap(),
+        pre_parsed,
+        "uuid parse differs"
+    );
     println!("# [A3] gate: pre-parsed sub_id equals per-request parse — OK");
 
     let m_before = measure(iters, || {
@@ -279,18 +301,41 @@ fn section_sub_parse() {
 fn build_request_headers() -> HeaderMap {
     // A representative authed browser request.
     let mut h = HeaderMap::new();
-    h.insert(header::AUTHORIZATION, HeaderValue::from_static("Bearer eyJhbGciOiJIUzI1NiJ9.payload.sig"));
+    h.insert(
+        header::AUTHORIZATION,
+        HeaderValue::from_static("Bearer eyJhbGciOiJIUzI1NiJ9.payload.sig"),
+    );
     h.insert(
         header::COOKIE,
-        HeaderValue::from_static("oxicloud_access=eyJ.payload.sig; oxicloud_csrf=3f2504e0-4f89-41d3-9a0c-0305e82c3301"),
+        HeaderValue::from_static(
+            "oxicloud_access=eyJ.payload.sig; oxicloud_csrf=3f2504e0-4f89-41d3-9a0c-0305e82c3301",
+        ),
     );
     h.insert(header::HOST, HeaderValue::from_static("cloud.example.com"));
-    h.insert(header::USER_AGENT, HeaderValue::from_static("Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36"));
-    h.insert(header::ACCEPT, HeaderValue::from_static("application/json, text/plain, */*"));
-    h.insert(header::ACCEPT_ENCODING, HeaderValue::from_static("gzip, deflate, br"));
-    h.insert(header::ACCEPT_LANGUAGE, HeaderValue::from_static("en-US,en;q=0.9"));
-    h.insert(header::REFERER, HeaderValue::from_static("https://cloud.example.com/files"));
-    h.insert("x-csrf-token", HeaderValue::from_static("3f2504e0-4f89-41d3-9a0c-0305e82c3301"));
+    h.insert(
+        header::USER_AGENT,
+        HeaderValue::from_static("Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36"),
+    );
+    h.insert(
+        header::ACCEPT,
+        HeaderValue::from_static("application/json, text/plain, */*"),
+    );
+    h.insert(
+        header::ACCEPT_ENCODING,
+        HeaderValue::from_static("gzip, deflate, br"),
+    );
+    h.insert(
+        header::ACCEPT_LANGUAGE,
+        HeaderValue::from_static("en-US,en;q=0.9"),
+    );
+    h.insert(
+        header::REFERER,
+        HeaderValue::from_static("https://cloud.example.com/files"),
+    );
+    h.insert(
+        "x-csrf-token",
+        HeaderValue::from_static("3f2504e0-4f89-41d3-9a0c-0305e82c3301"),
+    );
     h.insert(header::CONNECTION, HeaderValue::from_static("keep-alive"));
     h
 }
@@ -302,9 +347,13 @@ fn section_headermap_clone() {
     // Gate: the token extracted from a cloned map equals that from the borrowed map.
     let from_clone = {
         let c = headers.clone();
-        c.get(header::AUTHORIZATION).and_then(|v| v.to_str().ok()).map(str::to_string)
+        c.get(header::AUTHORIZATION)
+            .and_then(|v| v.to_str().ok())
+            .map(str::to_string)
     };
-    let from_borrow = headers.get(header::AUTHORIZATION).and_then(|v| v.to_str().ok());
+    let from_borrow = headers
+        .get(header::AUTHORIZATION)
+        .and_then(|v| v.to_str().ok());
     assert_eq!(from_clone.as_deref(), from_borrow, "authorization differs");
     println!("# [A4] gate: token from cloned map == token from borrowed map — OK");
 
@@ -358,7 +407,10 @@ fn section_caldav_rfc2822() {
     }
     println!("# [A5] gate: stack rfc2822_utc byte-identical to chrono to_rfc2822 — OK");
 
-    let dts: Vec<DateTime<Utc>> = secs.iter().map(|&s| DateTime::<Utc>::from_timestamp(s, 0).unwrap()).collect();
+    let dts: Vec<DateTime<Utc>> = secs
+        .iter()
+        .map(|&s| DateTime::<Utc>::from_timestamp(s, 0).unwrap())
+        .collect();
 
     let m_before = measure(iters, || {
         for dt in &dts {
@@ -372,7 +424,10 @@ fn section_caldav_rfc2822() {
         }
     });
 
-    println!("\n## [A5] CalDAV getlastmodified render ({} events, per REPORT/PROPFIND)", secs.len());
+    println!(
+        "\n## [A5] CalDAV getlastmodified render ({} events, per REPORT/PROPFIND)",
+        secs.len()
+    );
     header_footer("rfc2822 chrono/stack", &m_before, &m_after);
     if m_after.allocs_per_op >= m_before.allocs_per_op {
         eprintln!("GATE FAIL [A5]: stack render did not remove allocations — rollback");
@@ -389,7 +444,12 @@ fn section_caldav_href_etag() {
     let base_href = "/caldav/alice/personal/";
     // A page of events (uid, id) like write_report_page iterates.
     let events: Vec<(String, Uuid)> = (0..40)
-        .map(|i| (format!("event-uid-{i:04}-abcdef@oxicloud"), Uuid::from_u128(0x1000 + i as u128)))
+        .map(|i| {
+            (
+                format!("event-uid-{i:04}-abcdef@oxicloud"),
+                Uuid::from_u128(0x1000 + i as u128),
+            )
+        })
         .collect();
 
     // Gate: reused-buffer output identical to the per-event format! pair.
@@ -426,7 +486,10 @@ fn section_caldav_href_etag() {
         }
     });
 
-    println!("\n## [A6] CalDAV per-event href + etag ({} events/page, per REPORT/PROPFIND)", events.len());
+    println!(
+        "\n## [A6] CalDAV per-event href + etag ({} events/page, per REPORT/PROPFIND)",
+        events.len()
+    );
     header_footer("href+etag per page", &m_before, &m_after);
     if m_after.allocs_per_op >= m_before.allocs_per_op {
         eprintln!("GATE FAIL [A6]: reused buffer did not reduce allocations — rollback");

examples/bench_round14_queries.rs

@@ -57,7 +57,11 @@ fn bytes_to_embedding(b: &[u8]) -> Vec<f32> {
 /// hydrate the full 10-column row (decoding the 2 KiB embedding like
 /// `row_to_face`), then filter `user_id == caller` in Rust and keep only
 /// `(id, person_id, bbox)`.
-async fn boxes_before(pool: &PgPool, file_id: Uuid, caller: Uuid) -> Vec<(Uuid, Option<Uuid>, Vec<f32>)> {
+async fn boxes_before(
+    pool: &PgPool,
+    file_id: Uuid,
+    caller: Uuid,
+) -> Vec<(Uuid, Option<Uuid>, Vec<f32>)> {
     let rows = sqlx::query(
         "SELECT id, file_id, user_id, person_id, bbox, det_score, quality, embedding, blob_hash, created_at
            FROM faces.faces WHERE file_id = $1",
@@ -83,7 +87,11 @@ async fn boxes_before(pool: &PgPool, file_id: Uuid, caller: Uuid) -> Vec<(Uuid,
 }
 
