font-display: swap paints with the fallback font stack first, then
swaps to Inter once the 876KB variable font finishes downloading.
.loader-wordmark was already width-fixed from an earlier centering
fix, but .loader-tagline, .loader-message, and .progress-percent had
no explicit width and (more importantly) used line-height: normal,
which is derived from each font's own metrics - Inter and the fallback
stack disagree by 3-4px per line. Those three deltas summed exactly to
the .loader column's total 10px height change, which is what visibly
shifted the whole centered loading screen once the swap happened.
Gave the width-dependent elements width: 100% (matching the wordmark)
and gave all three text elements an explicit unitless line-height
(computed from font-size, not font metrics) so their box size no
longer depends on which font is currently active.
Verified empirically: forced the fallback font stack on the live page
and diffed getBoundingClientRect() against Inter for every loader
element - all width/height deltas are 0 after the fix (previously up
to 17.5px width / 10px cascaded height).
Root cause: CheckBox/Button reserves vertical space for a text
baseline even with an empty string, so no amount of custom_minimum_size
or stylebox content-margin tweaking produced a true square - it just
came out close-but-not-quite in some layouts and visibly elongated in
others.
Decoupled the visual square from the CheckBox entirely: a plain Panel
(exactly 20x20, toggled by direct click) now draws the border/fill,
while a real (invisible) CheckBox still holds .button_pressed for the
rest of the code to read. Also added left padding on the row and
vertically centered both the square and the consent text within it.
Verified with a direct size dump (20x20 in both checked/unchecked
states, where it previously measured non-square) and rendered
screenshots of both states.
The generic "style every Button" pass in _apply_theme also matched
CheckBox (it's a Button subtype) and ran after the checkbox-specific
styling, overwriting the intended blue border with the plain default
button box - and that box's own content margins were inflating the
checkbox well past its 26x26 minimum (actual rendered size was 60x52).
Skip CheckBox in that generic loop, zero out the checkbox stylebox's
content margins, shrink it to 22x22, and stop it stretching to the
row's full height. Also sized the consent text down (15->13) and gave
the row some vertical breathing room.
Verified with a rendered screenshot and a direct stylebox/size dump:
checkbox is now 22x24 with a real 2px blue border, matching the
provided reference.
Drop the bordered card - it wasn't the right direction. Instead:
- Plain small checkbox with just a blue border (no fill/icon when
unchecked), matching the reference exactly.
- "Terms of Service"/"Privacy Policy" render as inline blue links
within the sentence (RichTextLabel + BBCode [url], routed through
the same _open_legal_document as the footer), instead of a duplicate
link row underneath. meta_underlined = false to match the
reference's plain (non-underlined) link style.
Verified with real rendered screenshots at each step, including a
pixel-level crop confirming the underline was actually gone.
A bare CheckBox's word-wrap left-aligns every wrapped line flush with
the checkbox glyph, so "Privacy Policy." on its own line read as
unindented and sloppy next to an otherwise fully themed form. Split
the toggle from its label (a separate Label wraps within its own
column instead, so wrapped lines stay aligned under the first line)
and wrapped the row in a bordered card matching the app's field
styling. Clicking the label also toggles the checkbox now, for a
bigger hit target.
Verified with an actual rendered screenshot: wrapped text now indents
consistently, and both checked/unchecked states are legible.
Register and Steam-linking forms rendered their own Terms of
Service/Privacy Policy links right under the consent checkbox, on top
of the ones already pinned to the bottom of the panel. The consent
checkbox text alone is enough; the persistent footer covers the links.
Register, reset, and Steam-linking forms used a separate "Show
password" CheckButton row; the main login form already had the eye
icon embedded in the field itself. Generalized _add_password_reveal
(previously hardwired to a single tracked button) to support multiple
fields via password_reveal_entries, and used it for every password
field instead. Removed the now-unused _add_show_password.
Verified all 6 password fields (login, register x2, reset x2, steam)
get a working reveal button and there's no leftover "Show password"
checkbox anywhere.
.loader-wordmark had no explicit width or text-align, so as a flex
item (align-items: center) it shrink-wrapped to whatever width the
"RanaZe" inline spans needed. wordmark-trainer's display:block then
filled that same ambiguous width rather than the wordmark block's
true rendered box, leaving it off-center relative to "RanaZe" above
it. Giving .loader-wordmark an explicit width:100% and text-align:
center makes both lines center within the same known-width box.
Verified in a real browser: both elements now report identical
getBoundingClientRect() centers (x=960 in a 1920px viewport).
Making the auth panel a fixed height (to kill the scrollbar) left the
shorter views (login, reset) stranded at the top with dead space below.
Split panel_content into a main_column (header, message, mode views -
expand-fill, center-aligned) and the legal footer as a sibling that
stays pinned to the panel's bottom edge regardless of which view is
showing or how tall its content is.
Verified headlessly: top/bottom whitespace above and below the active
view's content is now equal (e.g. login 174/174, steam 2/2), and the
footer's position doesn't move between modes.
The auth panel was a fixed 710x820 (640x790 on narrower desktop
windows) regardless of which view was showing, but Register and the
Steam linking form need up to ~990px to lay out without scrolling -
so the inner ScrollContainer was always kicking in for those views.
- Tightened form spacing (_new_form separation 16->10, field height
68->56, panel_content separation 24->16, page margins 56->28) to
shrink how much height each view actually needs.
- Size the panel to the tallest view's real requirement (1000/1010),
clamped to whatever height the window actually has so a short window
shrinks the panel - and lets the ScrollContainer take over as a
fallback - instead of the panel overflowing off-screen.
Verified headlessly across 1920x1080, 1366x768, and a narrow
compact-desktop width: no scrollbar in any mode at the two common
resolutions, and the panel now stays within the window (falling back
to scrolling only for the two tallest forms on the 768px-tall case,
rather than clipping off both edges of the screen as a fixed height
would have).
A local category_label in _build_report_modal() shadowed the
class-level category_label field (the question's topic tag near the
stem). Renamed the local to report_category_label.
boot_splash/image was still on ranaze-loader-icon-v1.png, the mark
from before the app rebranded to the shield "RZ" logo (c91e70d). Point
it at assets/branding/ranaze-mark.png, the same mark already used for
the in-app header brand throughout the current UI.
project.godot's API URLs already pointed at UAT (us3q.ranaze.com), but
deployment_environment itself was still "development" - a leftover
label that never matched the actual origin. Since the UAT-only feature
gates (Dynamic Questions, CCS Cases, Flashcards) check for the literal
"uat" value, running from the Godot editor/debugger hid all three even
though it's UAT in every other respect.
Set it to "uat" to match reality, and update the two places that
sed-patch this value at build time (web/Dockerfile for the UAT/PRD web
builds, tools/export_steam_release.sh for the Steam build) so they
still match the source string they're replacing.
Same treatment as Dynamic Questions and CCS Cases: hide the nav
sidebar entry and the dashboard's "flashcards due" review card unless
step3/deployment_environment is "uat".
Docs, local release tooling (tools/), and the GDScript test suite
(never executed by this pipeline, and irrelevant to the API image,
the exported web build, or k8s manifests) don't need a full deploy.
workflow_dispatch still allows a manual run for any of these if ever
needed.
project.godot stays checked in against UAT (us3q.ranaze.com) so the
Godot editor/debugger keeps testing against UAT. This script patches a
local copy to prd/s3q.ranaze.com just for the Windows Desktop export -
the same swap web/Dockerfile already does for UAT vs PRD web builds -
runs the export, then always restores the original file.
Verified: after a run, `git diff project.godot` is clean and the
exported .pck contains s3q.ranaze.com with no trace of us3q.ranaze.com.
