changelog

Seven new words from term 5 (gyógypedagógus, Egészségedre!, szerelő, szerviz,
szennyezett, nej, pirítós), 98 more sentences from the notes below the vocabulary
tables, all voiced, and clips for six class-1 words that had none.

Co-Authored-By: Claude Opus 5.5 

Grammar gaps show the English and the missing word's dictionary form; a word already known is
asked once per round; new words arrive commonest first; a typed meaning a word short or a typo
off says "Nearly"; correct typed answers move on by themselves. Built from hungarian e25e128.

Co-Authored-By: Claude Opus 5.5 

Co-Authored-By: Claude Fable 5.1 

Built from hungarian 9d43c05. Eighteen new clips for the September class
words; two Forvo clips give way to Commons recordings.

Co-Authored-By: Claude Fable 5.1 

Co-Authored-By: Claude Fable 5.1 

Co-Authored-By: Claude Fable 5.1 

Turning over two face-down tiles and finding they do not match is the mechanic,
not a failure to know the word. A word now has to be missed three times before
it counts.

Co-Authored-By: Claude Opus 5 (1M context) 

200 cloze items across 50 teaching points, each one checked against the class
vocabulary before it shipped. Three more sentence clips.

Co-Authored-By: Claude Opus 5 (1M context) 

Quests instead of a streak, a daily challenge with a shareable row, an exam
per zone, and five rounds built on the instructor's own sentences -- fill the
gap, put it in order, what did they say, write the sentence, say it out loud.

286 new clips, one per sentence, synthesised locally. Everything else in
audio/ is untouched.

Co-Authored-By: Claude Opus 5 (1M context) 

Levels, skill lines and a dark theme, plus three rounds that ask for a form
rather than a headword: the word back in one of the instructor's own
sentences, listening for meaning rather than spelling, and the meaning typed
out.

Also offline for the first time -- the manifest has been here for a while and
a service worker had not, so "installable" meant an icon that showed a
connection error on a train.

Five new clips; everything else in audio/ is untouched.

Co-Authored-By: Claude Opus 5 (1M context) 

Rebuilt from hungarian/data/words.json after importing lesson plans
A1-4-11, A1-4-12 and A1-4-13 and syncing the reviewed backlog: 2247
words, 2234 with audio, up from 1062.

Co-Authored-By: Claude Opus 5 

NTFS rejects "?" in a filename, so these thirteen clips failed the backup of
the workspace share and failed the whole job with them. Fixed upstream in
hungarian and szojatek; this is the deployed copy.

Eleven were byte-identical duplicates of the clean-named clip already present
and were dropped, two were renamed. The bundle references the names inside
backtick template literals rather than quoted strings, so it needed updating
too - without that, those thirteen words would load a 404 on the live site.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01RABCtgV5xEf8r6swemwiTi

Two problems in scripts/serve.mjs, both live in this repo rather than theoretical.

The containment check compared a bare prefix, so any path resolving under a sibling
directory whose name merely starts with the root also passed. This repo has a
_site.stale directory sitting right next to _site, so `/../_site.stale/...` was served,
percent-encoded or not. Now compares against ROOT + separator, with the root itself
allowed.

The server also listened on every interface while printing a localhost URL, which put an
unpublished local copy of the site on the LAN for the duration of `npm run dev`. It now
binds 127.0.0.1, matching what it advertises.

Verified against both forms of the escape plus ordinary paths. check:pii:all and the
build stay green.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01CNgWMcdGJckkWCkKusLXKB

The CI section still described the old pipeline, and neither the --all mode nor the
content/sites/ exclusion was written down anywhere a future session would find it.
Records the cost of the exclusion explicitly, since a skipped directory that nobody
remembers is how something eventually ships unscanned.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01CNgWMcdGJckkWCkKusLXKB

Three blind spots meant a green tick was worth less than it looked.

git quotes any path containing a non-ASCII byte and octal-escapes it
("...Brass\303\263.mp3"). The diff parser matches a bare `+++ b/` prefix, so every such
file fell out of the scan: 1373 of 2585 tracked paths, more than half the repo, on a
public site. The git helper now passes -c core.quotepath=false, after which none are
quoted. scripts/build.mjs shells to git only for `rev-list --count`, so it needs no
equivalent change.

Adds a --all mode (npm run check:pii:all) that sweeps every tracked file rather than the
staged diff. The staged mode is right for the pre-commit hook but passes vacuously when
nothing is staged, which is the state of a fresh CI checkout: wiring the existing script
into CI would have reported success having read nothing. Skip rules are now shared by
both modes rather than inlined in the diff loop, and binaries are excluded by extension
plus a NUL check.

content/sites/ is now skipped whole, not just its assets/ subdirectories. Sub-sites are
self-contained published artifacts reviewed when they go up, and their payload IS the
data: coordinate arrays, and minified bundles where 4294967296 is 2^32 and a weight
array reads as GPS. Scanning them produced 498 hits, every one of them the thing the
page exists to show. The trade-off is recorded in the code: anything under content/sites/
is unscanned, so a new sub-site needs reviewing before it is added.

With those done and the last false positive allowlisted (1013904223, the Numerical
Recipes LCG increment in build.mjs, ten digits so the phone pattern matched), the
whole-tree sweep is clean. check:pii:all now runs in deploy.yml ahead of the build, so a
leak blocks the deploy instead of being noticed after it is live. Verified the full CI
sequence locally.

Also: .claude/settings.local.json was untracked but unignored on a public repo, one
`git add -A` from being published, and it carries machine paths and usernames. And CI
plus engines.node move from Node 20, which is EOL and no longer gets CVE fixes, to 24.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01CNgWMcdGJckkWCkKusLXKB

Three passes, each verified to leave the rendered text of all 337 pages
byte-identical to what was live before.

site.config.yaml is new: which sections exist and whether their index is a
timeline, the reserved FOIA routes, reading speed, excerpt and description
lengths, every image width and JPEG quality, RSS and keyword limits. All of
it was previously a literal somewhere in build.mjs.

copy.yaml gains a ui: block for chrome shared by more than one template, and
the strings the build itself was emitting: the unit words in a meta line, the
doc/pp counters. Templates no longer hardcode "browse by folder:", "Tags.",
"other years:", "no commits yet." or the reading-pane button labels. Five keys
that were defined but referenced by nothing are gone, and two category labels
that existed in copy while the template ignored them are now actually wired up.

97 inline style attributes become 69: the four that repeated are classes now
(.sp-crumb, .sp-eyebrow, .sp-page-head, .sp-backlink), each carrying exactly
the declarations its inline rule did. The FOIA back link kept its wider top
margin as .is-wide rather than being quietly flattened into the others. What
remains inline is the genuinely dynamic set, mostly background-image built
from a photo URL.

Co-Authored-By: Claude Opus 5 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

Work happens on home-server now, from /nas/workspace directly, so the UNC and
cmd.exe story behind "use execFileSync" no longer applies. The rule stays, on
the reason that outlives the share: execSync hands the command to a shell and
re-parses whatever is in a path or a commit subject.

Co-Authored-By: Claude Opus 5 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

Seven templates each had their own copy of the entry markup, which is why the
three sections had drifted into three different reading experiences: a card
stack, a dated stream and a ledger table.

They now all include templates/_entries.eta and get the same row, thumbnail,
pinned strip, sort toggle and reading pane. build.mjs normalizes a post or a
records request into one entry shape first, so the component never learns what
a musing or a FOIA request is. 512 lines of duplicated markup become 192 across
the seven callers plus one 215-line partial.

Musings keep their excerpt and tag line; they are longer-form and the excerpt
earns its place. FOIA loses the table: filed date, agency, volume and tracking
go on the meta line, status and fees on the sub-line, the blurb as the excerpt.
Agency pages show the statute where the index shows the agency, so the column
does not just repeat the page title. Requests have no featured image, so their
thumb slot stays empty rather than collapsing, which keeps the left gutter
aligned with the other two sections.

The reading pane now lifts .musings-main and strips .crumbs, .sp-postnav and
.sp-backlink, which works for all three page types instead of only musings.
S3-backed posts point at the 600px thumb they already have rather than baking
a second crop.

Co-Authored-By: Claude Opus 5 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

They were fixed in the working tree but never staged, so the homepage comment
and the first/last markers on FOIA release and document pages still carried
one. Caught by sweeping the deployed site rather than the local build.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

Site-wide rule from here on: period or comma, never an em dash. A period where
the dash joined two independent clauses, a comma where it set off an
appositive or introduced a list.

Swept prose, alt text, copy.yaml, template strings, the page title separator
and the two standalone static sites. Where the dash was an empty-value glyph
(an unpriced table cell, a missing stat) it becomes a hyphen; where it was a
leading marker before a parenthetical it is simply gone.

