Round 3 of benchmark-gated optimizations (benches/ROUND3.md; every change
gated by a before/after benchmark — an AFTER that did not beat its BEFORE
was to be rolled back; none needed it. Equivalence gates assert identical
row sequences / byte-identical output on every behavior-preserving rewrite):
DB hot paths (local PG16, EXPLAIN-verified):
- Web-UI listing (list_resources_paged): cursor pushed INSIDE the
folders/files UNION-ALL branches as sargable row-value comparisons with
per-branch ORDER/LIMIT + two partial expression indexes
(folder_id, LOWER(name), id). 20k-entry folder: 26.6 -> 1.3 ms/page
(19.5x); other sort modes at parity or better. New migration
20260918000000. [benches/LISTING-KEYSET.md section in ROUND3]
- Photos timeline (list_media_files): per-drive CROSS JOIN LATERAL top-N
on the timeline index, joins moved above the top-N. 50k-photo library:
97.4 -> 1.6 ms/page (55.7x). The old "LIMIT stops the scan early"
comment was refuted by EXPLAIN.
- PROPFIND sub-folders (both DAV surfaces): keyset list_folders_batch off
idx_folders_unique_name replaces COUNT(*) OVER() + LIMIT/OFFSET
(5k dirs: 79.7 -> 17.9 ms full walk, 4.5x).
Concurrency:
- Basic-auth cache single-flight (moka try_get_with): 8 concurrent DAV
connections at TTL expiry paid 8 Argon2id runs (2.6 s CPU + 8x64 MiB);
now 1 (300 ms). Failed verifications remain uncached.
- CachedBlobBackend per-hash single-flight + unique tmp names: 16
concurrent cold readers = 16 full remote downloads racing truncating
writes on ONE deterministic .tmp (corruptible cache); now 1 download
(16x less egress, 2.8x wall on a shared link) and torn files can never
be renamed into the cache.
I/O and allocations:
- Chunk-assembly reads 64K -> 512K buffers (2.3x, 8x fewer syscalls);
chunk-spool writes via BufWriter 512K (5.6x, 32x fewer syscalls).
- S3/Azure put_blob_from_bytes_unsynced overrides: dedup settle no longer
pays a HEAD probe per new chunk (2 RTT -> 1, 1.8x); Azure stops copying
every chunk (Bytes -> Body, -0.44 ms - 4 MiB alloc per 4 MiB chunk).
- Entity->DTO mapping: Arc<str> interning of closed-set display fields +
common MIMEs, 1-alloc etag/size formatting, FolderDto moves instead of
clones. File row: 11 -> 4 allocs; folder row: 11.8 -> 1 (2.1x faster).
- CardDAV REPORT: deleted dead per-contact vCard pre-generation and the
O(N^2) uid scan whose result was discarded (5k contacts: 55.7 -> 5.7 ms,
9.8x); byte-identical XML asserted.
- Search-results cache: byte weigher + 32 MiB budget
(OXICLOUD_SEARCH_CACHE_MAX_BYTES) replaces the 1000-ENTRY cap that let
~300 MiB of enriched rows sit in RSS; read latency parity.
- Dropped aws-config + aws-smithy-types (zero references; -82 dep-graph
nodes, three SDK stacks gone from every build). tokio "process" is now
an explicit feature (was enabled transitively by aws-config).
Frontend:
- Cached Intl.DateTimeFormat keyed by (locale, options) in formatDate and
4 sibling callsites: 20k dates 2612 -> 51 ms (51.6x); vitest gate
asserts output identity across locales and a 3x floor.
Validation: cargo fmt + clippy --all-features --all-targets -D warnings
clean; 518 unit + 548 integration-cfg tests green; new-shape endpoints
smoke-tested end-to-end over HTTP (all 5 listing sort modes with cursor
walks, WebDAV PROPFIND Depth-1, photos timeline, Basic-auth DAV login);
frontend npm run check clean, new vitest gates green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EBsU2qEzny3A8WQUEuMNCr
- add resource kind filter (file, folder, drive) in shared section (localStorage stored)
- add user preferences serverside store
- add client side dotfile filter (show/hide dotfiles) (user perf stored, default: dotfiles are shown)
for security trashed dotfile are always displayed
protection added: if a folder has only hidden items, a notification invite user to display it
if a user rename or create a hidden item, a notification tells it to user
Add an end-to-end and unit test suite for the SvelteKit frontend:
- Playwright e2e specs (tests/e2e/spa) with a throwaway container stack,
codegen scenarios, and an Istanbul-based coverage report pipeline.
