A review pass, and it found the one that would have cost somebody real chips.
Side pots were only ever cut in runout() — the path taken when the betting
stops because nobody is left able to bet. But a hand reaches a showdown with an
all-in player in it and the betting having finished perfectly normally: a short
stack shoves, two players who still have chips behind call, and then keep
betting past them street after street to the river. Nothing was cut. One pot,
everybody eligible, and the short stack takes the lot — every chip the deep
players put in after they were already all-in, money that could never have been
lost to them. All-in for 100 against two players who each put in 500, and the
best hand collects 1,100 instead of the 300 it was playing for.
Chip conservation never saw it. The chips balance perfectly; they just land in
the wrong seat. And every browser session went through runout(), because a
player shoving is what ends the betting. It took reading the code.
Also from the review: play() dereferenced a table it had just been handed as
null, the top-up button offered chips the wallet could not cover, and the
trainer's ETA was sixty thousand hands optimistic on the first line it printed.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
Phase 4. Hold'em, and it's the only table in the casino that is a session
rather than a game: you buy in, play as many hands as you like, and leave with
what's in front of you. So the live row spans hands and chips cross the border
exactly twice. Everything in between is inside the engine.
The bots move inside ApplyMove, as UNO's do, which is what keeps poker off a
socket: shove all-in and the flop, turn, river, showdown and payout all come
back in one response, as a script the felt plays back.
The CFR policy the plan called "the single highest-value asset in either repo"
was never read. Not once, in the whole life of the game: the trainer wrote its
info-set keys under IP/OOP and the runtime looked them up under BTN/SB/BB, so
every lookup missed and fell silently through to a pot-odds heuristic. Nothing
looked broken, because a policy miss is not an error. And it was the wrong
policy anyway — ten big blinds deep, trained on a tree where a call always ends
the street, which is not poker. So the trainer is rewritten to play the real
engine through the real reducer, at every stack depth the table deals, and the
trainer and the table now build the key with the same function so they cannot
drift apart again. A test fails if the bots stop finding themselves in it.
Three money bugs, and the tests earned their keep. Chip conservation across a
hundred sessions caught an uncalled bet that minted chips. A var-init ordering
trap meant every card was identical, every showdown tied and every bot believed
it held exactly 50% equity. And the browser caught the rake being silently
zero — the tier said 5 meaning percent, the casino handed it 0.05 meaning a
fraction, and integer division took the house's cut down to nothing.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
A code review of the uno table found the stuck guard had never once fired.
It counted how many bots had passed in a row and wanted more of them than
there are seats — but the bot loop hands the turn back the moment it comes
round to you, so the count could never get there, and your own empty-handed
pass was never in it. A dead table just passed the turn round forever. That
is not an ugly ending, it's a game you cannot finish, and a game you cannot
finish is chips you cannot cash out. So it asks the real question now: is
there anything to draw, and is anyone holding a card that goes.
And the table let go of itself too early. busy came off when the request
landed, not when the script it came back with had finished playing — so for
the seconds a bot lap takes, you could click a card at a board the server
had already moved past. It comes off at the end now, like the other tables.
Also: left: 0 was being dropped on its way out the door, which is the one
number that matters (the seat that just went out), the deck counter didn't
come back after a reshuffle, and hoisting fly() into flyNode() had quietly
flattened the chip arc on every other table in the room.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
Driven in a browser for the first time, which is where three bugs were.
Every visit to /games/uno was a 500: the page was never added to the list
server.go parses into the games template set, so render() answered "unknown
page". The casino tests all call their handlers directly and never go through
render(), so nothing saw it. TestEveryCasinoPageRenders now walks the mux and
asks for every page the casino routes to.
The play script hid the first card that lit up rather than the one you clicked,
so playing any other playable card made an innocent card vanish. And on a phone
the discard sized its box but not its card, which takes its size from --uno-h,
so a full-size card hung out of a small hole and covered the colour in play.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
UNO, played for chips. You stake once, sit down against one to three bots,
and going out first pays the table: 2.2x heads up, 3.6x against a full house.
Anybody else going out first takes the stake. The table size is the tier,
because it is the only dial UNO has.
The bots move inside ApplyMove. A game with opponents is normally where you
reach for a socket, and the plan says solo UNO must not — so one request plays
your move and every bot turn behind it, and hands back the whole lap as a
script the felt plays in order.
The RNG is in the state rather than an argument to it: the bots choose and a
spent deck reshuffles, so the engine needs randomness mid-game, and there is no
generator alive across requests to pass in. The seed rides in the state and each
step derives its own. The game still replays exactly as it fell.
The zero value of Color is Wild, and that is the whole point of it: a wild
played with the colour field missing from the JSON must be refused, not
quietly played as a red one. It was red for an hour.
The browser never sees a bot's card — not the deck, not a hand, not the face of
a card a bot drew, which is most of the deck. Seats cross the wire as a name and
a count.
The multiples are measured, not guessed: playing the first legal card you hold
wins 43/32/27% of the time against these bots, so the tiers price that to lose
about 8% a game and leave good play worth roughly the house's edge.
PeteFX.flyNode is the throw with the chip taken out of it, so a card can be
thrown across the felt the same way. fly() is now that with a chip in it.