832 visible em dashes down to 21, all of them on /changelog, which renders git
commit subjects from ten commits written before this rule existed. Clearing
those means rewriting history or faking the commit text at render time, so
they stay for now.

The rule is written into CLAUDE.md so it outlives this session.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

An entry with no featured image keeps its column so titles stay aligned, but
shows nothing in it. A pin has no column to hold, so its slot collapses and
the title sits flush.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

Pinned: `pinned: true` in frontmatter lifts an entry out of the dated stream
into a compact strip at the top. Costs & hours is the only one so far; Lessons
Learned joins it when it exists.

Sort: newest first is what ships in the HTML, so the no-JS default is already
right and the control stays hidden. The toggle rebuilds the // year rules
rather than just reversing rows, and remembers the choice.

SEO: the index was a bare list of links, which reads as thin content. It now
carries an intro paragraph and Blog + BlogPosting structured data for all 15
entries, and the title tag says what the page is instead of "rv12is".

The intro and the title are drafted -- they are marked in copy.yaml and want
George's own words.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

I compressed all nine on the way in. Four lost real content -- the technical
counselor, what the pre-made flaperon parts save you, the spars visible in the
garage shot, and "I'm not proud to show this, but if it helps somebody, that's
awesome," which was the entire reason that photo was published.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

Featured images are stored full-size, so the index bakes a 160px square crop
rather than pulling a dozen 1800px JPEGs. Entries with no featured image keep
the column so titles stay aligned.

The chevron is an ordinary link; with JS it fetches the post into a pane that
scrolls on its own. Esc, the scrim and browser Back all close it, and the URL
tracks what is open so a reload lands on the real page.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

Wings at 9.5/8.5/1.0. Two more replacement parts, and the #40-before-#30
lesson on the aileron counterbalance.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_015yJvzKQmSNts9HrKjY9HX9

Three sources, in descending order of preference: Wikimedia Commons, Forvo,
and only then synthesis.

The largest single gain was not new recordings. A headword is not always what
to ask a pronunciation site for -- `mi?`, `keres (pénzt)`, `rózsaszín(ű)`,
`bemegy vhova/vhonnan` -- and the fetchers had been querying the lemma
verbatim since the project began. 48 clips that Commons and Forvo had all
along, including every one of the question words. `lemmaGuesses` now strips
the annotation and both fetchers consult it.

It deliberately never offers the synonym inside a cross-reference. Taking the
shortest variant looks obvious and gives `gramm (= g)` -> `g`, which Commons
has a clip of -- of the letter G. `édesanya (= anya)` would get anya's voice
under édesanya's card. A card that says one word while the clip says another
is worse than silence, so only annotation is stripped, never substituted. A
test pins it.

The rest is synthesis, 224 clips from the macOS Hungarian voice, marked
audioSource: 'tts' and credited as synthetic so a later Commons run can still
upgrade one to a human. Local rather than cloud for a reason beyond cost: a
cloud voice means POSTing the locked deck's plaintext to somebody else's
server, which is precisely what its encryption exists to prevent.

The locked deck reaches 123/123, its clips named after their hashed ids. 2,200
clips ship under readable Hungarian names and doing that here would let anyone
read the vocabulary off a directory listing. build-data.mjs skips p-*.mp3 when
pruning, or it would delete all 123 on the next run for being unreferenced.

Twelve entries stay silent on purpose: two bound suffixes, and ten sentence
frames with gaps. `azért ... mert` read aloud becomes "azért mert", which is
not the thing being taught.

Also: A1-4 Class 11 vocabulary (hányadik, század, forradalom, világháború,
aranyérem and eight more), and pofoz, forgalom, körforgalom, gyökér.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01RCxF3MSEkr2nyzX2ahsw3B

A Connecticut request to the Norwich Police Department for its dealings with
the Law Enforcement Council of Connecticut and with PoliceApp, the platform
its hiring runs through. Fifth entry in the archive.

Fourteen pages, released as a scan with no text layer -- which means the PII
pre-flight could read none of it, so I read all fourteen by hand instead,
pulling the embedded page images out of the PDF. Worth recording what that
turned up, since a scanner would have found nothing:

LEC annual assessments climbing from $13,926 to $17,471 between 2022 and 2026.
Three Eastern Regional Police Academy invoices covering seven named recruits at
$3,300-$3,500 each -- municipal spend on sponsored recruits, which is a routine
public record. Two PoliceApp pipeline exports listing every application by date,
stage and outcome, already anonymised by the department. A vendor sales email.

The substance is on the first two pages, where the department answered each
item in red: Norwich HR rather than the police department holds the PoliceApp
contracts and pays those invoices, the department has never received statistics
reports from PoliceApp, and it does not possess any validation study or
disparate-impact analysis of the LEC written entry examination. That last one
was asked for specifically.

Status is partial. The email categories were set aside to get the discrete
documents moving and are still outstanding.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01SwfZCx4FMsF1CaWt9PUgQW

CLAUDE.md gains a /foia section: the content model, the traps, and the design
lesson. The traps are the useful part -- rights language is per-jurisdiction
and 17 USC 105 does not cover county records; pdf.js needs bucket CORS with
range headers exposed because it fetches bytes via XHR; a cross-origin
 is ignored so the viewer builds a Blob from bytes it already has;
and correspondence logs are timeline source, never content, because they carry
session tokens and home addresses.

Also recorded: shell out to git with execFileSync, never execSync. execSync
goes via cmd.exe, which refuses a UNC working directory and quietly runs from
C:\Windows. On this network share that left the PII hook unable to read its own
staged diff -- so, failing closed, it blocked every commit -- and dropped the
version to v1 with an empty changelog. That one cost an hour to find and would
have cost the next session the same.

The design note is the one I most needed and did not have. I built the first
pass as a stack of notice boxes over a seven-column table with the same records
repeated as cards underneath, on a site whose whole aesthetic is a narrow flat
column. "Don't add" now says it plainly: rules beat borders, a footnote beats a
callout, and check what an existing page type already does before inventing a
layout.

The skill picks up the AWS bits -- where the CLI actually lives, and that
authenticating needs a real console rather than a shell with no TTY.

Writing that down promptly tripped the scanner on the very number the note
cites as a false positive, so 2^32 is now allowlisted.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01SwfZCx4FMsF1CaWt9PUgQW

Public pack: forgalom (traffic), körforgalom (roundabout), gyökér.

gyökér is one card carrying both senses -- the literal root and the insult --
because the insult was already in the locked deck, and adding the literal sense
separately would have produced two gyökér cards the moment that deck was
unlocked: one word's progress split in half, and both able to appear in the
same multiple-choice round with no way to tell them apart. build-private.mjs
now refuses to build when a headword exists in both decks, which is the check
that would have caught it.

Locked deck, 117 -> 123. The vulva set is filled in (hüvely, vagina, nuni,
puni, puna, suna beside the existing picsa, pina, punci) and cigány is added
with roma beside it, the way buzi already points at meleg.

Every one of the 123 entries now carries a usage note. Eight Hungarian words
landing on one English word reads as pointless duplication until you notice
English has the same ladder -- vagina / privates / fanny / pussy / cunt -- and
simply does not label it. So the clustered words each carry a compact Ladder:
line naming the whole set in order: cock, arse, shag and hell get one too.
The notes also say what the near-synonyms actually disagree about, which is
the part a dictionary never tells you -- balfasz calls you useless where
faszfej calls you obnoxious, bunkó is about manners and hülye is about wits,
and fasza, alone in the fasz- family, means "great".

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01RCxF3MSEkr2nyzX2ahsw3B

The first pass was a dashboard: four bordered notice blocks stacked above a
seven-column table, and then the same four requests rendered a second time as
cards underneath it. The request page carried a facts grid and twenty timeline
entries each wearing a date column, a glyph, a type pill, a track pill, a
heading, a note and a footer row. This site is a narrow flat column and it
already had the right idiom in the build log's dated list; I ignored it and
built something that belonged on a different site.

So: the cards are gone, the ledger is three columns with everything secondary
on one muted sub-line, and the three notice boxes collapse into a single
footnote with the long version behind a disclosure toggle. The facts grid is
now a run of prose. Timeline entries lost the glyph column, the rail and the
bordered pills.

Procedural steps -- acknowledgements, extensions, follow-ups, the requester
narrowing again -- now collapse into one foldable line reading "MAY 28 - JUN 02
: 2 extensions", so the events that actually produced documents stand out. A
run never spans tracks, because the lane filter hides by attribute and could
not otherwise reach inside a fold. Anything carrying documents stays out of a
fold whatever its type, and narrowing-request stays prominent: an agency
telling you the request is too big is a substantive act, and here it always
arrived as a letter worth reading.