 /// AFTER: narrow projection, caller filter in SQL.
-async fn boxes_after(pool: &PgPool, file_id: Uuid, caller: Uuid) -> Vec<(Uuid, Option<Uuid>, Vec<f32>)> {
+async fn boxes_after(
+    pool: &PgPool,
+    file_id: Uuid,
+    caller: Uuid,
+) -> Vec<(Uuid, Option<Uuid>, Vec<f32>)> {
     let rows = sqlx::query(
         "SELECT id, person_id, bbox FROM faces.faces WHERE file_id = $1 AND user_id = $2",
     )
@@ -200,8 +208,12 @@ async fn section_face_boxes(pool: &PgPool) {
     let wire_after = n * (16 + 16 + 16 + 8);
     println!("\n## [Q1] Lightbox face boxes — group photo, {n} faces");
     println!("| arm | mean ms | p50 ms | p95 ms | ~bytes/req |");
-    println!("| BEFORE wide row (incl. embedding) | {wm:>7.3} | {wp50:>6.3} | {wp95:>6.3} | {wire_before:>9} |");
-    println!("| AFTER  narrow (id,person,bbox)    | {nm:>7.3} | {np50:>6.3} | {np95:>6.3} | {wire_after:>9} |");
+    println!(
+        "| BEFORE wide row (incl. embedding) | {wm:>7.3} | {wp50:>6.3} | {wp95:>6.3} | {wire_before:>9} |"
+    );
+    println!(
+        "| AFTER  narrow (id,person,bbox)    | {nm:>7.3} | {np50:>6.3} | {np95:>6.3} | {wire_after:>9} |"
+    );
     println!(
         "# {:.2}x faster; ~{} KiB embedding/columns off the wire per lightbox open (scales with face count)",
         wm / nm,