It checked for a "DynamicQuestionsIcon" mode card that no longer
exists in study_selector.gd (that mode moved to its own nav item and
screen a while back), so it always failed. CCS Cases is now also
UAT-gated and won't be present in the default test environment either,
so point the check at Question Bank, the one mode card that's always
rendered.
Both are still in progress and shouldn't be reachable in production
(or local dev) yet - hide every entry point (nav sidebar, the
dashboard's "dynamic concepts due" review card, and the CCS Cases
mode card on the study selector) unless step3/deployment_environment
is "uat", the same setting that already drives the UAT badge. Added
Brand.is_uat_environment() as the shared check.
The optional "FOLLOW-UP QUESTIONS" reinforcement section shown during
answer review is now gated behind a Settings toggle (on by default),
persisted locally the same way as dark mode and interface sounds.
The toolbar had followed the progress group down into the row below
the branding after it moved there - readers expect actions docked to
the top brand row, not hanging off the question count.
- Fix the toolbar wrapping onto two rows: its minimum width was tuned
for an older, narrower button set and never widened back out after
reverting to bare icon buttons.
- Drop the "Time elapsed"/"Ctrl+Enter" readout next to Submit - it was
clutter; the underlying timer logic keeps running unseen.
- Move the RanaZe brand mark above the question-count row instead of
sharing a line with it.
- Progress text drops the word "Question" everywhere ("1 of 10"), and
core/dynamic sessions now cycle against a configurable block size
(Settings > Question block size, default 10) instead of counting up
forever.
Replace the default Godot boot splash and exe icon with the RanaZe
mark so a Windows build reads as RanaZe from first launch, not the
engine it's built on.
Varying label widths (especially when a button hides itself, e.g.
"Labs" on a question with no lab results) broke even spacing between
icons. Every toolbar button already has a tooltip_text carrying the
same label as hover text, so drop the labels/wrapper containers
entirely and go back to plain, uniformly-spaced 44x44 icon buttons.
- A toolbar icon that hides itself at runtime (labs, notes,
flashcard, bookmark) left its text label behind, showing as an
orphaned label floating in a gap. The label's wrapper now tracks
the button's own visibility.
- The theme toggle used a distinct circular pill shape; changed it
to match the same rounded-square box every other toolbar icon uses.
The 11 labeled groups are wider than the old bare-icon buttons, so
the toolbar's allotted minimum width (tuned for the old layout) was
too narrow and forced a wrap onto two rows. Widened it and raised
the compact-mode breakpoint to match.
Core sessions show a per-question stopwatch that counts up (no fixed
deadline, just a pacing signal), not a countdown - it was incorrectly
labeled "Time remaining" like the simulation/Timed countdown; now
reads "Time elapsed" to match what the number actually does.
The 6-item + "More" overflow popup made the toolbar feel sparse and
hid frequently-used actions (flag, report, flashcard, theme) behind
an extra click. Reverting to a flat toolbar where every action is
directly visible and labeled, including text size (grouped A-/A+
under one "Text size" caption instead of behind a popup).
Adapts the quiz screen's header, toolbar, choice cards, and bottom bar
to the designer's new layout:
- Header: adds the RanaZe logo/wordmark and a question-progress bar +
percentage alongside the existing "Question N of M" text.
- Toolbar: consolidates 12 icon-only/chip buttons into 6 labeled items
(Text size, Calculator, Clear highlight, Notes, Bookmark, More), with
a new "More" popup holding the rest (labs, flag, report, flashcard,
theme toggle, feedback). The underlying buttons are reparented, not
duplicated, so every existing theming/visibility/state call site
keeps working unmodified.
- Category tag: shows the question's discipline (falling back to
system) above the stem.
- Choice cards: every choice now renders as a subtly-bordered rounded
card even before answering, not just a flat divider row; the
currently-selected pre-submit choice shows a checkmark.