Clow goes from twenty full-height entries to sixteen rows at roughly a third
the height. The stylesheet loses 267 lines.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01SwfZCx4FMsF1CaWt9PUgQW

pofoz (pofozni) joins Lili's pack, stored 3sg-first like the other 214 verbs.

bennünket and benneteket are added as accepted `forms` on minket and titeket.
They are not variants of a word, they are the same paradigm cell said the
other way -- which is exactly why they are forms rather than cards: two cards
would land in one "Finish the set" round with no way to tell which was being
asked for. As forms they grade correct while the wrong person still does not.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01RCxF3MSEkr2nyzX2ahsw3B

A new top-level section for records I've obtained under FOIA and state public
records law. No commentary -- the documents are the point.

Four requests to start: JWA Aviation at Clow International Airport (the
Village of Bolingbrook plus a parallel FAA request), two Chicago CDOT
intersections, and an Illinois Valley Airport fee-waiver fight with Josephine
County, Oregon.

The modelling decision worth knowing: a release IS a timeline event. There is
no separate list of releases -- any event carrying files is one. So the
timeline is the complete account of a request, and the per-dump file listing
falls out of it. Requests that fanned out across agencies or reference numbers
use tracks, which render a JS-free lane filter; the Clow entry interleaves
twenty events across three of them.

Documents live on S3 like the photos do, not in the repo -- GitHub Pages caps
a site at 1 GB and one release here is already 23 MB. Each file is recorded
with its size, page count and sha256, and the viewer is a small driver over a
vendored pdf.js so a scanned release renders the same on a phone as on a
desktop.

Machine-written scaffolding is labelled as such. The build refuses to publish
a summary with no generated_by/generated_on, and /foia carries a standing
disclosure that the timelines, titles and descriptions around the documents
are Claude's work while the documents themselves are untouched.

Two things that were quietly broken on this machine, both from execSync going
via cmd.exe, which cannot take a UNC working directory and falls back to
C:\Windows: the PII hook could not read its own staged diff, so -- failing
closed, correctly -- it refused every commit; and the version fell back to v1
with the changelog empty. Both now use execFileSync.

The phone pattern also no longer fires on records reference numbers and
sha256 substrings. A scanner that cries wolf on every FOIA commit just teaches
you --no-verify.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01SwfZCx4FMsF1CaWt9PUgQW

Ordering the queue was correct and still not enough. An admission takes a new
word only on the pace's beat -- one in four at "normal", one in eight at
"slow" -- and an admission only happens when nothing already in play is due.
A twenty-answer round therefore delivered one to three new words, so a
priority pack of eight took up to eight rounds to meet: right order, and
indistinguishable from no order at all.

A priority word now halves the beat. Halved rather than suspended, because
`fresh` is sorted priority-first and the locked deck is 116 words -- admitting
them on sight would rebuild the wall Session exists to avoid. Measured over a
round, ordinary -> priority goes 1 -> 3 at the slow pace and 3 -> 5 at normal;
"aggressive" is already at the floor of 2 and does not move.

Round length becomes a setting rather than a constant, defaulting up from 20
to 25 answers with 12 and 45 either side. It sits beside the new-word pace
because the two together decide how much new material a sitting reaches, which
is the thing that was actually wrong.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01RCxF3MSEkr2nyzX2ahsw3B

Two new packs. `lili` is eight words she taught across a table -- teherautó,
kamion, rendszámtábla, jelzőlámpa, túlsúlyos, újságos, pára, fülledt -- and
lámpa is amended in place rather than duplicated, because two lámpa cards
would split one word's progress and then face each other in a multiple-choice
round.

Both packs are `priority`: their unseen words are introduced before other
unseen words and their due words reviewed first. Deliberately not ahead of
due-ness -- a week-overdue review still wins -- because priority is only free
inside each group. Among 900 unseen words that ordering is the difference
between meeting a word tonight and meeting it in March.

The second pack ships encrypted. The site is static, so "only I can see these"
cannot be a switch: a switch is a line of JavaScript anyone can read with the
words sitting next to it. AES-256-GCM under PBKDF2-SHA256 x 600,000, unlocked
by passphrase in Settings and cached per device. Word ids are hashes, because
exportSave writes ids into the file the app tells people to carry around, and
the passphrase itself is never stored.

Every entry in it carries a register -- mild, vulgar, very vulgar, slur -- shown
as a word rather than a colour, since buzi and meleg translate to neighbouring
English words and are not remotely interchangeable.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01RCxF3MSEkr2nyzX2ahsw3B

Belt to the previous commit's braces. `Match` and `Memory` built their tiles from
`question.grid[0]!` -- a non-null assertion over an optional field, which is a promise rather
than a check. The promise was broken exactly once, by a component being handed the next question
before it was swapped out, and the result was an error card rather than a missing round.

That race is fixed at the source. The assertion was still a lie: a mode that can throw on its own
props is a blank screen waiting for the next mistake to find it. Both now return early when there
is no grid to lay out.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01MCGSQM7Krukmjv1Y7VigDe

This is the error card George kept hitting, and it was mine. Loading the game components on
demand meant holding the component in its own `$state` and assigning it from an `$effect` -- but
`question` changes the moment an answer is recorded, and an effect does not run until after the
DOM has been updated. So there was one render where the *new* question was handed to the
*previous* question's component.

Most components survive that; the grid ones do not. `Match` and `Memory` read `question.grid[0]`
directly in an effect, so a question with no grid throws on the spot -- which is why it was
reliably "after the pair unflip game": whatever came after a memory round arrived while `Memory`
was still mounted. The word on the card, `fal` or `erdő`, was never the problem; it was just
whichever word came next. Pressing "Try again" worked because by then the effect had caught up,
which made it look like bad data rather than a race.

The component is now derived from the question rather than assigned alongside it, so the two
cannot come apart: it is either the right component or none at all. A chunk that has not arrived
yet renders nothing for a frame instead of rendering the wrong thing.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01MCGSQM7Krukmjv1Y7VigDe

Follow-on from the error card, and the other half of the same fault. `gamesFor` ends with a
floor -- "something must always be askable" -- that returned `choice-en` without checking whether
the learner had switched it on. So turning multiple choice off did not turn it off, and with most
games off the session leaned on that floor constantly.

That is also how the crash was reached. With nearly everything off, almost no word suits the one
remaining round, so `next()` fell through to its fallback on nearly every question -- and until
the previous commit that fallback served a round without checking it could be built. The matching
rounds read `grid[0]` directly, so the answer was an error card, over and over, in exactly the
place this was reported: after a memory round, on the question that followed it.

The floor now names a game only if it is allowed, and returns nothing when it cannot. `next()`
decides what to do with that, and it checks what it serves. `fallback.test.ts` plays every game
alone against the whole corpus and asserts two things: nothing invalid is ever served, and
nothing switched off is ever served.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01MCGSQM7Krukmjv1Y7VigDe

George hit the error card on `fal`, which is how I found this: when nothing in the rotation can
build a round for a word, `next()` fell through to a path that returned a round *without checking
it*. That was survivable while it was hard-coded to multiple choice, which can almost always
build. Once it started respecting switched-off games it could return any mode at all -- and an
unchecked round is a component rendering a payload it was promised would never arrive. `Match`
and `Memory` read `grid[0]` directly, so they throw outright.

Reaching it is easy: turn most games off in the settings panel and play. The fallback now tries
multiple choice, then anything else switched on, and only then introduces the word -- an
introduction always renders, because it needs nothing but the word itself. Showing a word is a
poor round; it is not a broken one. `fallback.test.ts` plays every game alone against the whole
corpus and asserts nothing is ever served that its own `valid()` rejects.

The card that appeared said "That round would not build", which is a sentence for whoever has to
fix it, not for someone in the middle of a session. It now says what happened, shows the word
anyway, and says the rest of the session is fine.

In the matching round the words shifted sideways whenever the key numbers appeared or a cleared
pair renumbered the column -- which is exactly when they are being read. The number's slot is
always in the layout now; only the glyph comes and goes.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01MCGSQM7Krukmjv1Y7VigDe

An audit of the whole app, then the fixes. Three findings were worth the exercise on their own.

`hungarian.ts` splits a word into Hungarian letters and nothing put them back: `letters('rossz')`
is r·o·sz·sz, and joining that plainly spells `rosszsz`. Unscramble deals those as tiles, so 31
words could not be solved at all, and the fill-the-blanks mask and the hangman board drew
spellings that do not exist. There is now an inverse, and a corpus test asserts the round trip
over all 2,214 words -- the test that would have caught it.