examples/bench_round15_micro.rs

@@ -0,0 +1,268 @@
+//! Round-15 CPU/alloc micro-pack (no Postgres).
+//!
+//! Each section is BEFORE (verbatim replica of the shipped-before shape) vs
+//! AFTER (the shipped-after shape, or the shipped function itself), with an
+//! equivalence gate and a `GATE FAIL … rollback` check that exits non-zero if
+//! the AFTER arm fails to beat its BEFORE — the round's roll-back rule encoded
+//! into the benchmark.
+//!
+//!   [B1] exif Make/Model — `display_value().to_string().trim_matches('"')
+//!        .trim().to_string()` allocates the display String, then throws it away
+//!        to allocate the trimmed copy (2 allocs). The shipped
+//!        `exif_service::display_value_trimmed` trims in place on the owned
+//!        buffer (`drain` + `truncate`) — 1 alloc. Per ingested photo.
+//!   [B2] content-index worker `supports()` — `text_extractor::supports`
+//!        (lowercases the MIME + extension, 1–2 allocs) was called TWICE per
+//!        file per drain batch: once in the wanted-hashes filter, once in the
+//!        records loop. The shipped code classifies each file once into a
+//!        `Vec<bool>` and threads it through both. Per reseed batch (every
+//!        file in the library).
+//!
+//! Run:
+//!   cargo run --release --features bench --example bench_round15_micro
+//! Tunables (env): BENCH_ITERS (200000), BENCH_BATCH (256)
+
+use std::alloc::{GlobalAlloc, Layout, System};
+use std::env;
+use std::hint::black_box;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::Instant;
+
+use oxicloud::infrastructure::services::search_index::text_extractor;
+
+static ALLOC_CALLS: AtomicU64 = AtomicU64::new(0);
+
+struct CountingAlloc;
+
+unsafe impl GlobalAlloc for CountingAlloc {
+    unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
+        ALLOC_CALLS.fetch_add(1, Ordering::Relaxed);
+        unsafe { System.alloc(layout) }
+    }
+    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
+        unsafe { System.dealloc(ptr, layout) }
+    }
+    unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
+        ALLOC_CALLS.fetch_add(1, Ordering::Relaxed);
+        unsafe { System.realloc(ptr, layout, new_size) }
+    }
+    unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
+        ALLOC_CALLS.fetch_add(1, Ordering::Relaxed);
+        unsafe { System.alloc_zeroed(layout) }
+    }
+}
+
+#[global_allocator]
+static GLOBAL: CountingAlloc = CountingAlloc;
+
+fn env_or<T: std::str::FromStr>(key: &str, default: T) -> T {
+    env::var(key)
+        .ok()
+        .and_then(|v| v.parse().ok())
+        .unwrap_or(default)
+}
+
+struct Measured {
+    wall_ns_per_op: f64,
+    allocs_per_op: f64,
+}
+
+fn measure<F: FnMut()>(iters: usize, mut f: F) -> Measured {
+    let a0 = ALLOC_CALLS.load(Ordering::Relaxed);
+    let t = Instant::now();
+    for _ in 0..iters {
+        f();
+    }
+    let wall = t.elapsed().as_nanos() as f64 / iters as f64;
+    let allocs = (ALLOC_CALLS.load(Ordering::Relaxed) - a0) as f64 / iters as f64;
+    Measured {
+        wall_ns_per_op: wall,
+        allocs_per_op: allocs,
+    }
+}
+
+fn print_row(label: &str, m: &Measured) {
+    println!(
+        "| {:<42} | {:>12.1} | {:>10.2} |",
+        label, m.wall_ns_per_op, m.allocs_per_op
+    );
+}
+
+fn header_footer(name: &str, before: &Measured, after: &Measured) {
+    println!("| arm | ns/op | allocs/op |");
+    print_row(&format!("BEFORE {name}"), before);
+    print_row(&format!("AFTER  {name}"), after);
+    println!(
+        "# {:.2}x wall, {:.2} fewer allocs/op",
+        before.wall_ns_per_op / after.wall_ns_per_op,
+        before.allocs_per_op - after.allocs_per_op
+    );
+}
+
+// ────────────────────────────────────────────────────────────────────────────
+// [B1] exif Make/Model trim — 2 allocs (throwaway display String) vs 1 (in place)
+// ────────────────────────────────────────────────────────────────────────────
+
+/// BEFORE: the shipped-before chain. `raw` stands in for the field's rendered
+/// display value; `to_string()` mirrors `display_value().to_string()` (the one
+/// unavoidable alloc), then `.trim_matches('"').trim().to_string()` allocates a
+/// second time for the trimmed copy.
+fn trim_before(raw: &str) -> String {
+    raw.to_string().trim_matches('"').trim().to_string()
+}
+
+/// AFTER: verbatim replica of `exif_service::display_value_trimmed` — trims in
+/// place on the already-owned buffer, so only the display String is allocated.
+fn trim_after(raw: &str) -> String {
+    let mut s = raw.to_string();
+    let trimmed = s.trim_matches('"').trim();
+    let start = trimmed.as_ptr().addr() - s.as_ptr().addr();
+    let len = trimmed.len();
+    s.drain(..start);
+    s.truncate(len);
+    s
+}
+
+fn section_exif_trim() {
+    let iters: usize = env_or("BENCH_ITERS", 200_000);
+    // Representative EXIF Make/Model display values: the widely-seen quoted
+    // form, plus a padded one and an already-clean one.
+    let samples = ["\"Canon\"", "\"NIKON CORPORATION\"", "  Apple  ", "SONY"];
+
+    // Gate: byte-identical output to the old chain across every shape.
+    for s in samples {
+        assert_eq!(trim_before(s), trim_after(s), "trim differs for {s:?}");
+    }
+
+    let before = measure(iters, || {
+        for s in samples {
+            black_box(trim_before(black_box(s)));
+        }
+    });
+    let after = measure(iters, || {
+        for s in samples {
+            black_box(trim_after(black_box(s)));
+        }
+    });
+
+    println!("\n## [B1] exif Make/Model trim (4 sample values/op)");
+    header_footer("exif trim", &before, &after);
+    if after.allocs_per_op >= before.allocs_per_op {
+        eprintln!("GATE FAIL [B1]: in-place trim did not reduce allocations — rollback");
+        std::process::exit(1);
+    }
+}
+
+// ────────────────────────────────────────────────────────────────────────────
+// [B2] content-index worker supports() — 2× per file vs 1× (memoized)
+// ────────────────────────────────────────────────────────────────────────────
+
+/// One drained file row: (name, mime, size). Mirrors the worker's
+/// `FileIndexRow` projection (only the fields `supports` + the size gate read).
+struct FileRow {
+    name: &'static str,
+    mime: &'static str,
+    size: i64,
+}
+
+fn corpus(n: usize) -> Vec<FileRow> {
+    // A realistic reseed mix: text/markdown/pdf/office (supported) interleaved
+    // with images/video/binaries (unsupported — the fast reject).
+    const MIX: &[(&str, &str, i64)] = &[
+        ("notes.txt", "text/plain", 4_000),
+        ("readme.md", "text/markdown", 8_000),
+        ("report.pdf", "application/pdf", 250_000),
+        ("photo.jpg", "image/jpeg", 3_000_000),
+        (
+            "sheet.xlsx",
+            "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
+            120_000,
+        ),
+        ("clip.mp4", "video/mp4", 40_000_000),
+        ("data.bin", "application/octet-stream", 1_000),
+        ("page.html", "text/html; charset=utf-8", 20_000),
+    ];
+    (0..n)
+        .map(|i| {
+            let (name, mime, size) = MIX[i % MIX.len()];
+            FileRow { name, mime, size }
+        })
+        .collect()
+}
+
+fn section_supports() {
+    let iters: usize = env_or("BENCH_ITERS", 200_000) / 20; // heavier op
+    let batch: usize = env_or("BENCH_BATCH", 256);
+    let max_bytes: u64 = 10 * 1024 * 1024;
+    let files = corpus(batch);
+
+    // BEFORE: `supports` is evaluated in the wanted-hashes filter AND again per
+    // file in the records loop — twice per file.
+    let run_before = |files: &[FileRow]| -> (usize, usize) {
+        let wanted = files
+            .iter()
+            .filter(|f| text_extractor::supports(f.name, f.mime) && f.size as u64 <= max_bytes)
+            .count();
+        let mut supported_files = 0;
+        for f in files {
+            if text_extractor::supports(f.name, f.mime) {
+                supported_files += 1;
+            }
+        }
+        (wanted, supported_files)
+    };
+
+    // AFTER: classify each file once into a `Vec<bool>`; both the filter and the
+    // records loop read the flag.
+    let run_after = |files: &[FileRow]| -> (usize, usize) {
+        let supported: Vec<bool> = files
+            .iter()
+            .map(|f| text_extractor::supports(f.name, f.mime))
+            .collect();
+        let wanted = files
+            .iter()
+            .zip(&supported)
+            .filter(|&(f, s)| *s && f.size as u64 <= max_bytes)
+            .count();
+        let mut supported_files = 0;
+        for (_, &s) in files.iter().zip(&supported) {
+            if s {
+                supported_files += 1;
+            }
+        }
+        (wanted, supported_files)
+    };
+
+    // Gate: identical (wanted, supported) tallies.
+    assert_eq!(
+        run_before(&files),
+        run_after(&files),
+        "supports tally differs"
+    );
+
+    let before = measure(iters, || {
+        black_box(run_before(black_box(&files)));
+    });
+    let after = measure(iters, || {
+        black_box(run_after(black_box(&files)));
+    });
+
+    println!("\n## [B2] content-index supports() ({batch} files/batch)");
+    header_footer("supports/batch", &before, &after);
+    if after.allocs_per_op >= before.allocs_per_op || after.wall_ns_per_op >= before.wall_ns_per_op
+    {
+        eprintln!("GATE FAIL [B2]: single-classify did not beat the double call — rollback");
+        std::process::exit(1);
+    }
+}
+
+fn main() {
+    println!("#################################################################");
+    println!("# Round-15 CPU/alloc micro-pack");
+    println!("#################################################################");
+
+    section_exif_trim();
+    section_supports();
+
+    println!("\nGATE PASS (all sections)");
+}