- Bottom bar: merges the Skip/Submit row and the timer row into one -
Skip as a plain link, a centered "X of Y" progress bar, and a
time slot that shows the existing countdown when a limit applies or
a new count-up stopwatch ("Time elapsed") otherwise, plus a
Ctrl+Enter/Cmd+Enter submit shortcut with a visible hint.
Replaces the OS/window icon with the new dark-background logo design,
and derives a transparent shield-mark cutout (via chroma-key against
the flat background) for the in-app login/sidebar mark and the web
loading-screen logo. Brand.ranaze_mark_texture() no longer needs the
AtlasTexture crop since the new source asset is already tightly
cropped to the mark.
Previously only the correct choice and any choice the student actually
selected got a percentage badge once reviewing; never-selected
distractors were left blank. Now every choice shows its selection
percentage as soon as the question is answered, matching the earlier
change that made unselected distractors render red like the rest of
the review UI.
node_added already fires once per node as it enters the tree (including
every descendant of a newly-added subtree), so the recursive call_deferred
in _on_node_added was redundant and held a raw Node reference across a
frame boundary. If that node got freed before the deferred call fired
(e.g. a container torn down and rebuilt the same frame), the engine's
message queue couldn't marshal the argument and threw "Cannot convert
argument 1 from Object to Object". Wiring synchronously removes the
window entirely.
Also silences the intentional integer division in _duration_text,
matching the same annotation already applied to _elapsed_simulated_text.
Renames locals/params that shadowed base-class properties (Node.name,
Control.size, BoxContainer.alignment), prefixes an intentionally-unused
parameter, and annotates a deliberate integer division so the console
stays clean during development.
The connected-accounts "Connect" button in progress_profile.gd
referenced res://assets/icons/lucide/link.svg, which didn't exist,
causing a resource-load error whenever the Steam/Google linking row
rendered.
Steam uses a theme-aware monochrome glyph (light/dark variants) since
the button background flips between navy and white; Google keeps its
official 4-color mark, which reads fine on either background.
Guarded the answer-review render path (teaching cards, review
follow-ups, distractor explanations) on current_answer/pending_correct_advance
directly rather than solely on correct_review_panel.visible, and added
the same guard inside _render_unselected_distractor_explanations
itself. This closes a path where a fresh, never-answered question
could show wrong-answer explanations immediately, before any
selection or submission.
Also: unselected distractors now render in the same red "wrong" card
style as selected/locked-wrong choices once a question is being
reviewed, instead of staying in the plain default row style. And
removed the redundant percentage badge that was floating above the
in-progress follow-up quiz prompt, since the choice button above it
already shows the same percentage.
Two real gaps meant the account-synced zoom (from the previous
"sync through the account" change) never reached the screen:
_load_cache() updated AccountProfileService.preferences but never
emitted profile_changed, so question_player.gd/ccs_service.gd (already
built and listening before login even happens) never found out a
cached value had loaded. And nothing forced a server refresh
right after login - it only fired if the user happened to open the
profile/progress screen, so a fresh device/session had no path to
the true synced value at all.
main.gd now calls AccountProfileService.refresh() once per login
(not on every silent token refresh) so the real preferences GET
always runs right after signing in, on any device.
Also added Question bank/CCS case text size steppers to the Settings
screen next to dark mode, per user request, reusing the same
AccountProfileService/CCSService plumbing as the in-screen A-/A+
buttons.
Drop the redundant header row (duplicate percentage badge + collapse
chevron), the "Why this is appropriate/less appropriate:"/"Why this
is incorrect:" heading labels, and the nested sub-box - now that
these cards live inside their own choice's colored border, all of
that was repeating what the color and position already say. Both the
selected-wrong-choice resolved follow-up and the never-selected
distractor cards now just show the plain explanation sentence, with
the leading "Correct."/"Incorrect." verdict word stripped too.
Removing the toggle chevron removes the only way to reach the
collapse/expand behavior, so that machinery (_collapse_resolved_followups,
_toggle_followup_state, the "collapsed" state field) comes out too
rather than linger unreachable.