The `accent-miss` verdict is the idea the whole grading design is built around, and it was
invisible. Four games computed a sentence to explain it and nothing rendered that sentence;
`--near` was the least readable colour in the app at 2.29:1. Worse, in the one round built to
separate `hal` from `hál`, typing the *other word of the pair* was greeted with "right word,
check the accents" and passed as a stumble. The note now shows, every colour that carries text
clears AA, right and wrong carry a mark rather than only a hue, and the partner is graded wrong.

Every mode reported the same rating to FSRS, so a two-option guess and a word written from its
meaning were identical evidence and a word grew in stability on recognition alone until it
stopped being asked. `intro` graded too, advancing a card for a word that had only been *shown*.

Two new rounds, both built from vocabulary the class already has: vowel harmony (`-ban vagy
-ben?`, 1,155 words) and `hol? / hova? / honnan?` (18 complete place families). Harmony refuses
to classify mixed-vowel words -- the rule says the back vowel wins and it usually does
(`papírban`) and sometimes does not (`ábécében`) -- because a drill that is confidently wrong
one time in twenty teaches the exception as the rule.

Also: an error boundary, so a round that throws no longer blanks the page; progress can be saved
to a file and restored, since one browser key was the whole of someone's work; a round has a
finish line; a mis-tap can be undone; the 19 game components load on demand rather than before
the home screen can draw a button; and four headwords that arrived from the .docx extractor with
their `ő` or `ű` replaced by a space are repaired and pinned.

Co-Authored-By: Claude Opus 5 (1M context) 
Claude-Session: https://claude.ai/code/session_01MCGSQM7Krukmjv1Y7VigDe

Four real rounds prompted this, two good and two not, so I sampled the builders against the real
corpus -- 590 group sorts and 614 odd-one-outs -- rather than argue from four screenshots. Two in
five sorts had a bucket that is a grab-bag: "Describing things" is 306 unrelated adjectives, and
you cannot decide from a word's meaning whether it belongs there, which is the whole exercise.
One in three had a chip nobody could sort: "az is tetszik, hogy…", a sentence opener, or
"állomás (= vasútállomás)" with its cross-reference showing. Two in five odd-one-outs had a verb
or an adjective sitting inside a family of things.

The buckets were the easy half. VAGUE_TOPICS was written for exactly this and wired only to
odd-one-out, so the sort round had no topic filter at all. It has one now, and a tile filter that
every grid shares -- no phrases, no ellipsis frames, no cross-references, no bare affixes, in
sorting, odd-one-out, match, memory and the speed round alike. The sort round also says outright
where each tile belongs instead of leaving the component to pick the first bucket the tags happen
to match, which is how "antipatikus" -- an adjective meaning unpleasant -- became the correct
answer for "People & family".

The hard half was upstream. Tags are keyword matches on the English gloss, and the boundary was
[^a-z]: ASCII only, so every accent read as a word break. `bus` matched inside "Busó-walking" and
a carnival was filed under transport. `one's` made three verbs numbers. "(first person of kér)"
gave `kérek` the tags people, numbers and feelings, three of its four. And a compound is not its
first half: a hot dog was an animal AND weather, hand cream was a body part, a coffee table was a
drink, "to suit somebody" was one of only 25 clothing words. A general head-of-compound rule was
tried and rejected -- it moves bikavér out of drink and into body, worse than what it fixes -- so
twelve words are named outright instead.

Fifty everyday nouns had no keyword at all and sat in `general`, which every themed round strips:
pohár, kanál, kés, hűtő, garázs, tyúk, festő. A third of the collection was invisible to both
games, which is the other half of why rounds repeated.

Re-tagging now REPLACES rather than unions. Union is only safe while the tagger is too quiet; the
moment it is too loud, a union can never take the wrong tag back off, and that is how virsli
stayed an animal through three earlier passes. Words the grammar tables wrote by hand are left
alone. 127 words changed. `body` went from ten words to six actual body parts.

Nine headwords lost their ő or ű on the way in from the glossary -- `id` glossed "time",
`sz nyeg` glossed "carpet" -- and all nine shipped as unanswerable cards. Seven duplicated an
entry that came through intact; two were the only record of their word.

Every rule here was chosen so as not to suppress the two rounds that were good, and both are
pinned as regression tests. Excluding category names would have killed the first (ruhadarab among
clothes) and requiring nouns would have killed the second (egészséges among nouns) -- so neither
rule exists, and the 42% figure is fixed upstream instead, by `áll` no longer being clothing.

Also, two things asked for directly:

The speed round now holds its options back for a beat once a word is well known. The mechanic
already existed -- `recallPause`, which the ordinary choice rounds have used for months -- and had
simply never been pointed at the speed round. The word and the audio arrive at once and the four
answers land a second and a half later, so you have to retrieve the meaning instead of scanning
for it. The clock keeps running. No new switch: it turns itself on at level 3.

And a new round with no English on it anywhere. `cipőbolt` is a shop, not a shoe; `vörösbor` is a
wine, not a colour. Hungarian compounds are right-headed without exception, so the last element
says what the thing *is* -- one fact that unlocks a lot of vocabulary, because someone who knows
`bolt` and `cipő` can read a word nobody taught them. The round shows the compound and asks which
word it is a kind of, with the modifier always offered beside the head, because that is the answer
you give when you have read the spelling and not the meaning. 129 of them, both halves already in
the deck, nothing added. Asking instead which two words make it up would have been a spelling
puzzle -- the letters give that away without any meaning being involved.

Odd-one-out has also been playable from the keyboard the whole time except that it wasn't: it drew
1-4 on the tiles and had no key handler.

Co-Authored-By: Claude Opus 5 (1M context) 

Forvo is a crowd: every clip a different room, a different microphone, a different person's idea
of how loud to be. Wikimedia Commons has Lingua Libre, where native speakers record whole word
lists to one protocol in one sitting, and it is the difference between a set that sounds like a
recording and a set that sounds like twelve strangers. 927 of the deck's 2221 words had one, so
they use it now, and three words that had no audio at all finally do.

The filenames came out of the dump already downloaded for the conjugation work, so nothing had to
ask an API where anything lives -- Commons files sit under the first two hex digits of the md5 of
their name, which is arithmetic. One request per second with the backoff Wikimedia asks for, ogg
to mono mp3 through ffmpeg, about twenty minutes for the lot.

Reversible by construction: the Forvo clip keeps its own filename and only the reference moves,
so putting one back is a database edit rather than another download. `audioSource` records which
is which. One word, magas, 404s on Commons and keeps what it had.

Nobody was going to listen to nine hundred recordings, so they were checked by duration instead:
mean 1.14 seconds, none under a quarter of a second, one at 4.6. That is what a list of single
spoken words looks like, and it is what catches a whole example sentence hiding among them.

Attribution ships with them. audio-credits.json names the Commons file behind every clip, rebuilt
from the database each run rather than appended to -- an accumulating list drifts the moment a run
dies before writing it, which happened twice, and a credits file missing entries is worse than
none. Settings carries the one line the licence actually requires.

Two things the share taught along the way, both already written down for npm and both true here:
writing nine hundred small files straight to it stops dead after fourteen, and copying them one
at a time dies with EBADF after three hundred. Build on a local disk, hand the directory to
rsync. The corpus build now retries a failed clip copy for the same reason -- a transient failure
there does not fail loudly, it just leaves a card silent.

Also: verbs open in the word browser. Clicking a Hungarian word shows its six present-tense forms,
folded away by default because 207 of 2221 words have a table and six extra lines each turns a
list you can scan into one you have to read.

Co-Authored-By: Claude Opus 5 (1M context) 

Six present-tense forms for 207 of the deck's 214 verbs, and none of them is a new word. They
ride on the verb as a field, so the deck is still 2221 cards and not 3463, and a round reports
against the verb -- getting szeretunk wrong is a mark against szeret, where it belongs.

Harmony cannot be read off the last vowel. lakik takes lakunk and fekszik takes fekszunk; both
end in -ik and neither announces which, and a last-vowel rule gets 42 of the deck's 65 -ik verbs
wrong. alszik becomes *alszunk with an umlaut -- a form that looks entirely plausible and is not
a word, which is the worst kind of mistake to drill. So the class comes from the data: a 1.6 GB
English Wiktionary dump, streamed once into a 6.3 MB table of conjugation class, Commons audio
filename and IPA. hu-conj-ok on lakik says back harmony despite the ending, and alszik has a
class of its own so no detection is needed at all.

Sixteen classes implemented, three of them regular and covering 157 verbs, the rest one-offs
where jon swaps its stem for jov- in half the table. The phonology is where the care went: sz is
one letter, so kivisz is not a consonant cluster and takes kivisztek rather than *kiviszetek --
counting characters gets every digraph verb wrong. A doubled consonant is awkward only after a
short vowel, so all takes allsz and hall takes hallasz, and Wiktionary marks the same five verbs
the same way with short=y against long=y. Stems that drop a vowel drop it only before a vowel:
csorgok but csorogsz.