- Vitest unit tests across API endpoints, components, stores and composables.
- `data-testid` hooks on interactive elements (AppShell, FileViewer,
ShareDialog, search, photos, files breadcrumbs, login/Nextcloud flows,
public share pages) so the e2e suite can target them deterministically.
- Serve the SPA app-shell CSP from a <meta> policy (svelte.config.js) plus a
middleware that skips the CSP header on HTML; move the Nextcloud Login Flow
v2 grant page to the SvelteKit /nextcloud/login route.
- `just front-codegen` recipe and start-server-spa.sh harness.
Make the test environment robust and consistent:
- Install a deterministic in-memory localStorage/sessionStorage in the Vitest
setup so storage behaves identically across Node versions (Node 26 ships a
native Web Storage global that otherwise shadows jsdom's).
- Pin devenv to Node 26 + PostgreSQL 18 and pin every CI job to Node 26.3.0
so the dev shell and CI run the same toolchain versions.
Repair the API/WebDAV (hurl) suite, which had drifted from the backend:
- Migrate the removed `/api/folders/{id}/listing` endpoint to `/resources`
(cursor-paginated `{items:[{resource_type,resource}]}` shape) across the
batch-copy, grants, nested-group, and WebDAV NC tests + the dav_helpers
wipe routine.
- Stop photos_etag from uploading the dedup-tracked fixture so the dedup
blob-lifecycle test can own its content-addressed blob exclusively.
- dedup_create now asserts the idempotent same-content re-upload (201 +
existing file id) instead of the stale 409 expectation.
Generated coverage reports, nyc output and the e2e server runtime data dir
are gitignored rather than committed.
Checkpoint of the in-progress frontend toolchain work (Vite pinned to ^6 after
the 7/8 rolldown build break, eslint-plugin-svelte v3 navigation/reactivity
fixes, CI/Dockerfile/manifest updates) together with three UI performance
optimizations (verified on the Vite 6 build):
- Critical CSS: move auth.css/music.css off the global path into their route
chunks (login/device/nextcloud-login, music) -- -25% gzipped critical CSS
(~5.4 KB) on every non-auth/non-music page load.
- relativeTimeAgo: cache the Intl.RelativeTimeFormat (was rebuilt per call, once
per row per render) -- 22.7x faster date formatting in large lists.
- Virtualize search results and grouped trash (list view) via VirtualList -- DOM
rows mounted stay ~constant (~27) instead of O(N) (94.6% fewer for 500 hits).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Bundles the backend+frontend delta-upload (content-dedup) feature with a
batch of frontend fixes from this session.
Upload / dedup:
- Client-hashed delta & instant upload (deltaUpload, hashWasm vendor shim)
- Backend dedup batch endpoint (dedup_service, dedup_handler, routes)
- session store owned-hash helpers; unit tests + upload-strategy bench
Frontend UI/UX:
- Colour file-type icons in grid/list (per-type tinted tiles + glyph hue)
- Robust thumbnail fallback; PDFs now show their type icon (backend
generates no PDF thumbnails) instead of a blank tile
- Fix PDF preview: load via a same-origin blob: iframe — the API URL is
blocked by the global X-Frame-Options: DENY in the browser's framed
PDF viewer, matching the existing CSP `frame-src blob:` design
- Groups: localized virtual-group description (no DB schema-note leak),
add nav.groups to the 15 missing locales, fix primary-button contrast
- Repoint --color-text-light → --color-on-accent (was faint grey on accent)
- Nudge the admin role badge off the user-menu header divider
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Backend — tail latency & throughput:
- FileContentCache, image transcode, and search now use moka single-flight
(try_get_with / get_or_load) so N concurrent misses for the same key
collapse to one disk read / transcode / query instead of a thundering herd.
Microbenchmark (128 concurrent on one hot key): 128 loads / p99 ~1023ms
before vs 1 load / p99 ~32ms after.
- DB: configurable per-statement timeout on the primary pool
(OXICLOUD_DB_STATEMENT_TIMEOUT_SECS, default 30; maintenance pool exempt) so
a runaway query can't pin a connection and starve the pool.
- DB: background pool-saturation monitor
(OXICLOUD_DB_POOL_MONITOR_INTERVAL_SECS) that WARNs as the primary pool nears
exhaustion — the early signal before tail latency cliffs.
- mimalloc: set MIMALLOC_PURGE_DELAY=0 (Dockerfile + compose) so freed pages
return to the OS and RSS tracks the live working set; benchmarked on
musl/aarch64 at ~400MB reclaimed vs 0MB with the default.
Frontend — UI / i18n fixes:
- i18n: fix literal "{{count}}" and "{{percentage}}/{{used}}/{{total}}" in the
selection toolbar and storage line — the call sites passed param names that
didn't match the locale placeholders; unify on `count` and pass the storage
template its params. Add es files.selected_count.
- sidebar: hide the drive picker when there's only one drive (the redundant
"Personal" row); remove the coloured left accent on the active nav item.
- logo: stop clipping the cloud's left bulge — viewBox recentred on the cloud's
true bbox with proportional SVG size so it keeps the same rendered scale.
- user menu: drop the default <a> underline on the link rows.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Defer loading components that are off the initial render path until they
are first needed, keeping them out of the bundle that loads on page entry.
- Add `lazyComponent` composable: a tiny rune-based holder that dynamic-
imports a component on first `load()` and exposes it for `{@const}`
rendering, with the component type inferred from the module so prop/
binding type-checking stays intact.
- CommandPalette: loaded on the first Cmd/Ctrl+K and mounted open via a
new `autoOpen` prop. It was previously imported by AppShell, i.e. in the
initial chunk of every authenticated route (~27 KB). AppShell now owns
the shortcut that triggers the load.
- FileViewer + WopiEditor: loaded when a preview/editor first opens, across
files, recent, favorites and shared-with-me.
- PhotoLightbox / PlacesMap / PeopleView: loaded on first lightbox open or
when the places/people tab is selected (the latter also defers maplibre).
Once loaded a component stays mounted and behaves exactly as before, so
behavior is unchanged on the second use onward.
- About modal overlay used a raw z-index: 1200, below the semantic
--z-modal (3000)/--z-toast (4000) layers, so token-based surfaces would
cover it. Consume var(--z-modal) instead.
- .btn-primary and .search-button sit on the accent gradient but coloured
their text with --color-danger-text (only correct by coincidence, #fff).
Use the purpose-built --color-on-accent so retinting danger can't break
them. The other --color-danger-text uses are white-on-red and stay.
- --color-item-hover-blue/--color-item-hover-sky were light-only literals
that would read wrong in dark mode; wrap them in light-dark().
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013J5koaCrHDMvS7uwpawLBN
The per-section icon colours used :nth-child(1..6), but the nav grew to 8
items (files, shared, shared-with-me, recent, favorites, photos, music,
trash). The colours had drifted out of sync with their labels, and music
and trash got no colour at all.
Map each item's colour off a stable `data-section` key instead, covering
all eight sections with distinct curated calendar-dot hues (sibling blues
for the two share directions, purple for music, red for trash). Reordering
or adding nav items can no longer desync the colours.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013J5koaCrHDMvS7uwpawLBN
Two related parity gaps from the VanillaJS → Svelte migration (issue #500):
internal-vs-external users weren't badged, and external users in a share's
member list rendered as a bare UUID with a static icon — no avatar, no email.
Both share one root cause: there was no shared user vignette and no resolver
for non-directory (external) users (the system address book lists internal
users only, and ShareDialog hardcoded isExternal=false).
- lib/api/endpoints/users.ts: resolveUser(id) — cached GET /api/users/{id}
(the authenticated per-user profile lookup) → {name, email, image,
isExternal}; returns null when the profile isn't visible so callers keep
their fallback label.