Local user://settings.cfg meant zoom level didn't follow a student
across devices/browsers. Question bank and CCS text scale now live on
the user's study profile in Mongo (QuestionTextScale/CcsTextScale),
synced through a new PUT /api/profile/me/text-scale endpoint and the
existing preferences GET/cache already used for study preferences.
Both question_player.gd and ccs_service.gd read the synced value at
startup and react to AccountProfileService.profile_changed so a value
pulled down after login (or changed from another session) applies
without a restart. Backend validates the 0.85-1.35 range and updates
only the field(s) provided via a targeted $set, so it can't clobber
the other scale or the unrelated study-profile preferences.
Question bank and CCS case text-size (A-/A+) previously reset to 1.0
on every app launch. Both now load their saved scale from
user://settings.cfg at startup and save on every adjustment, mirroring
the existing pattern used for dark/light mode and sound preferences.
The flag-media-questions CLI command only caught questions that
existed in the DB when it ran once. Any question created, imported,
or edited afterward could add media without ever being flagged for
review, since nothing gated new writes going forward. CreateAsync,
UpdateAsync, and UpsertManyByExternalIdAsync now all run new/updated
questions through ApplyMediaReviewGate, which flags any question that
has media and isn't already flagged.
The "WHY OTHER OPTIONS ARE INCORRECT" aggregate section duplicated
explanations that selected wrong choices already showed inside their
own resolved follow-up quiz. Removed that section entirely; choices
the user never picked now get their "why incorrect" explanation
docked directly inside their own card (reusing the same followup
panel/attached-card styling), so every distractor's explanation lives
with its choice instead of in a separate list at the bottom.
The choice button above the panel already shows the letter badge,
choice text, and percentage, so repeating them in the panel header
was redundant now that the panel is docked inside the choice's card.
Keeps the collapse toggle, now named FollowupToggle since the old
FollowupSummary button it lived on is gone.
Now that the explanation panel is docked inside the correct choice's
own green card, the heading was redundant with the badge/border color
already conveying it's the correct answer.
Flag question now uses a clean vector flag (dark/light SVG variants,
replacing a hand-drawn pixel-art icon that got muddied by the shared
toolbar icon-tint override). Report question is now a megaphone
instead of a second flag glyph, since both toolbar buttons looked
identical before.
The correct-answer explanation and the wrong-answer follow-up quiz
previously rendered as separate bordered boxes sitting below the
entire choice list. Both now dock inside their own choice's card:
the explanation panel reparents onto the correct choice's row, the
follow-up panel was already there but visually detached. A shared
_attached_panel_box style drops the shared border/corner between the
choice button and its panel so they read as one continuous card.
Also drop a stale custom_minimum_size(0, 54) reset left over in
_set_choice_button_text from before the answer row redesign - it was
silently overriding the button's real 68px minimum height on every
question load and on every elimination/strikethrough toggle, which is
why choice rows never actually got taller and why striking a choice
through visibly shrank its card.
Fixes the large empty gap below the answer choices reported after the
answer-row redesign - the stem+choices VBox was pinned to the top of an
expand-filled scroll panel with all the leftover height sitting unused
below choice E. Now centers that block when unanswered and reverts to
top alignment once the explanation/review content grows underneath it.
StyleBoxFlat-drawn circles kept looking rough at this size even after
tuning corner detail/antialiasing, so the badges are now two small SVG
assets (outline and filled) tinted per-state via modulate instead -
crisp regardless of size. Also increased row height, badge size, and
margins across the board; the previous pass's padding bump was too
timid and the rows still read as cramped.
icon_circle_box's perfectly-circular StyleBoxFlat (corner radius equal
to half the box size, no straight edges between corners) is a known
Godot degenerate case that draws a faint seam through the middle when
antialiasing is on and looks faceted at low corner_detail. Bumping
corner_detail and disabling antialiasing renders a clean circle at
these small badge/icon sizes instead - affects every icon_circle_box
call site (answer badges, stat tiles, avatars, etc.), not just the
question player.