Then the whole thing is checked against Wiktionary's own form pages, 35,141 of them, tagged
structurally: alszunk is "first-person plural indefinite present of alszik" and alszunk-with-an-
umlaut is not a page at all. 357 forms confirmed, and in the three places where the rules and the
dump disagree the dump wins and the swap is printed. Absence proves nothing -- Wiktionary has
visztek and not kivisztek -- so only a contradiction counts.

Nineteen hobby verbs had no page: borozik, fenykepezik, sorozik, taxizik, the whole free-time
chapter. Those read their harmony off the suffix that derived them, since -ozik is back and
-ezik front by the same rule that governs every other suffix. Three of the nineteen have a page
for the bare stem and all three agree.

Two rounds, no new components: pick the form for a pronoun, and finish the table with one row
blank. Both are the paradigm rounds with a different table behind them.

Also, the pairs the learner actually named. csipos against cipos was missing because edit
distance calls it two substitutions -- the same score as six thousand pairs of unrelated words.
It is not two mistakes; it is one word written the way it sounds, so two near misses in one word
now count as one confusion. hal/hall was missing twice over: under the length floor, and rejected
as a stem with a suffix, which a doubled consonant is not. And amikor/mikor is gone -- both tiles
said "when", so the round had no answer.

Four verbs were filed under their infinitive by a phrase split and shadowed the real entry, so
szallitani and szallit were two cards for one verb. Merged, keeping the sense each loser added.
hal, cipos and fekszik added by hand as the missing halves of pairs, each checked against the
724,942-title Hungarian Wiktionary list rather than guessed.

Last: both drills were nearly unreachable. A pair round needs a word with a partner and a learner
who has met both, which is about one question in a hundred -- right for a mixed session, and no
way to ship a feature someone asked for. "Easy to mix up" and "Verb endings" now sit above the
topic chips and hand the whole list straight to a session.

Co-Authored-By: Claude Opus 5 (1M context) 

vilag and virag, ajto and auto, anyuka and apuka, bors and gyors. One letter apart and nothing
to do with one another, and an ordinary round never puts them together -- distractors are chosen
by topic, and these pairs almost never share one. So the mistake survives.

122 pairs, found by edit distance over Hungarian letters rather than characters, which matters:
sz is one letter, so sor and szor differ by one substitution and not by an insertion. Counting
characters gets every digraph word wrong in both directions. 119 of the 122 have a recording on
both sides, and six differ only by an accent -- bal/bal, ot/ot, fel/fel -- which the grader
already had a separate verdict for.

Three filters, all using signals the data already carried: a word and its own stem (asztal /
asztali) is related, not confusable; a form a word already lists is not a trap; and two cells of
one grammatical table are drilled by the paradigm games, so re-presenting them as a mix-up
teaches the wrong thing about them. Two cells of one verb are deliberately kept -- szeretek
against szeretnek is exactly the confusion worth having.

A minimum length of four letters, because below that one letter is most of the word: ad/az and
agy/ar are a letter apart and share nothing, and over half of all distance-1 hits were words of
three letters or fewer. Accent twins are exempt, since ot/ot is short and is the point.

Four rounds: hear one and pick which, see both and pick the one that means it, hear one and
write it, and see the difference said out loud -- "sz, not s" -- because a letter position would
be true and useless.

Computed at build time, about a second, and the answer does not change between page loads. The
engine's own letter tokeniser does the work, so the tricky part is written once.

Also: "Only what we've covered in class" is now a switch. 1933 of 2222 words were taught in a
lesson, a MagyarOK chapter or a class transcript; the other 289 are the grammar tables and the
old Anki import. It is off by default, and it ANDs with the pack switches rather than replacing
them -- packs choose which material, this chooses how far ahead of the class to go. That rule now
sits at the top of both CLAUDE.md files, because several things being built here could pull in
the whole language if left alone, and none of them is allowed to add anything on its own.

Co-Authored-By: Claude Opus 5 (1M context) 

Hungarian splits place three ways where English mostly manages with one: bent is "inside", be is
"inside" as somewhere you are going, bentrol is "inside" as somewhere you have come from. Every
word of hol/hova/honnan, itt/ide/innen and ott/oda/onnan was already in the database and none of
them belonged to anything, so nothing ever drilled them together -- which is the only way that
distinction is learnable.

Thirteen sets now: the question words themselves, here, there, home, inside, outside, up, down,
front, back, elsewhere, near and far. 22 new words, and 11 that were already present simply
joined a family and kept their own glosses.

Two judgements worth recording.

The question set gets its own row headers. Every other set is labelled hol? hova? honnan?, which
is ideal for a postposition and useless for the question words themselves, where the row would
hand over its own answer.

"Outside" has no middle slot. The hova? form of kint is ki, which is spelled exactly like the
question word "who?" and is a different word that happens to share a spelling. The database
models that as two entries; hu paradigms deliberately never forks one, so the set stays
two-thirds built rather than quietly stealing the question word out of its own family.

Also fixed on the way: the fourth clip of "hazulrol" has no Forvo recording, and one of the
twenty-two new words likewise -- both are marked rather than silently missing.

Co-Authored-By: Claude Opus 5 (1M context) 

"Three of these are numbers" is not a puzzle, and that is what the game was building. A round
offered brassoi apropecsenye against het, mobilszam and van -- "is, there is" -- because all
three carry a numbers tag, and another offered ertekezlet against nalam, mennyibe and
tizenegyedik. Nothing links those. There is no odd one out.

Two causes, neither in the game.

Tags are keyword-matched against the English gloss, and 946 words from the textbook import had
none at all: they were written with categorize(gloss, pos) when it takes a single object, so
every one of them got the "nothing matched" fallback. scripts/ was outside the tsconfig include,
so nothing checked the call. Both are fixed, and 518 words gain real tags -- colours goes from 3
to 15, nature 5 to 14, health 10 to 13, and three topics become offerable that were too thin
before. tag-words.ts was also documented as safe to re-run while it assigned tags wholesale,
which would have deleted all 213 tags the grammar tables declare; it is additive now.

Families are also chosen better: they must share a topic that names a kind of thing -- food,
drink, animals, clothing, transport, and so on. Time, work, places, people and numbers are
states and abstractions, and three of them together look like a group without being one. The
stranger stays unrestricted, because a pronoun against three health words is a perfectly good
question and restricting it throws away a third of the rounds for nothing. Topic choice is
weighted by the square root of the topic's size, so a four-word topic is not as likely as a
two-hundred-word one, and the round is not always about food.

Odd one out can now be built for 51% of asks rather than 31%, and the four reported rounds are
all unbuildable.

Four headwords were also repaired. The double-acute o and u drop out of one of the PDF's fonts,
which was known for the example sentences -- they were left out -- but it hit "csipos", "szuk",
"szolo" and "buszsofor" too, and they slipped through as phrases because a space reads as one.
The import's own validation could not have caught them: it matched 719 extracted headwords
against the database, which says the extraction is sound and says nothing about the broken
ones, since a word with a hole in it matches nothing by construction. The importer now refuses
a noun, adjective or verb that contains a space.

Co-Authored-By: Claude Opus 5 (1M context) 

Match up was a fixed board of five pairs at every level, and with the columns equal the last
few pairs are free: whatever is left has to go together, so the end of every round tested
nothing at all.

Now it grows -- four pairs, then six -- and at the hardest level some tiles have no partner.
Which side carries the spares changes from round to round, so "the longer column is the one
with extras" does not become the new shortcut either. A decoy is always a word already met,
because an unfamiliar one is not a distraction so much as noise, and never reads the same as a
real tile on its side, which would make one of them unanswerable rather than merely hard.

The board says so when there are spares. Finishing with tiles left over is the correct outcome
there, and without a word for it the round just looks broken.

No special case in the matching itself: a decoy's id matches nothing in the grid, so it can
never clear and pairing one is wrong through the path that was already there.

Co-Authored-By: Claude Opus 5 (1M context) 

Hungarian counts cs, dz, dzs, gy, ly, ny, sz, ty and zs as single letters of the alphabet, each
with its own sound and its own place in the ordering. Every part of the app that worked letter
by letter was splitting them with [...word], which is wrong in three places at once: hangman
offered "s" and lit up half of "orszag", unscramble dealt "s" and "z" as separate tiles for a
word containing neither, and the fill-the-blanks mask could hide one half of a letter and ask
for something that does not exist.