examples/bench_round15_tantivy.rs

@@ -0,0 +1,224 @@
+//! Round-15 tantivy zero-hit snippet skip (no Postgres).
+//!
+//! `TantivyContentIndex::search_blocking` builds a `SnippetGenerator` from the
+//! query right after the `TopDocs` search — but a `SnippetGenerator::create`
+//! compiles the query against the index (term lookups + weight build), and when
+//! the query matched NO documents that generator is never used (the per-hit
+//! loop is empty). The shipped fix returns `Ok(Vec::new())` as soon as
+//! `top_docs.is_empty()`, before the create.
+//!
+//! This bench reproduces the exact skipped operation on a RAM index built with
+//! the public tantivy API (same crate + version the service uses):
+//!   BEFORE = search (→ 0 hits) + `SnippetGenerator::create` (+ `set_max_num_chars`)
+//!   AFTER  = search (→ 0 hits) + `top_docs.is_empty()` early return
+//! The delta is the wasted create the fix removes from every no-hit content
+//! search. A sanity arm confirms a term that DOES hit still yields a snippet, so
+//! the skip only ever triggers on a genuine zero-hit query.
+//!
+//! Run:
+//!   cargo run --release --features bench --example bench_round15_tantivy
+//! Tunables (env): BENCH_ITERS (50000), BENCH_DOCS (400)
+
+use std::alloc::{GlobalAlloc, Layout, System};
+use std::env;
+use std::hint::black_box;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::Instant;
+
+use tantivy::collector::TopDocs;
+use tantivy::query::QueryParser;
+use tantivy::schema::{STORED, STRING, Schema, TEXT, Value as _};
+use tantivy::snippet::SnippetGenerator;
+use tantivy::{Index, TantivyDocument, doc};
+
+static ALLOC_CALLS: AtomicU64 = AtomicU64::new(0);
+
+struct CountingAlloc;
+
+unsafe impl GlobalAlloc for CountingAlloc {
+    unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
+        ALLOC_CALLS.fetch_add(1, Ordering::Relaxed);
+        unsafe { System.alloc(layout) }
+    }
+    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
+        unsafe { System.dealloc(ptr, layout) }
+    }
+    unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
+        ALLOC_CALLS.fetch_add(1, Ordering::Relaxed);
+        unsafe { System.realloc(ptr, layout, new_size) }
+    }
+    unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
+        ALLOC_CALLS.fetch_add(1, Ordering::Relaxed);
+        unsafe { System.alloc_zeroed(layout) }
+    }
+}
+
+#[global_allocator]
+static GLOBAL: CountingAlloc = CountingAlloc;
+
+fn env_or<T: std::str::FromStr>(key: &str, default: T) -> T {
+    env::var(key)
+        .ok()
+        .and_then(|v| v.parse().ok())
+        .unwrap_or(default)
+}
+
+const SNIPPET_MAX_CHARS: usize = 200;
+
+fn main() {
+    let iters: usize = env_or("BENCH_ITERS", 50_000);
+    let docs: usize = env_or("BENCH_DOCS", 400);
+
+    println!("#################################################################");
+    println!("# Round-15 tantivy zero-hit snippet skip");
+    println!("#################################################################");
+
+    // ── Build a RAM index: a stored content field + a name field, the shape
+    //    the service indexes. Fill it with realistic prose so create() has real
+    //    terms to weigh. ────────────────────────────────────────────────────
+    let mut schema_builder = Schema::builder();
+    let name = schema_builder.add_text_field("name", STRING | STORED);
+    let content = schema_builder.add_text_field("content", TEXT | STORED);
+    let schema = schema_builder.build();
+    let index = Index::create_in_ram(schema);
+
+    const WORDS: &[&str] = &[
+        "informe",
+        "trimestral",
+        "ventas",
+        "region",
+        "norte",
+        "presupuesto",
+        "reunion",
+        "proyecto",
+        "cliente",
+        "factura",
+        "contrato",
+        "entrega",
+        "calendario",
+        "documento",
+        "resumen",
+        "analisis",
+        "resultados",
+        "equipo",
+    ];
+    {
+        let mut writer = index.writer(15_000_000).expect("writer");
+        for i in 0..docs {
+            let body: String = (0..40)
+                .map(|j| WORDS[(i * 7 + j * 13) % WORDS.len()])
+                .collect::<Vec<_>>()
+                .join(" ");
+            writer
+                .add_document(doc!(
+                    name => format!("doc-{i}.txt"),
+                    content => body,
+                ))
+                .expect("add");
+        }
+        writer.commit().expect("commit");
+    }
+    let reader = index.reader().expect("reader");
+    let searcher = reader.searcher();
+    let parser = QueryParser::for_index(&index, vec![content]);
+
+    // A multi-term query of words that appear in NO document → zero hits, but
+    // valid tokens (so the real code reaches the search, not the empty-token
+    // guard). These are plausible-but-absent search terms.
+    let miss_query = parser
+        .parse_query("zzznonexistent quuxfoobar wibblewobble")
+        .expect("parse");
+    // A query that DOES hit — the sanity arm.
+    let hit_query = parser.parse_query("informe ventas").expect("parse");
+
+    // ── Correctness gates ──────────────────────────────────────────────────
+    let miss_hits = searcher
+        .search(&miss_query, &TopDocs::with_limit(32).order_by_score())
+        .expect("search");
+    assert!(
+        miss_hits.is_empty(),
+        "miss query must return zero hits (got {})",
+        miss_hits.len()
+    );
+
+    let hit_hits = searcher
+        .search(&hit_query, &TopDocs::with_limit(32).order_by_score())
+        .expect("search");
+    assert!(!hit_hits.is_empty(), "hit query must return hits");
+    // The generator the fix keeps for real hits still produces a fragment.
+    let generator = SnippetGenerator::create(&searcher, &*hit_query, content).expect("gen");
+    let (_, addr) = hit_hits[0];
+    let d: TantivyDocument = searcher.doc(addr).expect("doc");
+    let preview = d
+        .get_first(content)
+        .and_then(|v| v.as_str())
+        .unwrap_or_default()
+        .to_owned();
+    assert!(
+        !generator.snippet(&preview).fragment().is_empty(),
+        "a real hit must still yield a snippet fragment"
+    );
+
+    // ── BEFORE: search + build the snippet generator even on zero hits. ──────
+    let a0 = ALLOC_CALLS.load(Ordering::Relaxed);
+    let t = Instant::now();
+    for _ in 0..iters {
+        let top = searcher
+            .search(&miss_query, &TopDocs::with_limit(32).order_by_score())
+            .expect("search");
+        let sg = SnippetGenerator::create(&searcher, &*miss_query, content).map(|mut g| {
+            g.set_max_num_chars(SNIPPET_MAX_CHARS);
+            g
+        });
+        black_box((top.len(), sg.is_ok()));
+    }
+    let before_ns = t.elapsed().as_nanos() as f64 / iters as f64;
+    let before_allocs = (ALLOC_CALLS.load(Ordering::Relaxed) - a0) as f64 / iters as f64;
+
+    // ── AFTER: search + the shipped early return on an empty result. ─────────
+    let a1 = ALLOC_CALLS.load(Ordering::Relaxed);
+    let t = Instant::now();
+    for _ in 0..iters {
+        let top = searcher
+            .search(&miss_query, &TopDocs::with_limit(32).order_by_score())
+            .expect("search");
+        if top.is_empty() {
+            black_box(top.len());
+            continue;
+        }
+        // Unreached for the miss query; present so the arm is structurally the
+        // shipped code, not a stripped one.
+        let sg = SnippetGenerator::create(&searcher, &*miss_query, content).map(|mut g| {
+            g.set_max_num_chars(SNIPPET_MAX_CHARS);
+            g
+        });
+        black_box(sg.is_ok());
+    }
+    let after_ns = t.elapsed().as_nanos() as f64 / iters as f64;
+    let after_allocs = (ALLOC_CALLS.load(Ordering::Relaxed) - a1) as f64 / iters as f64;
+
+    println!("\n## zero-hit content search ({docs} docs indexed)");
+    println!("| arm | ns/op | allocs/op |");
+    println!(
+        "| {:<40} | {:>12.1} | {:>10.2} |",
+        "BEFORE search + snippet create", before_ns, before_allocs
+    );
+    println!(
+        "| {:<40} | {:>12.1} | {:>10.2} |",
+        "AFTER  search + is_empty skip", after_ns, after_allocs
+    );
+    println!(
+        "# {:.2}x wall, {:.2} fewer allocs/op",
+        before_ns / after_ns,
+        before_allocs - after_allocs
+    );
+
+    if after_ns >= before_ns || after_allocs >= before_allocs {
+        eprintln!(
+            "GATE FAIL [B3]: zero-hit skip did not beat building the snippet generator — rollback"
+        );
+        std::process::exit(1);
+    }
+
+    println!("\nGATE PASS");
+}

frontend/src/lib/components/ResourceList.svelte

@@ -78,6 +78,7 @@
 	import { formatBytes } from '$lib/utils/format';
 	import { formatDate, iconNameFromClass, fileIconKindClass } from '$lib/utils/display';
 	import { gridColumns } from '$lib/utils/grid';
+	import { ResourceSectionsBuilder } from '$lib/utils/resourceSections';
 	import { fileThumbnailUrl, thumbSizeForView } from '$lib/api/endpoints/files';
 	import {
 		canThumbnailClientSide,
@@ -369,29 +370,24 @@
 	/**
 	 * Partition the visible items into grouped sections when a `bucketOf` is
 	 * active. Server order is preserved within and across buckets (first-seen).
+	 *
+	 * `ResourceSectionsBuilder` re-buckets only the freshly-appended page rather
+	 * than the whole accumulated list, and hands `VirtualList` the same rows
+	 * array reference for every untouched bucket so it skips re-rendering it. An
+	 * infinite-scroll drain of a grouped listing (trash / recent / favorites /
+	 * shared-with-me) collapses from Σ O(N²/page) to O(N) bucketing work
+	 * (benches/ROUND15.md §F1). Held off the reactive graph — a plain
+	 * accumulator keyed by the append cursor, not $state; `sync` is idempotent,
+	 * so if the derive re-fires without an actual append it safely full-rebuilds
+	 * to the same output the pure `buildResourceSections` reference produces.
 	 */
-	const sections = $derived.by(
-		(): Array<{ key: string; label: string; rows: Array<FileItem | FolderItem> }> => {
-			const bucketOf = activeGroup?.bucketOf;
-			if (!bucketOf) return [{ key: '', label: '', rows: visibleItems }];
-			const order: string[] = [];
-			// Transient bucketing map computed inside $derived.by — not reactive state.
-			// eslint-disable-next-line svelte/prefer-svelte-reactivity
-			const map = new Map<string, Array<FileItem | FolderItem>>();
-			for (const item of visibleItems) {
-				const k = bucketOf(item, ctxOf(item.id)) ?? '∅';
-				if (!map.has(k)) {
-					map.set(k, []);
-					order.push(k);
-				}
-				map.get(k)!.push(item);
-			}
-			return order.map((k) => ({
-				key: k,
-				label: activeGroup?.labelOf?.(k) ?? k,
-				rows: map.get(k)!
-			}));
-		}
+	const sectionsBuilder = new ResourceSectionsBuilder<FileItem | FolderItem, ItemContext>();
+	const sections = $derived.by(() =>
+		sectionsBuilder.sync(visibleItems, {
+			bucketOf: activeGroup?.bucketOf,
+			labelOf: activeGroup?.labelOf,
+			ctxOf: (item) => ctxOf(item.id)
+		})
 	);
 	const grouped = $derived(!!activeGroup?.bucketOf);
 

