Source data wraps every math fragment in display-mode $$ delimiters,
which KaTeX rendered as centered blocks mid-sentence. scripts/fix_math.py
normalizes the corpus: $$/$ pairs become inline \\(...\\) (block
derivations with newlines or environments stay display), legacy
\\matrix blocks become begin/end matrix, _\\limits/^\\limits,
\\raise/\\lower boxes, \\root/\\of, \\tag and \\, are rewritten
to KaTeX-supported forms. Batch-validated 60k inline spans: 99.98% render
clean; remaining handful are source pathologies that degrade to text.
import_examgoal.py runs the same normalization at write time.
deploy/termux-{start,stop,status}.sh manage the release binary on the
tablet; deploy/cloudflared-whaleprep.yml is the dedicated tunnel config
for whaleprep.oiupoyt.space. README gains a deploy section.
import_examgoal.py required /tmp/pyqeval helper modules that were never
committed, so it failed on any fresh machine. The two upstream model
files now live in scripts/examgoal_model/ (relative import fixed) and
the default pickle path is scripts/examgoal_db.pkl.
GET /paper/:id redirects the owning user to their attempt screen (404
otherwise, login wall otherwise). Shared /s/:token pages now end in a
real action instead of a dead list: signed visitors get a take-this-paper
button, anonymous visitors get a sign-in link preserving the share URL
as next. Verified live: 1/302/1 across the three states.
seen_questions gains seen_at (with ALTER TABLE migration for existing
DBs, index created after the migrate step). Submit stamps seen_at with
an upsert so repeats refresh recency. build_paper tops up short
unseen-only pools from longest-ago-seen in scope instead of erroring.
Verified live: 1-question subtopic served the same seen question on
repeat builds (303, not 400).
Result page shows a why-lost row (concept gap / calculation slip /
misread / out of time / blind guess) on every wrong or blank question.
POST /result/:attempt/tag/:qid toggles one tag; ownership checked,
vocabulary enforced server-side. Dashboard gains a 'why you lose marks'
section aggregating tags across all submitted attempts. Verified live:
tag persists, rollup renders.
load_chapter_blocks used a correlated COUNT subquery per chapter row:
~5s on the 9.8k-question bank. Rewritten as one GROUP BY joined once
(16ms p50 on /build). Nine CREATE INDEX statements added to init_db
so fresh and existing DBs both get them. build_paper had a dead
placeholder branch issuing a subtopic_id=0 lookup per chapter; removed.
scripts/import_examgoal.py pulls the MIT-licensed examgoal-derived
14,973-question pickle and keeps the 9,813 fully-text MCQs with single
correct answers (2002-2025, all subjects). chapterGroup becomes our
chapter, chapter slug becomes our subtopic. Idempotent: skips stems
already present.
views.rs gains rich() — a read-side sanitizer matching the importer's
allowlist — and all six stem/option/solution render sites use it.
KaTeX auto-render (trust:false) added to head() for $$..$$ math.
db.rs gains nine CREATE INDEX statements; the chapter loader query
drops from ~5s to 6ms on the 9.8k-question bank. DATA.md rewritten
around the real pipeline with attribution.
scripts/expand_seed.py generates synthetic questions across the full JEE Main
syllabus taxonomy (15 chapters, 45 subtopics, correct MCQ/NUM split). Stems
are deliberately sparse so no-one mistakes them for NTA content.
DATA.md documents why there is no real scraper: NTA PDFs render questions as
images; pdftotext extracts only IDs. Every aggregator with structured data
built it through manual entry pipelines. The document lists realistic paths
to a real bank and the SQL schema for loading one.
The builder previously exposed only chapters. Now each chapter expands to show
its subtopics as individual checkboxes: selecting a chapter checks all of its
subtopics, selecting the subject-level box checks everything, and selecting
individual subtopics sets the chapter to indeterminate. The exam query filters
by subtopic_id when any are given; the SQL path deduplicates so a question
matching two subtopics only appears once.
JEE mode sets duration to 60 × unique-subject-count (180 min for a full paper).
On the exam screen this renders per-subject countdown strips: Physics 60:00,
Maths 60:00, Chemistry 60:00. Each strip counts down independently; when one
hits zero the screen auto-jumps to the first unanswered question of the next
subject. The last strip exhausting time submits the form.
Single-subject papers keep the flat timer and show no strips.
Also: all audit targets now genuinely 44px (the .opt labels were 42px).
The screen had no mobile state at all: the clock scrolled away mid-paper, the
only navigation was save-and-next so a candidate could not go back, the palette
sat below the question where it pushed submit off-screen, and the nav row lost
three of its four links under 700px.
Sticky header carries the clock and a progress bar; previous/next are real
buttons; the palette is a drawer below 1024px opened by a floating button that
sits above the tab bar. Primary destinations move to a bottom tab bar on
phones, data tables scroll inside their own card, and every control meets the
44px touch minimum.
Verified with a headless browser across five viewports rather than by reading
the stylesheet: scripts/audit_mobile.py fails on horizontal overflow, tap
targets under 44px, and content hidden behind a fixed bar; scripts/shots.py
captures each page at both widths. Both caught bugs I would otherwise have
shipped, including a palette toggle that lived inside the drawer it opened and
so could never be pressed on a phone.
fmt_dur drops a leading zero hour, so a 24 minute paper reads 24m 00s.
The subscriber was pinned to DEBUG, so RUST_LOG=warn was ignored and sqlx
logged every statement to stderr in production. Now an EnvFilter defaults to
info with sqlx at warn; RUST_LOG=debug restores the query log.
Three benchmarks, so the hosting numbers come from measurement rather than
guesswork: scripts/bench.py for latency and throughput, bench_write.py for the
autosave and submit path under concurrent students, and bench_scale.py which
builds a database at the real bank's size (10,905 questions, 2,000 students,
12,000 papers, 900k answers) to report disk cost and query timings.
Two paths, one session. The GSI popup posts an ID token to /student/google;
where a popup is refused a PKCE authorization-code flow redeems server side at
/student/google/callback, keeping the client secret out of the browser.
ID tokens are verified here rather than trusted from the browser: RS256
signature against Google's published JWKS (cached an hour), audience pinned to
this client, issuer restricted to Google, expiry enforced, email must be
verified. State is single use and expires in ten minutes. Without
GOOGLE_CLIENT_ID the button is hidden and the email form is offered instead.
13 tests cover the verification rules using locally generated RSA keys,
including rejection of a foreign signature, a wrong audience, a wrong issuer,
an expired token and a rewritten payload. An example signs a token from an
untrusted key so the rejection path can be checked without a Google account.
Axum + SQLite full-stack copy of the ZenithPrep PYQ practice app.
Paper builder with unseen-question exclusion, timed CBT screen with
palette and autosave, +4/-1 grading with numerical tolerance, per-question
analysis, history, dashboard with weak-subtopic tracking and coach notes.