New engine/hungarian.ts tokenises a word the way the language does. The subtle part is that a
long digraph doubles only its first character -- "asszony" is a-sz-sz-o-ny, not a-s-sz-o-ny --
so a naive reading hands a guess of "s" a free hit. The one case it cannot resolve is a compound
seam, where "kisszek" really is kis + szek; telling those apart needs morphology the app has
not got, and doubling is much the commoner case in ordinary words.

Hangman's alphabet now offers the digraphs, and typing still works: a key that could begin one
is held for a moment, so pressing s then z guesses sz rather than burning a life on s and then
guessing z. Otherwise nine letters of the alphabet would have been mouse-only.

The scramble test asserted character-level behaviour and now asserts the letters of the word --
cseresznye is seven tiles, not ten. The reachability test was pinned to a single seed and broke
because tokenising differently consumes the rng differently, which says nothing about whether a
game can be reached; it now checks across several seeds.

Co-Authored-By: Claude Opus 5 (1M context) 

execSync buffers the whole diff and throws past its 1 MiB default, which the hook never
raised. Committing 780 audio files produced a 1.3 MB diff, execSync threw, and the catch
reported it as "not in a git repo? bailing" and let the commit through -- so the scan opted
out of precisely the commits most worth scanning, while printing something that read like a
harmless note.

Raised the buffer, and made the failure path fail closed: a scanner that cannot read the diff
and stays quiet is worse than no scanner, because it looks like it passed. --no-verify is
still the deliberate way past.

Checked afterwards: nothing sensitive reached the published site. The only pattern-matches in
everything added today were two FSRS weight vectors inside the minified bundle, which look
like coordinates and are not.

Co-Authored-By: Claude Opus 5 (1M context) 

"MagyarOK 1 · 3" asked the reader to know the series numbering. The level is the part that
means something, and it keeps reading correctly when the A2 book arrives.

Co-Authored-By: Claude Opus 5 (1M context) 

Most of this was already written and simply unreachable. `themedQueue`, `WordIndex.topics` and
`topicProgress` were finished and tested when themed rounds went in, but `game.topic` could only
ever be set from the argument `start()` had just been handed, so nothing could reach them.
`Progress.days` has been recorded on every session since the first version, documented as
"Drives the streak", while `computeStats` took it as a parameter and never touched it.
`sfx.levelUp()` was written for a word graduating and never called, though the store already
knew the exact moment. `Session.remaining` had no reader. `Stats.due` was computed every render
and thrown away.

So: a streak and a "due now" count on the home screen, both from data already on disk -- an
existing save shows a real number immediately. Yesterday still counts as unbroken; a run that
dies at midnight before you have had a chance to play is a punishment for sleeping. The best
run is shown beside the current one so a break is not the whole story.

A word list, which the app has never had. Pick a topic or a chapter and see every word in it,
weakest first, with how well you know each one -- and practise just those, or just the ones you
have never seen. It could tell you 939 words were unmet but not which ones.

Topic chips on the home screen, with two-tone progress: FSRS needs four reviews before a word
counts as learned, so a bar driven by that alone sits empty for a fortnight across every topic
at once. "Describing things" and "Pronouns" are filtered out -- 179 and 128 words, and neither
is a theme anyone would choose.

The trouble list is now something you can act on: each word starts a round, or practise all
eight. It had been eight inert pills on the one screen meant to answer "what am I bad at".

Co-Authored-By: Claude Opus 5 (1M context) 

1256 -> 2202 words. The corpus was being outrun -- 897 answers spread over 123 words while 939
sat untouched -- and a weekly class cannot fill it fast enough. This is the glossary of
MagyarOK 1, the book the class actually uses, which is a far better source than any frequency
list: it is the vocabulary these lessons are built on, and it arrives already divided into the
chapters they are taught in.

946 new words, and 760 words we already had now record the chapter that introduces them, so a
chapter can be practised or switched off whether or not it came from the book.

Eleven switches on the home screen, one per chapter plus class lessons, grammar tables and
everything else. Switching one off removes those words everywhere -- the queue, the counts, and
the pool wrong answers are drawn from, which is the one that would have gone unnoticed: a
chapter you had turned off would have gone on supplying distractors. Choices are stored as
explicit yes/no rather than a list of exclusions, so a MagyarOK 2 pack can later ship switched
off until it is asked for; a deny-list cannot say that.

Extraction was checked rather than assumed: 719 of the glossary's headwords match words already
in the database exactly, which a mangled parse could not do. Example sentences are deliberately
left out -- the double-acute o and u drop out of the font they are set in, turning "baratnomnek"
into two words. Headwords are unaffected.

The book's own English is kept as the gloss. It is the meaning the class will be tested on, and
rewriting it would only introduce drift. Verb government becomes the case it asks for, so
"ad vkinek vmit" reads "to give sth to sb (+ -nak/-nek, -t)". 781 of the new words have a Forvo
recording; 68 have none and are marked.

Co-Authored-By: Claude Opus 5 (1M context) 

1176 -> 1256 words. The corpus was starved exactly where a beginner needs it most: nine words
marked num. and four marked ppn. across the whole collection, with "de", "vagy" and "ha"
missing outright.

Added as sixteen more grammatical families plus twenty-eight plain entries:

  ordinals            elso..nyolcadik, then kilencedik..ezredik
  postposition triads alatt/ala/alol, mogott/moge/mogul, and four more -- the row headers are
                      hol? hova? honnan?, which is how the distinction is actually taught
  the vala- series    valahol / sehol / mindenhol / barhol / akarhol, one family per question
                      word, because the prefix carries the meaning and the stem never moves
  joining words       nelkul, szerint, miatt, ota, mulva, de, vagy, ha, amig, mielott...

Two judgements worth recording. The series tables leave slots out rather than inventing
symmetry: there is no *semmikor and no *mindenhogy, so "when" takes soha and mindig and "how"
has no universal at all. And the suffix is appended to a gloss only where it distinguishes one
near-identical family from another -- "second (-dik)" and "under (at) (under)" are noise.

Every one of the eighty new words has a Forvo recording; none was missing.

Co-Authored-By: Claude Opus 5 (1M context) 

Two changes, one of which is a bug fix that matters more than the feature.

`Session.admit()` took every due word before it would look at a new one. A word only retires
after two consecutive correct answers, so a backlog of R words cost at least 2R answers before
the first introduction -- in practice, 897 answers spread across 123 words while 939 had never
been shown once. Fifteen minutes of play could produce no new word at all. New words now
interleave on a ratio, at a pace the learner sets: slow, normal or eager. The regression test
plays a session with a 150-word backlog and fails if nothing new appears.

The corpus goes 1062 -> 1176 words. The new ones are the closed grammatical tables: the
personal pronouns declined for case, the ez/az demonstratives, and the ki/mi question words --
114 cells across 16 families, hand-authored because an invented form is indistinguishable from
a real one and would be drilled as fact. Four games teach them, and every round draws only from
one family, so the ending is the question rather than something you can skip past.

Also: the trouble list finally says what the words mean. It had been eight bare Hungarian words
with the English in a title tooltip, which does not exist on a phone -- on the one screen meant
to answer "what am I bad at".

Co-Authored-By: Claude Opus 5 (1M context) 

The pre-commit hook came back from a Windows checkout rewritten with CRLF endings, which
turns its shebang into a request for an interpreter literally named "bash\r". The hook then
does not run at all. It is the PII scan for a public repo, so it fails in the quietest and
least useful way possible: the commit succeeds and nothing is checked.

Pinning it to LF costs nothing -- it is the only shebang file in the repo, and the working
tree and index are already LF, so this changes no content today. *.sh is covered too, for the
next script that gets added.

Co-Authored-By: Claude Opus 5 (1M context) 

The published corpus goes 879 -> 1062 words, all vetted, 988 of them with a pronunciation
clip. 246 new clips ship alongside.

The words come from the 50 lesson plans, which imports now leave where they are instead of
moving them out of an inbox as they are read. Extraction offered 791 proposals: 522 inherited
a decision already recorded in words.json, 57 were dropped again, and 212 were decided by
hand. That pass fixed 21 typos carried in from the source documents (füllhallgató ->
fülhallgató, küldföld -> külföld, öccs -> öcs, trolli -> troli), collapsed infinitives to the
third-person singular the rest of the deck is written in, and dropped 73 entries that were
headings, example sentences or duplicates rather than vocabulary.

Only the built site lives here; the word database and the tooling that fills it are in the
hungarian repo.

Co-Authored-By: Claude Opus 5 (1M context) 

Odd one out was coming up blank. Its builder correctly returns nothing when there are not four
words from topics the learner has actually met -- which got much more common after last change
tightened what counts as a fair family -- but the empty round was served anyway and the
component had nothing to draw. Group sort could fail the same way.