frontend/src/lib/utils/resourceSections.bench.test.ts

@@ -0,0 +1,211 @@
+import { describe, expect, it } from 'vitest';
+import {
+	ResourceSectionsBuilder,
+	buildResourceSections,
+	type SectionGrouping
+} from './resourceSections';
+
+/**
+ * Benchmark gate for the incremental swimlane builder (ResourceSectionsBuilder)
+ * that replaced ResourceList's `sections` `$derived.by`.
+ *
+ * Audit finding (ROUND14 deferred flagship): every grouped listing (trash,
+ * recent, favorites, shared-with-me) pages in via `raw = [...raw, ...page]`,
+ * and `sections` re-bucketed the WHOLE accumulated list on every page — Σ ≈
+ * O(N²/page) `bucketOf` + `ctxOf` calls during an infinite-scroll drain, and a
+ * brand-new rows array for EVERY bucket each page (so VirtualList re-diffed
+ * every swimlane every page). The builder re-buckets only the fresh page and
+ * hands back the same array reference for untouched buckets.
+ *
+ * Gates:
+ *  1. Equivalence — at EVERY page of the drain, the incremental output is
+ *     deep-equal to the verbatim full-rebuild reference (buildResourceSections),
+ *     for a contiguous group-by (date, bucket aligned with order) AND a
+ *     non-contiguous one (trash-by-drive: name-ordered, drive-bucketed); plus
+ *     group-by switch, deletion and the flat pass-through fall back correctly.
+ *  2. Reference stability — untouched buckets keep their exact array reference
+ *     across a page append (the property VirtualList relies on to skip them),
+ *     while a grown bucket gets a fresh one.
+ *  3. Perf — bucketing work collapses from Σ O(N²/page) to O(N) across the
+ *     drain (deterministic `bucketOf`-call count) and wall drops ≥3x.
+ */
+
+interface Item {
+	id: string;
+	name: string;
+	driveId: string;
+	/** ms epoch; descending with index (newest-first, like the server pages). */
+	date: number;
+}
+
+interface Ctx {
+	date: number;
+	driveId: string;
+}
+
+const DAY = 86_400_000;
+
+/** Item `i`: newest-first date, name in a fixed lexical order, round-robin drive. */
+function item(i: number): Item {
+	return {
+		id: `it-${i.toString().padStart(6, '0')}`,
+		// Zero-padded so lexical name order is a stable, well-defined sequence.
+		name: `file-${i.toString().padStart(6, '0')}`,
+		driveId: `drive-${i % 4}`,
+		date: 1_700_000_000_000 - i * (DAY / 2)
+	};
+}
+
+const contextMap = new Map<string, Ctx>();
+function ctxOf(it: Item): Ctx | undefined {
+	let c = contextMap.get(it.id);
+	if (!c) {
+		c = { date: it.date, driveId: it.driveId };
+		contextMap.set(it.id, c);
+	}
+	return c;
+}
+
+/** Month bucket key from a ctx date (contiguous under date order). */
+function monthKey(d: number): string {
+	const dt = new Date(d);
+	return `${dt.getUTCFullYear()}-${`${dt.getUTCMonth() + 1}`.padStart(2, '0')}`;
+}
+
+/** Contiguous group-by: date-ordered pages, date buckets. Counts bucketOf calls. */
+function dateGrouping(counter?: { n: number }): SectionGrouping<Item, Ctx> {
+	return {
+		bucketOf: (_it, ctx) => {
+			if (counter) counter.n++;
+			return ctx ? monthKey(ctx.date) : null;
+		},
+		labelOf: (k) => `📅 ${k}`,
+		ctxOf
+	};
+}
+
+/**
+ * Non-contiguous group-by mirroring trash "by drive": pages arrive in NAME
+ * order but bucket by driveId, so a fresh page sprays items across every
+ * already-emitted drive bucket. Equivalence must still hold.
+ */
+function driveGrouping(counter?: { n: number }): SectionGrouping<Item, Ctx> {
+	return {
+		bucketOf: (_it, ctx) => {
+			if (counter) counter.n++;
+			return ctx ? ctx.driveId : null;
+		},
+		labelOf: (k) => `💾 ${k}`,
+		ctxOf
+	};
+}
+
+const PAGE = 50;
+const PAGES = 50; // 2 500-item drain
+
+describe('incremental resource sections (benchmark gate)', () => {
+	for (const [name, mk] of [
+		['contiguous date buckets', dateGrouping],
+		['non-contiguous drive buckets', driveGrouping]
+	] as const) {
+		it(`stays deep-equal to the full rebuild at every page — ${name}`, () => {
+			const all = Array.from({ length: PAGE * PAGES }, (_, i) => item(i));
+			const builder = new ResourceSectionsBuilder<Item, Ctx>();
+			// ONE stable grouping across the drain — mirrors the component, where
+			// `activeGroup.bucketOf` is a fixed closure from the page's once-defined
+			// `groupBys`. This is what lets the builder take its incremental path,
+			// so this loop genuinely exercises it (not the rebuild fallback).
+			const g = mk();
+			for (let p = 1; p <= PAGES; p++) {
+				const cumulative = all.slice(0, p * PAGE);
+				const incremental = builder.sync(cumulative, g);
+				const reference = buildResourceSections(cumulative, g);
+				expect(incremental, `page ${p}`).toEqual(reference);
+			}
+		});
+	}
+
+	it('keeps untouched bucket arrays reference-stable and refreshes grown ones', () => {
+		const all = Array.from({ length: 600 }, (_, i) => item(i));
+		const builder = new ResourceSectionsBuilder<Item, Ctx>();
+		const g = dateGrouping();
+
+		const first = builder.sync(all.slice(0, 300), g);
+		const refBefore = new Map(first.map((s) => [s.key, s.rows]));
+
+		const second = builder.sync(all.slice(0, 350), g);
+		let stable = 0;
+		let refreshed = 0;
+		for (const s of second) {
+			const prev = refBefore.get(s.key);
+			if (prev === undefined) continue; // brand-new bucket
+			if (prev === s.rows) stable++;
+			else refreshed++;
+		}
+		// Date-ordered append only grows the boundary bucket(s): most earlier
+		// buckets must be handed back by the SAME reference (VirtualList skips
+		// them), and at least one bucket must be refreshed (it grew).
+		expect(stable).toBeGreaterThan(0);
+		expect(refreshed).toBeGreaterThan(0);
+		expect(stable).toBeGreaterThan(refreshed);
+	});
+
+	it('falls back to a correct full rebuild on group-by switch, deletion and flat', () => {
+		const all = Array.from({ length: 600 }, (_, i) => item(i));
+		const builder = new ResourceSectionsBuilder<Item, Ctx>();
+		const byDate = dateGrouping();
+		const byDrive = driveGrouping();
+
+		// Drain a few pages under date grouping, then switch to drive grouping
+		// (a different bucketOf reference → rebuild).
+		builder.sync(all.slice(0, 300), byDate);
+		expect(builder.sync(all.slice(0, 300), byDrive)).toEqual(
+			buildResourceSections(all.slice(0, 300), byDrive)
+		);
+
+		// Deletion under the SAME grouping (list shrinks / prefix changes) →
+		// rebuild via the append check, not a grouping-ref change.
+		const shrunk = all.slice(0, 300).filter((_, i) => i % 7 !== 0);
+		expect(builder.sync(shrunk, byDrive)).toEqual(buildResourceSections(shrunk, byDrive));
+
+		// Flat pass-through (no bucketOf) yields one section and doesn't wedge the
+		// next grouped sync.
+		const flat: SectionGrouping<Item, Ctx> = { ctxOf };
+		const flatOut = builder.sync(shrunk, flat);
+		expect(flatOut).toEqual([{ key: '', label: '', rows: shrunk }]);
+		expect(flatOut[0].rows).toBe(shrunk); // pass-through, no copy
+		expect(builder.sync(shrunk, byDate)).toEqual(buildResourceSections(shrunk, byDate));
+	});
+
+	it('collapses bucketing work from Σ O(N²/page) to O(N) and runs ≥3x faster', () => {
+		const N = PAGE * PAGES;
+		const all = Array.from({ length: N }, (_, i) => item(i));
+
+		// AFTER: incremental — each item is bucketed exactly once across the drain.
+		// ONE stable grouping (fixed bucketOf), exactly as the component supplies.
+		const afterCounter = { n: 0 };
+		const gAfter = dateGrouping(afterCounter);
+		const builder = new ResourceSectionsBuilder<Item, Ctx>();
+		const t1 = performance.now();
+		for (let p = 1; p <= PAGES; p++) builder.sync(all.slice(0, p * PAGE), gAfter);
+		const afterMs = performance.now() - t1;
+
+		// BEFORE: full rebuild per page — re-buckets the whole cumulative list.
+		const beforeCounter = { n: 0 };
+		const gBefore = dateGrouping(beforeCounter);
+		const t0 = performance.now();
+		for (let p = 1; p <= PAGES; p++) buildResourceSections(all.slice(0, p * PAGE), gBefore);
+		const beforeMs = performance.now() - t0;
+
+		console.info(
+			`resource sections ${PAGES}×${PAGE}: before ${beforeCounter.n} bucketOf calls / ${beforeMs.toFixed(1)} ms — after ${afterCounter.n} calls / ${afterMs.toFixed(1)} ms (${(beforeCounter.n / afterCounter.n).toFixed(1)}x fewer calls, ${(beforeMs / afterMs).toFixed(1)}x wall)`
+		);
+
+		// Incremental buckets each item once: exactly N calls.
+		expect(afterCounter.n).toBe(N);
+		// Full rebuild is quadratic: Σ_{p=1..P} p·PAGE.
+		expect(beforeCounter.n).toBe((PAGES * (PAGES + 1) * PAGE) / 2);
+		expect(afterCounter.n).toBeLessThan(beforeCounter.n / 5);
+		expect(afterMs).toBeLessThan(beforeMs / 3);
+	});
+});