- lib/components/UserVignette.svelte: reusable identity chip — avatar (photo
or coloured initials), name, email, and a building-circle-xmark badge for
external users; resolves lazily and falls back to a caller-supplied label.
- lib/utils/avatar.ts: userInitials() + avatarColorIndex() extracted from
AppShell (now shared by both — no duplicated logic) so vignette and account
button render identically.
- ShareDialog: user member rows now render <UserVignette>; groups keep their
icon+label. Drops the dead hardcoded isExternal.
Backend already exposes everything (UserDto.email/image/is_external via
GET /api/users/{id}); no backend change. Frontend gate green (svelte-check
0/0, eslint, stylelint, prettier) + 47 Vitest.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Places fell back to a flat themed background (blankStyle) when no Protomaps
.pmtiles basemap was present, so the map was just a grey void behind the
photo markers.
Replace that with worldStyle(): a MapLibre GL v8 style that draws land masses
and country borders over a themed ocean, sourced from a bundled Natural Earth
110m outline (public domain). The asset is geometry-only (properties stripped,
coords rounded to 2 decimals) — 177 features (148 Polygon + 29 MultiPolygon),
~165 KB — served same-origin at /geo/world-110m.geojson, so no external tiles
(CSP connect-src 'self' friendly) and no per-instance basemap install. A full
Protomaps basemap still takes precedence when one is installed.
Also add a quiet map 'error' handler: if the outline is ever missing, the SPA
fallback serves index.html (text/html) which MapLibre can't parse as GeoJSON;
markers still render over the ocean layer, so swallow the error rather than
logging to the console (mirrors the checkBasemap content-type guard).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two issues kept the Photos → Places map blank once the SPA could boot:
- CSP `worker-src 'self'` blocked MapLibre GL, which spawns its web worker
from a blob: URL, so the map never constructed. Allow `worker-src
'self' blob:` ('self' still covers same-origin workers like delta-upload).
- `checkBasemap()` trusted `res.ok`, but the SPA fallback serves index.html
(HTTP 200, text/html) for any missing path — so a missing basemap.pmtiles
read as "present", and pmtiles.js then choked on HTML ("Wrong magic
number for PMTiles archive"). Reject text/html responses so an absent
basemap falls back cleanly to the themed blank style.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The files browser re-fetched the first 200 favorites AND the first 200 outgoing
shares on every folder navigation (two round-trips each time) just to render the
star / shared badges — work that grew with how much the user browsed, for data
that barely changes.
Move both id sets into a session-scoped `badges` store: `ensureLoaded()` fetches
once (concurrent callers share one in-flight request) and every later navigation
reads from cache, so browsing costs zero extra requests. Mutations keep the cache
in sync optimistically:
- favorite toggle / batch-favorite → `setFavorite` (revert on failure),
- share creation → `markShared`, wired through a new optional `onshared` callback
on ShareDialog (fired when a grant or public link is created). This also makes
the shared badge appear immediately instead of only after re-navigating.
Net effect per session: badge fetches drop from O(navigations) × 2 to 2 total.
The 200-item ceiling is unchanged from before; the fully-correct fix is per-item
flags on the listing endpoint (a backend change, noted in the store).
Verified: new badges store unit tests (load-once, concurrent de-dupe, optimistic
favorite/share, reset) and a headless render of the real files route in list and
grid (virtualization intact, no runtime errors).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M8Vb9QHmLZnEMzHz7MrFy6
Extends windowing to the remaining O(n)-DOM surfaces: the card-grid view of
ResourceList (recent / favorites / shared / shared-with-me / trash / search) and
the main file browser (`files/[...path]`), in both list and grid layouts.
- VirtualList gains a real grid mode: its inner window carries the caller's grid
class (`files-grid-view`) and lays out `columns` cards per windowed row. The
row pitch is auto-measured (grid card height tracks column width via the 4/3
aspect-ratio thumbnail) and re-measured on resize.