Every game now declares what a built round needs to render, and the session tries its
candidates in turn and serves the first that actually builds, falling back to multiple choice.
Fixing it per game would have left the next builder free to do it again.

Choice questions now hide their options for a few seconds once a word is familiar: three at
level 2, four at level 3, none at level 1 where the word is still new enough that staring at a
blank card teaches nothing. A list of answers on screen turns recall into recognition -- you
spot the right one instead of remembering it, which is the easier task and the less useful one.
Enter cuts the wait short. Speed round is deliberately left alone, since a deliberate pause is
the opposite of what it is for.

63 component words of multi-word entries are now entries themselves, written by hand rather
than fetched -- Wiktionary is good at citation forms and poor at what actually turns up inside
a phrase (`vagyok`, `benne`, `órakor`, `szénsavas`). Phrases with every part covered went from
10 of 55 to 51, and none is left as an opaque lump. Each new word records the phrase it came
out of, so it can be audited or dropped later.

Odd one out was "sometimes a stretch", and it was, for two reasons. It could pick `describing`
as the family -- 163 words whose only shared trait is being adjectives, which nobody can see as
a category. And it only checked that the odd word lacked the one tag the family was grouped by,
so a word carrying a second shared tag could honestly belong. Now vague topics cannot be a
family, and the stranger must share no topic at all with any of the three.

That surfaced worse tagging underneath. `sárgarépa` was filed under colours, because its gloss
says "carrot (lit. yellow root)"; so were parsnip, apricot, both wines and black pepper. And
Hungarian lemmas were being matched against English keyword lists, which made `fog` weather
rather than a tooth. Tags now come from the gloss alone, colours needs a sense that is a bare
colour, and a few English idioms are dropped first because their words are not about their
subject -- "on the other hand" is not anatomy. Colours honestly has 3 words, not 11.

Hangman had no "I don't know", so being stuck meant burning lives on letters you knew were
wrong. The speed round now says each word as it comes up rather than after the answer, which
was a beat too late to be any use.

Twenty-two of the 816 glosses quote the word they define -- "lángos (deep-fried flatbread)",
"fork (also: villa, large house)", "take-out (elvitelre: to take away)". That is fine in a
dictionary, where you see both sides at once, and useless in a game, where the gloss is the
prompt and the word is the answer.

engine/gloss.ts trims them at the point of asking rather than editing the corpus, since Anki
still wants the full entry. It drops bracketed asides that quote the word (matching inflected
forms too, because Hungarian agglutinates) and a leading repetition of the word itself. The
trailing guard matters: an earlier cut turned "deciliter" into "liter" and "motorcycle" into
"cycle", which is worse than the leak. Nineteen of the twenty-two now read cleanly; the three
that cannot -- hotel, modern, park, where the word simply is the English -- are kept out of the
games that ask you to write it.

Phrases are now taught as their parts. Fifty-six entries are multi-word, and meeting "egész
hétvége" as one opaque lump wastes the fact that "egész" and "hétvége" are already cards. The
intro and the answer card list whichever parts exist, each speaking when tapped. Grammar words
and punctuation are skipped, so "Hány óra van?" resolves to hány? and óra rather than to van.

Audio now plays wherever the Hungarian is on screen: Pár / Nincs pár, the prompt side of
choice-en, group sort as each word lands, and -- as asked -- after correctly picking the
Hungarian in choice-hu, which waits for the clip instead of cutting it off.

A single Enter did two jobs. The input submitted the answer, and the same keystroke carried on
bubbling to the window listener that dismisses the answer card -- so the card was built and torn
down inside one keypress, and no typed answer ever showed a result. It affected all four
free-form games: type it, fill the blanks, unscramble, hangman. The dismiss handler now ignores
keys that arrive in the same breath as the answer, which fixes the class rather than the four
instances.

Group sort printed the English gloss under every word, so sorting needed no Hungarian at all.
The meaning now appears once the word is committed to a group: before that it is the answer,
after it is the thing worth learning.

Affixes are no longer asked in games that want the word written out. "-féle" is glossed "-like,
of the sort of", and there is no route from that back to the spelling.

Balloon pop is gone -- it did not work and was not worth repairing.

Settings now record which games are switched *off* rather than which are on. The allow-list
meant a save written before a game existed pinned it off forever, which is why four new games
shipped and nobody who had already played ever saw them.

Also drops the PII hook's scan of compiled site bundles: minified output where 4294967296 (2^32)
reads as a phone number and an FSRS weight array reads as GPS coordinates.

The difficulty bands were keyed to a familiarity no word ever reached. A word retires from a
session after two correct answers, so it never climbed past ~3 -- which put nine of the fifteen
games behind a gate nothing could open. A first session served four game types all evening.

Games now open at the familiarity they are actually useful at, and only the pure recognition
ones close again: picking from a list is worth nothing once you know the word, whereas a puzzle
or a timed round stays worthwhile. Twelve of fifteen games now appear in a first session.

Also adds four Wordwall-style modes (balloon pop, group sort, hangman, speed round), scales
option counts, timers and lives with familiarity, and stops free-form modes from silently
skipping past a correct answer -- they now show the word, its meaning and how it sounds.

rotation.test.ts plays real sessions and counts what comes out, because testing gamesFor() with
a made-up familiarity is exactly how this got through the first time.

A static word game built from the lesson-plan vocabulary database in the
hungarian project. 816 reviewed words, 741 with native Forvo pronunciations,
16 topic rounds and 10 game modes.

Progress lives in localStorage and is scheduled with FSRS, so everyone who
opens the link keeps their own private spaced repetition with no account and
nothing to maintain.

Co-Authored-By: Claude Opus 5 (1M context) 

Co-Authored-By: Claude Fable 5 

- ink-3 darkened/lightened per theme to pass WCAG AA (was ~3.1:1)
- footer h5 -> h3 for sequential heading order
- local post images get width/height at build (no CLS)
- home page prefetches 3 shuffle photos instead of all (~46MB -> ~2MB)
- build log entries count as musings in the numbers box
- build log copy touch-ups

Co-Authored-By: Claude Fable 5 

Co-Authored-By: Claude Fable 5 

Co-Authored-By: Claude Fable 5 

Co-Authored-By: Claude Fable 5 

Co-Authored-By: Claude Fable 5 

Co-Authored-By: Claude Fable 5 

images:
- process-staging.mjs now renders 6 keys per upload: jpeg + webp for
  full/med/thumb. metadata.yaml entries get s3_webp/s3_med_webp/
  s3_thumb_webp/lqip in addition to the existing jpeg urls.
- makeLQIP renders a 24px-wide blurred jpeg and embeds it as base64
  (~600 bytes per image) in metadata as `lqip`.
- new helper `pictureFor(im, size, alt, extraAttrs)` in build.mjs:
  emits  with WebP source + JPEG fallback + inline lqip
  background-image when available.
- photo page main image, gallery feed view, and post-inline images
  all use pictureFor (or its inline equivalent in the markdown body
  rewriter). older entries without webp variants fall through to a
  plain .

new scripts:
- scripts/backfill-webp-lqip.mjs: walks all metadata + post
  frontmatter, downloads each S3 jpeg that's missing webp, renders
  + uploads variants, writes lqip base64, patches the yaml/md.
  idempotent.
- scripts/check-images.mjs (`npm run check:images`): HEADs every
  S3 url referenced from content. exit 1 if any are unreachable.
- scripts/check-links.mjs (`npm run check:links`): crawls _site
  html and HEADs every external href. opt-in, run after build.
- scripts/dev.mjs (`npm run dev`): watches content/templates/
  design/copy.yaml + scripts/build.mjs; rebuilds on change;
  serves _site on :4321.

performance:
- preconnect to s3 bucket in 
- width/height attrs on every  (uses exif dims already in
  metadata) — eliminates layout shift as photos load
- fetchpriority="high" on the photo-page hero

accessibility:
- prefers-reduced-motion: kills hover-grow transforms + cuts
  transition durations to 0.01ms for users who set the preference
- :focus-visible: 2px accent outline + accent-3 link bg
- skip-to-content link (visible on focus)

feeds + favicon:
- /atom.xml alongside /rss.xml (some readers prefer atom)
-  for both feeds in head
- design/favicon.svg: simple ~/red monogram, wired in head + copied
  to _site at build

new pages:
- /musings/tags index — alphabetical list of all tags + counts
- /musings/YYYY year archives — group posts by year, with cross-
  links to other years
- related-posts section on each post page (up to 3 by shared-tag
  score, then recency)

random-post link:
- "take me somewhere" link in the footer; clicking picks a random
  musing url client-side. all post urls are emitted as a json
  attribute on the link.

each gallery page that has sub-albums now renders a small toggle
(top-right of the gallery) with three views:

- folders (default): sub-albums above, then loose photos in this
  folder below
- grid: every image in this subtree (this album + every descendant)
  in a single tight VSCO grid
- feed: same set, but as a vertical scrolling stream — large centered
  images with caption + album link below each

implementation:
- build.mjs computes `allDescendantImages` (deduped by src) per node
- template renders all three modes; CSS toggles which is visible via
  :checked sibling-selector — no JS required for the toggle
- a tiny JS hook persists the chosen mode in localStorage so it
  survives page navigation
- empty galleries (no sub-albums, no photos) fall through to the
  empty-state message, and single-folder galleries skip the toggle.