frontend/src/lib/utils/resourceSections.ts

@@ -0,0 +1,189 @@
+/**
+ * Incremental swimlane bucketing for `ResourceList`, extracted from the
+ * component so the O(N²) accumulation of its `sections` `$derived` can be
+ * replaced with an append-aware builder (and unit/benchmark-tested off the
+ * Svelte reactive graph).
+ *
+ * `ResourceList` pages its list in via infinite scroll (`raw = [...raw,
+ * ...page]`), and `sections` was `$derived` over the WHOLE accumulated list —
+ * so paging to item N re-buckets everything loaded so far, Σ ≈ O(N²/page)
+ * main-thread work during the scroll (the same class ROUND6 fixed for the
+ * files listing and ROUND14 §F2 fixed for favorites, and PhotoTimeline fixed
+ * for the photos grid).
+ *
+ * Grouped listings sort by the active group's `orderBy`, so a fresh page only
+ * ever extends existing buckets or appends new ones — it never reorders an
+ * already-emitted bucket. {@link ResourceSectionsBuilder} exploits that: an
+ * append re-buckets only the fresh page and hands back the SAME array
+ * reference for every untouched bucket (so `VirtualList`, which diffs its
+ * `items` prop by reference, skips re-rendering it) while emitting a fresh
+ * array for each bucket the page actually grew.
+ *
+ * Correctness does not depend on bucket contiguity: even a group-by whose
+ * `bucketOf` is not monotonic in server order (e.g. trash grouped by drive but
+ * ordered by name) stays byte-for-byte equal to the full rebuild — it just
+ * touches more buckets per page. The pure {@link buildResourceSections} is the
+ * verbatim reference (what the old `sections` derive produced); the benchmark
+ * gate asserts the incremental builder stays deep-equal to it at every page.
+ */
+
+/** One swimlane: a bucket key, its (possibly async-resolved) header label, and its rows. */
+export interface ResourceSection<T> {
+	key: string;
+	label: string;
+	rows: T[];
+}
+
+/**
+ * The grouping inputs the builder needs, mirroring `ResourceList`'s active
+ * `GroupByDef` plus its per-item context accessor. `bucketOf` undefined means
+ * "flat list" (a single unlabelled section). Generic over the item type `T`
+ * and the per-item context envelope `C` so the module stays independent of the
+ * component's concrete types.
+ */
+export interface SectionGrouping<T, C> {
+	/** Map an item + its context to a bucket key; null → the `∅` catch-all bucket. */
+	bucketOf?: (item: T, ctx: C | undefined) => string | null;
+	/** Map a bucket key to its header label; identity when absent. */
+	labelOf?: (key: string) => string;
+	/** Resolve an item's context envelope (e.g. `contextMap.get(item.id)`). */
+	ctxOf: (item: T) => C | undefined;
+}
+
+/** The `∅` catch-all key the old derive used for a null bucket (kept byte-identical). */
+const NULL_BUCKET = '∅';
+
+/**
+ * Verbatim reference: the `ResourceSection[]` the old `sections` `$derived.by`
+ * produced for `items` under `grouping`. Bucket order is first-appearance;
+ * within a bucket, server order is preserved. The benchmark gate holds the
+ * incremental builder equal to this.
+ */
+export function buildResourceSections<T, C>(
+	items: T[],
+	grouping: SectionGrouping<T, C>
+): ResourceSection<T>[] {
+	const bucketOf = grouping.bucketOf;
+	if (!bucketOf) return [{ key: '', label: '', rows: items }];
+	const order: string[] = [];
+	const map = new Map<string, T[]>();
+	for (const item of items) {
+		const k = bucketOf(item, grouping.ctxOf(item)) ?? NULL_BUCKET;
+		let arr = map.get(k);
+		if (arr === undefined) {
+			arr = [];
+			map.set(k, arr);
+			order.push(k);
+		}
+		arr.push(item);
+	}
+	return order.map((k) => ({ key: k, label: grouping.labelOf?.(k) ?? k, rows: map.get(k)! }));
+}
+
+/**
+ * Incremental swimlane builder. Call {@link sync} with the current (already
+ * dotfile-filtered) item list and grouping on every change; it detects the
+ * common case — the list grew by appending a page while the group-by is
+ * unchanged — and re-buckets only the fresh items, reusing every untouched
+ * bucket's array reference so `VirtualList` skips it. Any other change
+ * (group-by switch, deletion, filter toggle, non-append) falls back to a full
+ * rebuild, so the result is always deep-equal to {@link buildResourceSections}.
+ *
+ * Header labels are recomputed on every sync (never cached) because a
+ * group-by's `labelOf` may resolve asynchronously — owner / sharer names
+ * arrive after the rows do, and a cached label would freeze the header at its
+ * fallback. Only the `rows` arrays are reference-stabilised; that is what
+ * `VirtualList` diffs.
+ */
+export class ResourceSectionsBuilder<T, C> {
+	/** Last synced list — the append cursor and the append-detection baseline. */
+	#items: T[] = [];
+	/** Bucket keys in first-appearance order. */
+	#order: string[] = [];
+	/** key → the bucket's rows array (a fresh reference whenever it grows). */
+	#rows = new Map<string, T[]>();
+	/** The `bucketOf` identity of the last grouped sync; a change forces a rebuild. */
+	#bucketOf: SectionGrouping<T, C>['bucketOf'] = undefined;
+	/** False until a grouped sync has populated the accumulation state. */
+	#grouped = false;
+
+	/** Whether `next` extends `prev` (same prefix objects + strictly longer). */
+	#isAppend(prev: T[], next: T[]): boolean {
+		if (next.length <= prev.length) return false;
+		// Prefix identity via the boundary object — O(1); the list is only ever
+		// mutated by appending a page or by replacing it with a filtered copy
+		// (which preserves element identity).
+		return prev.length === 0 || next[prev.length - 1] === prev[prev.length - 1];
+	}
+
+	#rebuild(items: T[], grouping: SectionGrouping<T, C>): void {
+		const bucketOf = grouping.bucketOf!;
+		this.#order = [];
+		this.#rows = new Map();
+		for (const item of items) {
+			const k = bucketOf(item, grouping.ctxOf(item)) ?? NULL_BUCKET;
+			let arr = this.#rows.get(k);
+			if (arr === undefined) {
+				arr = [];
+				this.#rows.set(k, arr);
+				this.#order.push(k);
+			}
+			arr.push(item);
+		}
+		this.#items = items;
+	}
+
+	#extend(items: T[], grouping: SectionGrouping<T, C>): void {
+		const bucketOf = grouping.bucketOf!;
+		const fresh = items.slice(this.#items.length);
+		// Collect the fresh page's items per touched bucket, preserving order and
+		// first-appearance for brand-new buckets. Each touched bucket's array is
+		// then rebuilt exactly once (a fresh reference so VirtualList re-renders
+		// it); untouched buckets keep their existing reference untouched.
+		const freshByKey = new Map<string, T[]>();
+		const newKeys: string[] = [];
+		for (const item of fresh) {
+			const k = bucketOf(item, grouping.ctxOf(item)) ?? NULL_BUCKET;
+			let arr = freshByKey.get(k);
+			if (arr === undefined) {
+				arr = [];
+				freshByKey.set(k, arr);
+				if (!this.#rows.has(k)) newKeys.push(k);
+			}
+			arr.push(item);
+		}
+		for (const [k, add] of freshByKey) {
+			const existing = this.#rows.get(k);
+			this.#rows.set(k, existing ? existing.concat(add) : add);
+		}
+		for (const k of newKeys) this.#order.push(k);
+		this.#items = items;
+	}
+
+	sync(items: T[], grouping: SectionGrouping<T, C>): ResourceSection<T>[] {
+		if (!grouping.bucketOf) {
+			// Flat list: a single pass-through section. Reset accumulation so a
+			// later switch back to a grouped view rebuilds from scratch.
+			this.#grouped = false;
+			this.#bucketOf = undefined;
+			this.#items = items;
+			return [{ key: '', label: '', rows: items }];
+		}
+		if (
+			this.#grouped &&
+			this.#bucketOf === grouping.bucketOf &&
+			this.#isAppend(this.#items, items)
+		) {
+			this.#extend(items, grouping);
+		} else {
+			this.#rebuild(items, grouping);
+		}
+		this.#grouped = true;
+		this.#bucketOf = grouping.bucketOf;
+		return this.#order.map((k) => ({
+			key: k,
+			label: grouping.labelOf?.(k) ?? k,
+			rows: this.#rows.get(k)!
+		}));
+	}
+}

src/infrastructure/services/exif_service.rs

@@ -61,23 +61,13 @@ impl ExifService {
 
         // ── Camera info ──
         if let Some(field) = exif.get_field(Tag::Make, In::PRIMARY) {
-            let val = field
-                .display_value()
-                .to_string()
-                .trim_matches('"')
-                .trim()
-                .to_string();
+            let val = display_value_trimmed(field);
             if !val.is_empty() {
                 meta.camera_make = Some(val);
             }
         }
         if let Some(field) = exif.get_field(Tag::Model, In::PRIMARY) {
-            let val = field
-                .display_value()
-                .to_string()
-                .trim_matches('"')
-                .trim()
-                .to_string();
+            let val = display_value_trimmed(field);
             if !val.is_empty() {
                 meta.camera_model = Some(val);
             }
@@ -115,6 +105,29 @@ impl ExifService {
     }
 }
 
+/// Render an EXIF field's display value, then strip surrounding quotes and
+/// whitespace (the shape `Make`/`Model` want) in a SINGLE allocation.
+///
+/// `display_value().to_string()` is the one unavoidable allocation — the field
+/// value is materialized to text. The old `…to_string().trim_matches('"')
+/// .trim().to_string()` chain then threw that `String` away and allocated a
+/// second time for the trimmed copy. Here the same two-stage trim is applied
+/// in place on the already-owned buffer (`drain` drops the prefix, `truncate`
+/// the suffix — both reuse the allocation), so a quoted `"Canon"` costs one
+/// allocation instead of two.
+fn display_value_trimmed(field: &exif::Field) -> String {
+    let mut s = field.display_value().to_string();
+    // Same order the old chain used: strip `"` first, then whitespace. The
+    // result is a contiguous subslice of `s`; capture its byte range before
+    // mutating the owned buffer (the borrow ends at these two reads).
+    let trimmed = s.trim_matches('"').trim();
+    let start = trimmed.as_ptr().addr() - s.as_ptr().addr();
+    let len = trimmed.len();
+    s.drain(..start);
+    s.truncate(len);
+    s
+}
+
 /// Parse EXIF datetime string "YYYY:MM:DD HH:MM:SS" into DateTime<Utc>.
 fn parse_exif_datetime(s: &str) -> Option<DateTime<Utc>> {
     // EXIF dates use ":" as separator for date parts

src/infrastructure/services/search_index/content_index_worker.rs

@@ -264,13 +264,21 @@ impl ContentIndexWorker {
         let found: HashSet<Uuid> = files.iter().map(|f| f.0).collect();
         deletes.extend(upsert_candidates.iter().filter(|id| !found.contains(id)));
 
+        // `supports` lowercases the MIME (and, on a generic MIME, the extension)
+        // — 1–2 allocations per call. Classify each file ONCE here and thread the
+        // flag through both the wanted-hashes filter and the per-file records
+        // loop below, where it used to be re-derived a second time per file.
+        let supported: Vec<bool> = files
+            .iter()
+            .map(|(_, _, name, _, mime, _)| text_extractor::supports(name, mime))
+            .collect();
+
         // Per-blob text: batch-read the extraction cache, extract misses.
         let wanted_hashes: Vec<String> = files
             .iter()
-            .filter(|(_, _, name, _, mime, size)| {
-                text_extractor::supports(name, mime) && *size as u64 <= self.max_extract_file_bytes
-            })
-            .map(|f| f.3.clone())
+            .zip(&supported)
+            .filter(|&(f, sup)| *sup && f.5 as u64 <= self.max_extract_file_bytes)
+            .map(|(f, _)| f.3.clone())
             .collect();
         let mut text_by_hash: HashMap<String, Option<String>> = HashMap::new();
         if !wanted_hashes.is_empty() {
@@ -287,8 +295,9 @@ impl ContentIndexWorker {
         }
 
         let mut records = Vec::with_capacity(files.len());
-        for (file_id, drive_id, name, blob_hash, mime, size) in files {
-            let supported = text_extractor::supports(&name, &mime);
+        for ((file_id, drive_id, name, blob_hash, mime, size), supported) in
+            files.into_iter().zip(supported)
+        {
             let content = if !supported {
                 None
             } else if let Some(cached) = text_by_hash.get(&blob_hash) {

src/infrastructure/services/search_index/tantivy_content_index.rs

@@ -349,6 +349,14 @@ impl TantivyContentIndex {
             .search(&query, &TopDocs::with_limit(limit.max(1)).order_by_score())
             .map_err(|e| DomainError::internal_error("ContentIndex", format!("search: {e}")))?;
 
+        // No hits → no documents to highlight. `SnippetGenerator::create`
+        // compiles the query against the index (term lookups + weight build);
+        // for a query that matched nothing that is pure waste on the search
+        // request path, and the per-hit loop below never runs. Return early.
+        if top_docs.is_empty() {
+            return Ok(Vec::new());
+        }
+
         // Snippets highlight CONTENT matches; an empty fragment means the hit
         // came from the name (or a fuzzy variant) — no snippet then.
         let snippet_generator = SnippetGenerator::create(&searcher, &*query, fields.content)