- `useVirtualWindow` now distinguishes scroll from resize and exposes a
`resizeTick`, so size-dependent layout (the grid pitch) is recomputed only when
it can actually change.
- `gridColumns(width)` (new util) mirrors the CSS `auto-fill` / `--grid-card-min`
/ gap so the windowed row count matches the browser's real wrapping exactly;
shared by both grid callers.
- The files browser flattens folders-then-files into one discriminated `entries`
list rendered through VirtualList (list: columns=1; grid: columns from width).
Grouped (swimlane) views stay fully rendered, as before — they're bounded.
ResourceList GRID, headless Chromium (1280x900), synthetic rows, before/after:
rows | mount→paint | DOM nodes | JS heap | scroll frame | jank frames
------+-------------+-----------+---------+--------------+------------
1000 | 979→75 ms | 19k→879 | 16→3 MB | 16→17 ms | 0→0
5000 | 4005→67 ms | 95k→879 | 72→3 MB | 37→17 ms | 26→0
20000 |12015→65 ms | 380k→879 |281→7 MB |197→17 ms |1249→19
Rendered DOM and heap are flat (O(visible)) regardless of dataset size; mount is
~185x faster and scroll holds ~60fps. Verified visually mid-scroll (5 columns,
4/3 thumbnail tiles, cards land at the expected indices). The files browser
shares the same VirtualList path (it can't be mounted headless — it depends on
$app routing/session — so it's validated via svelte-check + the production build
+ the shared, separately-benchmarked component).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M8Vb9QHmLZnEMzHz7MrFy6
The Photos "moments" grid rendered every tile into the DOM, so a 20k-photo
library mounted ~140k nodes / 20k <img> elements, held ~196MB JS heap, took
~3.5s to first paint and scrolled at ~6fps. It also ran the justified-layout
maths inside the template (recomputed on every reactive change) and generated
client-side video thumbnails for every off-screen video, not just visible ones.
Introduce `VirtualRows` — a variable-height, section-aware sibling of
`VirtualList` — and flatten the grouped timeline into one list of fixed-height
rows (a date header or a strip of explicitly-sized tiles) shared by both the
square and justified layouts. Only the rows near the viewport are mounted; a
prefix-sum offset table + binary search find the visible band, and a spacer
reserves the full height so the sticky header and load-more sentinel are
unchanged. The justified packing now runs once per groups/width/layout change
in a $derived, not per render.
To avoid duplicating the scroll-tracking logic across the two windowing
components, extract it into a `useVirtualWindow` composable (scroll-ancestor
detection + rAF-throttled aboveBy/viewportH signals); `VirtualList` is
refactored onto it with identical measured numbers.
Measured in headless Chromium (1280x900), synthetic photos, before/after:
SQUARE | mount→tiles | DOM nodes | <img> | JS heap | scroll frame
------------+-------------+-----------+-------+---------+-------------
2000 | 416→94 ms | 14k→629 |2000→96| 21→5 MB | 29→29 ms
5000 | 916→114 ms | 35k→629 |5000→96| 50→9 MB | 62→26 ms
20000 | 3455→220 ms | 140k→629 |20k→96 |196→29 MB|152→33 ms
JUSTIFIED 20000: mount 245 ms · DOM 315 · <img> 44 · heap 33 MB · ~60fps
Rendered DOM, mounted <img> count and heap are now flat (O(visible)) regardless
of library size; mount is ~16x faster and scroll jank drops from 413 to ≤24
frames. Off-screen video-thumbnail generation no longer fires for non-visible
tiles. Correctness verified by probing a deep scroll in both layouts (tiles
land within the viewport band; square cells equal-width, justified rows
aspect-preserving).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M8Vb9QHmLZnEMzHz7MrFy6
Large folders/collections rendered every row into the DOM, so memory and
scripting time grew linearly with item count. A 20k-row list took ~12.5s to
first paint, held ~380k DOM nodes / ~281MB JS heap, and scrolled at ~4fps.
Add a reusable, dependency-free `VirtualList` that windows rows against the
nearest scrollable ancestor (the existing `.content-area`), so the
single-scrollbar UX, sticky header and end-of-list sentinel are unchanged. It
reserves the full scroll height with a sized spacer and translates a small
window of rows; row height is auto-measured for the single-column case.
Wire it into ResourceList's flat list view (recent, favorites, shared,
shared-with-me, trash, search). Grouped sections and grid view are unchanged
for now and are the next callers of the same primitive.
Measured in headless Chromium (1280x900) with synthetic rows, before/after:
rows | mount→paint | DOM nodes | JS heap | scroll frame | jank frames
------+-------------+-----------+---------+--------------+------------
1000 | 632→70 ms | 19k→351 | 16→2 MB | 17→17 ms | 0→0
5000 | 4401→54 ms | 95k→351 | 72→3 MB | 47→17 ms | 136→0
20000 | 12577→56 ms | 380k→351 |281→6 MB |258→17 ms | 1443→0
Rendered DOM and heap are now flat (O(visible)) regardless of dataset size;
scroll holds 60fps with zero jank. Correctness verified by probing a mid-list
scroll (rows land at the expected indices, positioned within the viewport).
eslint: disable core `no-undef` for `.svelte` (TypeScript/svelte-check already
resolve identifiers, including `<script generics>` type params it can't see).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M8Vb9QHmLZnEMzHz7MrFy6
The merge brought in main's ReBAC→role-grants migration, which changed the
grant contract the Svelte sharing UI (branched before it) was written
against: GrantDto dropped `permission` and now carries an explicit `role`
(owner/editor/viewer/commenter/contributor), and the "admin" role was
renamed "owner". Left unchanged, the share UI derived roles from a
now-absent `permission` field and showed every member as "viewer".
- grants.ts: ShareRole is now viewer|editor|owner; Grant carries `role`
(not `permission`); `roleFromPermissions` → `displayRole`, which collapses
the unexposed commenter→viewer and contributor→editor.
- ShareDialog.svelte: read each subject's role directly (role-grants emits
one row per subject); role picker exposes Owner instead of Admin.
- shared/+page.svelte (My Shares): same owner rename; role badges run
through displayRole so server-only roles render sensibly.
Create/update already POST `role`, so only the read/display path and the
role literal needed fixing. npm run check, test:unit (36) and build pass.
Bring the Photos/People/Places UI that main added (in the legacy vanilla
frontend) into the SvelteKit rewrite, wired to the now-merged backend
(/api/photos/geo, /api/people/*).
Photos page (routes/photos/+page.svelte):
- Moments | Places | People sub-tabs (the People tab appears only when the
faces feature is enabled, via a /api/people capability probe), mirroring
the vanilla photos sub-nav.
- Square ↔ justified layout toggle. Justified uses a Flickr-style
row-packer over the width/height the photos list endpoint returns
(PhotoItem), falling back to 1:1 when dimensions are missing.
New components:
- PhotoLightbox.svelte — the lightbox extracted from the photos page into a
reusable component (items + bindable index, onDelete callback) so the
grid, People and Places all share one implementation (no duplication).
- PlacesMap.svelte — MapLibre GL map with server-clustered markers; the
vector basemap is optional (probed at /basemaps/basemap.pmtiles, themed
fallback otherwise). Cluster click zooms in or opens the lightbox.
- PeopleView.svelte — identity-cluster grid → per-person photo grid, with
rename via the in-app prompt dialog.
Supporting:
- api/endpoints/people.ts (+ peopleEnabled probe); photos.ts gains
fetchPhotosGeo + GeoCluster + PhotoItem; fileThumbnailUrl takes a size.
- lib/vendor/maplibre.ts — minimal typings + lazy loader for the vendored
MapLibre GL + pmtiles globals (kept any-free for ESLint).
- utils/media.ts — shared isVideo / photoTimestamp / minimalPhotoItem.
- Vendored maplibre-gl 5.24.0 + pmtiles 4.4.1 under static/vendors and an
optional static/basemaps dir, matching the PR's vendored-asset pattern.
- New photos.tab_*/layout_*/map_* + people.* keys in en.json.
Verified: npm run check (svelte-check + eslint + stylelint + prettier),
npm run test:unit (36 pass), and npm run build all green.
Claude