Unselected/eliminated choices now sit in a flat list with a thin
bottom divider instead of each getting its own bordered box, and only
the selected/correct/wrong choice pops out as a rounded, tinted card -
closer to the reference mockup and less visually busy than every
option looking like an separate boxed card.
A choice that was crossed out (right-click) and also happened to be
locked as the submitted-wrong choice, or revealed as the correct
answer, silently lost its strikethrough because the percentage-badge
branches took precedence over the elimination check in the same
if/elif chain. Decoupled the two so a choice can show both the
strikethrough and its reveal percentage at once.
The warning was purely a token-expiry countdown with no concept of
user activity, so it interrupted people who had been clicking through
questions the whole time. AuthState now tracks real input activity
globally and silently refreshes the session in the background when
the user has interacted recently, only surfacing the "Stay logged in?"
dialog for sessions that are genuinely idle. A failed background
refresh still surfaces the dialog rather than silently logging the
user out.
Flagged questions disappear from every student-facing surface (search,
exam simulation sampling, the missed-question review queue, bookmarks,
direct fetch) while staying published - flagging and unpublishing are
now separate, reversible actions with their own audit trail (who
flagged it, when, why). Backend-only for now: SUPER-only flag/unflag/
list endpoints, no client UI. Includes a one-time CLI sweep
(flag-media-questions) to flag every existing question with an image
so they can be reviewed.
Matches the mockup: the "Show Answer" button is replaced by an in-card
"Tap card to reveal answer" hint, and clicking anywhere on the card
reveals it (guarded the same way Space already was - only while at the
live card and not yet revealed).
System decks/notes are owned by a reserved sentinel user id, reusing
all existing per-user store/index plumbing unchanged. Regular users
see them under a separate "Official Decks" section, can study them
with fully independent per-user FSRS progress (lazily materialized on
first visit), but can never edit or archive them - enforced both in
the client (no Edit/Archive controls) and server-side (ownership guard
on every mutating endpoint). Includes a small idempotent seed command
(`dotnet run -- seed-flashcards`) for testing, since there's no in-app
authoring UI yet.
The study screen was still treated as a fullscreen immersive view (like
the question player), hiding the main nav sidebar entirely. The mockup
and the user's request both call for keeping the app's left menu while
studying, with the deck panel living inside the screen alongside it.
Add real session-history navigation (Previous/Next/Shuffle), compact
side panels, and a segmented progress row to match the mockup layout.
Flashcards now opens straight into Study, with library management
moved into the Options menu ("Manage Flashcards") and Export hidden
from the UI (Import stays). Shuffle is backed by a real, due-before-new
randomized pick on the backend, not a decorative toggle.
Study screen now matches the provided mockup's layout: a "Current
Deck" panel with a circular mastery-progress ring, and a "Spaced
Repetition" panel (New/Due/Learning/Review counts, streak, daily
goal) on either side of the card. Kept the app's real global sidebar
rather than the mockup's own nav, per direction.
Mastery/Streak/Daily Goal weren't tracked anywhere, so this adds real
backend support rather than showing fabricated numbers:
- Mastered = graduated to review state AND FSRS stability >= 21 days
(FlashcardScheduleStore.MasteryStabilityDaysThreshold), stricter
than the plain Review-state count
- CurrentStreakDays = consecutive UTC days with a review, computed
from review-log timestamps in the new pure FlashcardScheduling
.ComputeStreakDays (unit tested, habit-streak semantics: doesn't
reset to 0 the instant a new day starts before today's first review)
- DailyGoal/ReviewsToday reuse the account's existing ReviewsPerDay
setting rather than inventing a new one
Also removes the exit "X" button in favor of the same Escape/pause
menu the question-player flow uses (main.gd's shared app_menu, with
an "Exit Study Session" action), and adds a new circular_progress.gd
reusable ring control plus a real "Options" menu (bury/suspend the
current card) wired to schedule endpoints that had no UI before.