- /license page: full CC BY-NC 4.0 reuse terms with body + bullets +
  commercial section, all driven by copy.license in copy.yaml.
- footer adds a 'license' link, an inline copyright line
  '© George Main · CC BY-NC 4.0', drops the 'est ' line.
- changelog: cleaner ol-based row layout, version+date+hash inline at
  top (hash is a clickable github.com/.../commit/ link), subject
  on its own line, optional body collapsed in a 
 summary.
- outback restructure: misc is no longer a sub-album. All 19 misc
  photos moved (S3 mv) into /media/australia/outback/ as loose
  photos. content/media/australia/outback/misc/ deleted.
- new content/media/australia/tasmania/metadata.yaml: cover =
  russell-falls (better than the default first-descendant).
- mona cover = fat-car (more striking than cement-truck for the
  album thumbnail).
- mona alts: noodles-on-pedestal renamed to rubber-bands-on-pedestal
  (s3 mv + alt); 'red text piece by an artist commenting on artists'
  → 'Red text piece, MONA.'; bit.fall flipped to 'looking down at';
  'Textural close-up resembling stalactites' → 'Wall of rock inside MONA.'

- build.mjs: new optional `cover:` field on a gallery's metadata.yaml
  selects which image becomes the album cover (instead of images[0]).
  cover lookup recurses into descendant galleries so virtual parents
  (e.g. /media/australia, /media/australia/outback) can pick a cover
  from any nested album.
- new content/media/australia/metadata.yaml + .../outback/metadata.yaml
  (parent nodes that were previously virtual-only) — both pick
  uluru-sunset as cover.
- misc covered by northern-territory-sign; kings-canyon by
  lost-city-overview.
- s3 renames + alt fixes for the misc outback gallery: memorial-bell
  → centre-of-australia (it's the Centre of Australia cast-iron
  monument); red-dirt-campground → yulara-coach-campground;
  desert-camp → yulara-camp-overview; glamping-tents-overview →
  yulara-eco-tents-overview; camp-from-above → yulara-camp-from-above;
  glamping-tents → yulara-eco-tents.
- kings-canyon: overhang → rim-walk-squeeze (alt: rim walk squeeze
  between two rocks); person-on-rim → canyon-rim-trail (no person
  in shot); shelby-on-rim deleted entirely.
- uluru: uluru-from-viewpoint deleted.
- photo page < / > arrows moved beside the image (flex row) instead
  of overlaid on top of it.

- drop the musings index subtitle ("Long-form notes...") — not needed.
  copy.musings.index.intro removed from copy.yaml and template.
- fix random-photo shuffle: allImages was including the same image
  twice when it appeared in both a real and a synthesized-from-post
  gallery. dedupe by src URL so consecutive shuffles always land on
  a different image.
- preload all shuffled photos via new Image() at homepage load so reroll
  swaps are instant from the browser cache.
- hover-grow added to: album cards (.sp-album), homepage random-photo
  widget (#photo-img), photo-page main image. previously only the vsco
  grid cells and post figures had it.

user feedback: v3 paper-card design felt worse (abrupt bottom edge,
broke the stats grid boxes). preferred the v2 chrome but liked the
new paper color. so revert tokens.css + site.css + templates/ to
commit 6ce9249 (v2-final), then patch in only the cool paper palette
(#f2f1ec bg, #b21f1a accent) from v3.

what stays from v3: just the color palette.
what's restored from v2: section markers, paper-card-free shell,
stats-grid boxes, all template structure.

design bundle update from claude design tool:
- new tokens.css: cool off-white paper (#f2f1ec), no beige. deeper
  burgundy red accent (#b21f1a). subtle SVG noise texture. gutter 40px.
- new site.css: content sits on a paper card (.site-shell now has
  background + border-left/right). section markers are now a centered
  '· · ·' ornament + small italic '— a random musing —' kicker, not
  '// 01 · words' mono captions. headings are serif weight 600, not
  mono. tags are serif italic. .sp-tree replaced by nested ul. new
  .sp-photo-frame + .sp-photo-meta + .sp-photo-counter classes.
- templates:
  - drop mono '~/musings · 67 posts' breadcrumbs everywhere.
  - musing/category/musings-index now share the .musings-layout with
    a tree sidebar (was: index didn't have the tree). class renamed
    musings-main -> musings-body.
  - serif italic crumbs with › separators (Musings › Travel › Japan).
  - written-out dates (May 11, 2026) instead of 2026·05·11.
  - drop file-path lines from random widgets — title is the title.
  - photo page switches table → 
 + n
    of total counter + keyboard hint.
- build.mjs: fmtLongDate helper, exposed to templates. shuffled musings/
  photos + recent + stats all carry dateLong now.
- footer: drop "est " line; version 
 gets class="version".

5 photos from sony rx100 vii, ~20mb of originals on s3, ~0 bytes in repo.
metadata.yaml at content/media/australia/tasmania/mt-field/ with full
camera/lens/settings + dimensions captured from sharp.

also:
- alt text is now optional with a warning, not a hard build failure
  (build.mjs + process-staging.mjs)
- build auto-creates virtual parent gallery nodes for any folder depth,
  so /media/australia and /media/australia/tasmania resolve as routes
  even without metadata.yaml at those levels

- homepage random photo: max-height 60vh
- photo page: max-height 82vh, width auto
- post figures full-width: max-height 75vh with object-fit contain
- inline-floated images keep natural sizing

scripts/process-staging.mjs:
- reads full EXIF (camera, lens, ISO, aperture, shutter, focal length, dimensions)
  in addition to date/GPS
- scrubs the original (strips all EXIF including GPS via sharp's default behavior)
- uploads scrubbed full-size original to s3://georgemain-com-media//
- writes a smaller 1400px local reference into content/media/ for build-time
  thumb/med generation (keeps repo lean)
- metadata.yaml entries now carry { file, s3, alt, title, date, location, exif }

templates/media-photo.eta:
- camera/lens/settings/dimensions table when exif present
- "view full size ↗" link to the S3 original when s3 is set,
  falls back to local file for older entries

- templates/home.eta: random/next/read/open buttons now read their labels
  from copy.yaml's home.random_musing.* and home.random_photo.*. previously
  hardcoded.
- design/site.css: dropped the .tg-* class block (~64 lines). these were
  styles for the token-viewer page in the design canvas, never used in
  production.
- scripts/build.mjs: error messages now lead with the file path and name
  the exact missing field — e.g. ": image 'foo.jpg' is missing
  required 'alt' text" instead of "Image alt missing in ...".

clicking a gallery cell now opens a dedicated photo page at
/media/// instead of the raw image file. each page:

- big image with caption (alt text)
- metadata table: title, date, where, album, original file link
- prev / next nav (wraps around at the ends)
- counter "3 / 12"
- left/right arrow keys also navigate
- ↑ back to  link

works for both real galleries and synthesized-from-musing galleries.

- before: synthesis was opt-in via gallery.path
- now: any post with images auto-gets a gallery. default path = post
  category minus first segment + slug. override with gallery.path,
  disable with gallery: false.
- trim verbose alt text on three images
- make nunobiki falls inline (small) instead of full-width
- add CLAUDE.md

- renderTreeHtml outputs nested 
 so long names like
  "great-barrier-reef" wrap naturally instead of being clipped by the
  fixed sidebar column.
- depth indent via padding-left + dashed left-border guide (cleaner than
  ASCII glyphs that depend on white-space: pre).
- mobile (<=900px): tree is hidden behind a "browse ▾" toggle (pure CSS
  checkbox hack). desktop unaffected — tree always visible.
- previous spaces+pre approach is gone; no more clipping or weird wrap.

also fixed earlier issue: vsco gallery captions now appear BELOW the
cell on hover, not overlaid on the photo.

opt-in via gallery: { path, title, subtitle, location, date } on a musing.
build walks ancestor segments to create virtual parent gallery nodes so
breadcrumbs and album-tree work. images stay single-source in the post's
media/ folder (no duplication). synthesized gallery + each card show a
"↩ from the musing: