rand_array seeds with Math.floor(time / 86400000), i.e. the same
shuffle for every caller during a UTC day — so every list endpoint
that runs through it is safe to share publicly:
- /items/all_brief, /items/category/[c], /items/tag/[t],
/items/icon/[i], /items/in_season/[m]
→ public, max-age=28800 (8h), s-maxage=28800, SWR=1d
The distinct-value lists (no shuffle, change only on recipe edit):
- /items/category, /items/tag, /items/icon
→ public, max-age=3600 (1h), s-maxage=86400 (1d), SWR=1w
Individual recipes change when their author edits them:
- /items/[name]
→ public, max-age=300 (5m), s-maxage=3600 (1h), SWR=1d
Fitness exercise-picker filters are identical for every logged-in
user but require auth:
- /fitness/exercises/filters
→ private, max-age=3600
Skipped the calendar page itself: its HTML embeds data.session via the
faith layout's <UserHeader>, so public caching would leak identity.
List endpoint previously returned full session documents minus GPS
tracks — two months × up to 200 sessions means ~60 KB of payload per
month with a lot of fields (notes, templateId/Name, mode,
activityType, endTime, session-level gpsPreview) that SessionCard
never reads.
Narrow the projection to exactly what the history page + SessionCard
use, and switch the query to .lean() so we skip the Mongoose document
overhead on deserialisation.
Detail view (/fitness/history/[id]) hits a separate endpoint that
keeps the full document.
yearDays was a 365-entry array (one per day in the LY window) with
{iso, name, rank, color, seasonKey} on each — the client only needed
the color (for the needle pin on the currently-selected day; RingView
re-did the feast filter itself). Split into:
- yearDays: {iso, color} — unchanged count, but ~60% smaller per entry
(drops name, rank, seasonKey)
- feastDots: {iso, name, rank, color} — new, pre-filtered to
rank > ferial server-side (~150 entries instead of 365)
RingView's `feastDots` derivation shrinks to filtering out just the
currently-selected day, and `activeFeasts` filters `feastDots` by arc
bounds instead of re-scanning yearDays. needleDay's color lookup still
works with the trimmed YearDay.
Also collapses a stray `locals.session ?? (locals.session ?? …)` the
earlier #5 sweep introduced in both calendar page loaders.
Endpoint previously pulled full WorkoutSession documents (including
gpsTrack, notes, kcalEstimate etc.) to count sets per muscle group.
Adds a projection that keeps only startTime + exercises.exerciseId +
whole set objects — safe (avoids the malformed-sub-array issue the
earlier narrower projection caused in the stats overview handler),
but still drops the bulky session-level fields.
Also swaps the per-session findIndex() over the weekly bucket array
for direct date-math against the first bucket's Monday, turning
bucket lookup from O(sessions × weeks) into O(sessions).
Four panel fetches (muscle heatmap, nutrition stats, own periods, shared
periods) are now returned as unawaited promises from load() and resolved
into $state-backed locals on the client via $effect. The load function
keeps awaiting only stats/goal/latest since the main charts, goal
header, and body-part cards depend on them immediately.
Rationale for the $state-backed resolution rather than {#await}: the
user wants the nutrition card shells and the muscle heatmap container
to render their skeleton shape on first paint and only fill in the
numbers once the data arrives. Defaults (`{}`, empty heatmap, `[]`)
match the previous error-fallback shapes so the existing `!= null`
checks inside each card cascade naturally to the "—" branches while
the promise is in flight. No template restructuring beyond dropping
the outer `{#if ns}` (which already hid everything when null).
stats (overview) is intentionally still awaited: it feeds ~30 $derived
chart expressions and wrapping it would mean recreating every Chart.js
instance after the promise settles.
Extends the previous loader-only sweep across the full tree: every
remaining `await locals.auth()` now falls back through
`locals.session ?? await locals.auth()`, so the hook's cached result
is reused.
68 files, 107 sites touched — loaders, form actions, and API
endpoints across cospend / tasks / fitness / faith / recipe / admin.
hooks.server.ts is intentionally left alone since it's the originating
call that populates locals.session in the first place.
hooks.server.ts already awaits auth() once and stores the result on
locals.session. In-scope loaders (recipe list + filter views, rosary,
prayers, calendar — already done — and fitness stats) were awaiting
locals.auth() a second time per request.
Switched to the existing `locals.session ?? await locals.auth()` pattern
so the hook's result is reused. Also pulls session out of Promise.all
legs since it's now synchronous when the hook ran.
Scope: loaders only — actions, /admin, /edit, /add intentionally skipped.
Favorites page fetched both /favorites/recipes and /items/all_brief, then
stitched isFavorite flags onto allRecipes so Search could filter across
the full catalogue. But Search is invoked with favoritesOnly={true} and
hideFavoritesFilter (so it's pinned on), so it only ever returns matches
that are already in the favorites list — allRecipes was dead weight.
- drop allRes / allRecipes / favoriteIds / allRecipesWithFavorites
- Search now receives data.favorites directly
- filteredFavorites filters data.favorites by matchedRecipeIds
- use locals.session ?? locals.auth() to reuse the hook's auth lookup
Every recipe list endpoint wrapped its result in
`JSON.parse(JSON.stringify(...))` before handing it to `json()`, which
then serialises again — a full extra stringify+parse cycle per response.
`lean()` already returns plain objects and ObjectIds/Dates serialise
correctly through `json()`'s single `JSON.stringify`, so the extra round
trip was pure waste.
Removed from the 9 output-side call sites (all_brief, category,
category/[cat], tag, tag/[tag], icon, icon/[icon], in_season/[month],
search, favorites/recipes, offline-db, translate/untranslated).
Kept the two deep-clone-before-mutation usages in items/[name] and
json-ld/[name] — those are load-bearing.
Shuffle stays server-side: moving it to the client would need a hero
preload + hydration rework that's bigger than a perf tweak.
Chart.js (~244 KB) was a top-level import, so every route that referenced
FitnessChart.svelte transitively pulled it. Defer it to an async block
inside onMount so non-stats fitness routes (workout, check-in, nutrition,
history list) no longer ship chart.js.
- `ChartCtor` holds the async-loaded constructor
- `disposed` guard handles unmount during the import
- theme MutationObserver / matchMedia wiring moved inside the async
block so it only attaches once the chart actually exists
Barrel `from '@lucide/svelte'` imports pulled every referenced icon into
one shared 748 KB client chunk. Switch every call site to per-icon
subpaths (`@lucide/svelte/icons/<kebab-name>`) so Vite tree-shakes each
icon independently. Also logs the TODO list for the perf audit so we
don't lose track.
- 46 files, 106 unique icons
- single `Minus as MinusIcon` alias preserved
- Lucide-internal aliases (`AlertTriangle`, `BarChart3`) resolve through
Lucide's own re-export shims; no behavioral change
Body-measurement variation of <4 cm used to stretch the full chart
height, making normal weekly noise look dramatic. Now the y-axis
enforces a 4 cm floor centered on the data's midpoint; wider swings
render at their actual range as before.
- FitnessChart: new optional `yMin` / `yMax` props mapped to Chart.js
`suggestedMin` / `suggestedMax` — soft bounds, so data that exceeds
them still widens the axis.
- `/fitness/stats/history/[part]`: computes min/max across available
values (both sides if paired), enforces the 4 cm floor, passes to
FitnessChart. Tick distance stays on Chart.js auto — small ranges
get 0.5 cm ticks, wider ones scale up naturally.
Route slugs and nav label rename only — storage, API endpoints
(`/api/fitness/measurements`), and the `BodyMeasurement` Mongo model
keep their technical names.
- `/fitness/measure` → `/fitness/check-in` (EN)
- `/fitness/messen` → `/fitness/erfassung` (DE)
- Folder `[measure=fitnessMeasure]` → `[checkin=fitnessCheckIn]`
(git rename; history preserved).
- Param matcher `fitnessMeasure.ts` → `fitnessCheckIn.ts`, accepts
`check-in` / `erfassung`.
- `fitnessSlugs(lang).measure` and `fitnessLabels(lang).measure` code
keys are unchanged — value returns "check-in"/"erfassung" and
"Check-in"/"Erfassung" respectively, so no call site needs touching.
- slugMap language-detection updated to `erfassung ↔ check-in`.
- Service-worker cache list + the layout regex that gates the wider
content width now reference the new slugs.
- Nav icon swapped from `Ruler` to `NotebookPen` — reads as "logging
entries" and spans weight / composition / period better.
Bookmarks on the old URLs will 404; no redirect added.
Mirrors the weight chart pipeline (SMA + ±1σ confidence band) for
body-fat %, but emits deltas from the first displayed measurement so
the y-axis shows change instead of raw numbers. Title surfaces the
baseline (e.g. "Body Fat · Δ from 18.2%"), y-unit is "pp" (percentage
points), colours are purple trend on top of an orange raw-data line
so it reads differently from weight's green+blue at a glance.
FitnessChart gained two shared upgrades: `interaction.mode = 'index'`
on line charts so hovering the x-axis shows tooltips for every dataset
(including the trend line whose pointRadius is 0), and a `σ` dataset
filter so the confidence band doesn't clutter the tooltip. A new
optional `tooltipFormatter` prop lets callers format the hover label;
the BF chart uses it to show the signed delta + reconstructed
absolute % for raw points and additionally the ±1σ window for trend
points (e.g. "+0.30 ±0.45 pp · 18.5% (18.0–18.9%)").
- New "Same as last" pill below each step's stepper. Clicking fills
the input(s) with the prior recorded value(s) — for paired steps
in split mode, both L and R — and advances to the next step.
Only rendered when a previous measurement exists; the placeholder
already surfaces the exact number so the button text stays terse.
- Copy L→R button resized to match the same-as-last pill (0.88 rem
text, 0.55 × 1.1 rem padding) and given top margin. Unicode →
swapped for a proper ArrowRight icon between L and R.
- i18n: added `same_as_last` and split `copy_l_to_r` into
`copy_l_to_r_before` / `copy_l_to_r_after` so each language keeps
its natural wrapping around the arrow (EN "Copy L / R",
DE "L / R übernehmen").
- Server POST now upserts by (user, calendar day). Non-conflicting
fields merge silently; real overwrites (new non-empty value ≠
stored value) return 409 with the conflict list. Client retries
with `?overwrite=1` after a confirm dialog naming each field and
its old→new value. Null/empty payload fields are skipped, so
logging a body-fat entry on a day that already has weight merges
cleanly without flagging a phantom weight conflict.
- `summaryParts` in the history now includes a body-parts count,
e.g. "86 kg · 0.1% bf · 5 body parts" or "5 body parts" instead
of the flat "Body measurements only" fallback. Pluralised in EN
and DE.
- Inline quick-edit: "Full edit →" text replaced by a dashed primary
pill `Pencil · Edit all fields · ChevronRight`, inlined with the
X / ✓ action buttons on the same row. The label collapses to
icons only at ≤480px so the three controls stay on one line.
- Quick-edit date input swapped from native `<input type="date">`
to the site's `DatePicker` component.
- New i18n: `overwrite_title`, `overwrite_message`, `overwrite_confirm`.
- TODO.md marks features #2 and #3 done. CLAUDE.md carries a
policy note (no AI-attribution trailers on commits).
- Add SearchInput component for reusable search UI
- Add search functionality to prayers list with two-tier results:
- Primary matches (name/searchTerms) shown first
- Secondary matches (text content) shown after with reduced opacity
- Add individual prayer pages with language-appropriate slugs
(e.g., /glaube/gebete/ave-maria, /faith/prayers/hail-mary)
- Make prayer cards clickable to navigate to individual pages
- Fix language visibility for prayers without Latin (BruderKlaus, Joseph)
- Add Prayer wrapper to MichaelGebet for consistent styling
- Use CSS columns for masonry layout with dynamic reordering
Alexander