Compared side-by-side against Dashboard, the library screen used
entirely different visual primitives: a gradient primary button
Dashboard doesn't use anywhere, per-row bordered cards with low
contrast against the dark background, and a translucent pill badge
instead of the icon-circle pattern used everywhere else (stat tiles,
review queue, recent activity).
- Study/Save/Commit import now use Dashboard's actual primary button
style (flat ROYAL_BLUE, not the login/session-dialog gradient)
- Added a stat-tile row (Due today / New available / Decks) using
Dashboard's exact icon-circle + value + label construction, replacing
the due/new counts previously crammed into the deck dropdown's text
- The note list now lives in one bordered card with a section title
(matching Dashboard's chart/list panels) instead of each row being
its own separate low-contrast card
- Per-row card-type indicator is now an icon circle (clipboard-list /
notebook-pen) matching the Recent Activity list pattern, instead of
a translucent color-coded pill badge
refresh_theme_controls() colored the divider and the ANSWER/KEY
TAKEAWAY captions, but it only ran on an actual theme toggle - never
at initial build. Confirmed live: the divider was invisible (default
white on a white card) and the captions rendered as plain body text
instead of the intended accent color until you toggled dark mode
once. Call refresh_theme_controls() once at the end of _build() so
the correct colors apply immediately, which also fixes the same
latent gap in empty_label's color.
Study card: label the answer and "key takeaway" as distinct sections
with a real visible divider (previously an all-but-invisible
HSeparator) instead of one undifferentiated block of text. Also fixes
a real gap where a cloze card's extra field was silently dropped,
since only basic notes combined answer+extra into one label. Bigger
card, gradient Show-Answer button matching the app's signature CTA
style instead of a duller flat teal, and the Undo button now hides
itself when there's nothing to undo instead of sitting there disabled
looking identical to a real action.
Library: colored type badge per card (Basic/Cloze) instead of plain
gray text, more breathing room inside each row, Study promoted to the
same gradient CTA style, and Edit/Archive given a real secondary
button treatment instead of default theme buttons. Note editor fields
renamed ("Back / Extra takeaway" + a separate "Extra" field was
genuinely confusing) and now show which deck a card is being saved to.
They only closed via their own trigger button or a menu item -
clicking outside, pressing Escape, or opening the other dropdown
while one was open left it stuck on screen (confirmed opening both
stacked on top of each other). Add a shared invisible backdrop
(mirroring main.gd's drawer-backdrop pattern) that closes whichever
dropdown is open on outside click or Escape, and have opening one
close the other first.
current_study_day_start_utc() passed use_space=true to
Time.get_datetime_string_from_unix_time(), producing "2026-07-21
00:00:00Z" (space separator) instead of ISO-8601. The server's
System.Text.Json DateTime converter rejects that with a 400, which
the client surfaces as a generic "could not be loaded" error -
blocking every study session. Confirmed against the live API: the
space-separated value 400s, the corrected T-separated value returns
a card.
The panel's shared Brand.panel_box background is ~80% opaque, which
compounded with the dialog's own dim backdrop let the screen behind
show through and washed out the card against it. Force full opacity
on this dialog's panel specifically, since other panel_box overlays
still rely on the shared translucency.
Import and export were stubbed with status-text placeholders; wire them
to the existing backend endpoints with a real file picker (web JS
bridge or native FileDialog) and browser/desktop download, plus an
import preview with create/update/skip/error counts before commit.
Also fix the library and study screens' theme-refresh handlers, which
were no-op stubs so cards never restyled on a theme toggle, and add
responsive layout handling for narrow viewports.
Adds a full flashcards system: FSRS-based scheduling, cloze/basic
card types, CSV import/export, media storage, and study UI, wired
into navigation and the dashboard alongside existing question flows.
hz