Not yet driven in a browser, which in this room means not yet finished.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
The trivia ladder handled a walk before it looked at the clock, so the
timeout only ever bit if the browser volunteered it. Sit on a question,
look it up, answer if you find it and walk if you don't, and you never
lose a ladder. The clock is now the first thing that happens to a move.
The house's chip rack was wired up as bet buttons on blackjack and
hangman: it's four spans with data-chip on them and nothing said the
handler only wanted the real ones. Clicking the house's money raised
your bet.
Hangman had two definitions of "a letter you'd guess" — unicode in the
engine, ASCII in the renderer — and a phrase with an accent in it would
have had no tile to fill and no key to fill it with. One definition now.
Plus: trivia's countdown no longer freezes at zero when the server turns
down a timeout report it was early for, questions whose wrong answer
decodes into the right one are dropped at the door, and hangman bets on
PeteFX's spot like every other table instead of its own copy of it.
Trivia had every Go test passing and had never been in a browser, which
this plan's own rule says means nothing. So: play it.
The game itself holds up. The clock drains honestly and does not restart
on a reload, the multiple compounds, walking pays exactly what the felt
quoted, the reveal marks the right answer, and the auto-submit at zero
lands as a timeout rather than an illegal move. The next question's
answer never crosses the wire.
Two bugs only the browser could show:
- The spot printed double the stake after every settled game. standing()
set spot.amount and *then* poured the chips on, and pour grows the pile
from what it is told is already there. The money was always right; the
number under the chips was not, which is the one rule the felt is built
on.
- The house rack sat on top of the multiplier at 390px. Its 5.75rem inset
is not a margin, it is the width of blackjack's shoe — so on a phone the
rack sits in the middle of the felt. It now shrinks on small screens and
pulls into the corner where the corner is empty; data-at says which rack
is which, because pulling blackjack's to the edge slides it under the
deck.
The dev rig seeds its own question bank now (one real OpenTDB batch per
difficulty), because a fresh dev database 503s every start otherwise.
Solitaire, Vegas rules — the only shape solitaire has ever had as a
gambling game. You don't win or lose the deal: the stake buys the deck
outright, and every card you get home to a foundation pays a fifty-second
of the tier's multiple back. Cash the board whenever you like and keep
what you've banked, so a board that has gone dead is a decision rather
than a wall. No undo: the stake is spent the moment the deck is bought,
and an undo would be a way to walk a losing board backwards until it wins.
Three deals, and the two dials are the whole difficulty of Klondike.
Patient draws one with unlimited passes and pays 1.4x, so it takes 38
cards home to get square. Vegas draws three, three times round, 2.2x,
square at 24. Cutthroat draws three and gives you one pass, 3.4x, square
at 16 — most of those boards never clear, and you're ahead long before
they would.
internal/games/klondike is the same pure reducer as the other two, and
Pays() is one function for the same reason hangman's is. Two fuzzers hold
the deck together: no sequence of moves can lose or duplicate a card, and
the board stays well-formed. They earned their keep immediately — the
first thing they caught was a recycle that reversed the waste. It flips as
a block, so the card drawn first comes out first, and reversing it would
have dealt a different game on every pass and quietly broken the seed in
the audit log.
The browser never sees the stock or a face-down card, which here is most
of the deck rather than blackjack's one hole card: a column sends how many
cards are under it, never which.
The table re-renders and animates the difference. Blackjack plays back a
script because a hand only ever grows at one end; solitaire moves runs
from anywhere to anywhere and an auto-finish moves eleven cards at once,
so a script of "append this card there" would be a second engine over here
and it would be the one that's wrong. Instead the board on screen is
always exactly the board the server says exists, and each card is played
from where it just was to where it now is. The events supply only what a
diff can't: where a newly-revealed card came from, and what the board is
worth.
The rules are mirrored in JS on purpose, and only to light up the columns
a held card can go to. Being shown where a card goes is the game teaching
you; being told no after you commit is the game scolding you. The server
still decides, and a disagreement snaps the board back to what it says.
Two things came out into the open rather than being copied, which is the
rule this room runs on: casino-cards.js (the deck — faces, pips, the flip)
and PeteFX.spot() (the pile of chips and the number under it, which now
owns the rule that the number is a readout of the pile). Blackjack uses
both.
Not yet driven in a browser.
Hangman, and it plays for chips — which the plan had down as a free game, on
the grounds that trivia has no euro coupling in gogobee. But a free game in a
casino reads as a demo, so it stakes like everything else.
The idea that makes it a casino game rather than hangman with a wager stapled
on: the gallows is the payout meter. A wrong guess draws a limb *and* takes a
tenth off what a win is worth, because those are the same event and showing
them as one is the entire reason to bet on this. Short phrases pay 2.6x (fewer
letters, less to go on), long ones 1.6x — the floor is 1x, so a win never hands
back less than the stake, and the rake still comes out of winnings only.
State.Pays() is the number the felt quotes and the number settle() lands on.
They were briefly two sums, and the table spent an afternoon advertising a
pre-rake payout it didn't honour.
Two things the storage layer had already decided for us, and one it hadn't:
game_live_hands is keyed on the player, so "one game at a time" holds across
games for free (a live hangman 409s a blackjack deal, stake intact). But
table() unmarshalled every live row as a blackjack hand, which does not fail on
a hangman row — it quietly yields an empty hand. It dispatches on the game now.
commit() is the settle path both games share, and casinoRoutes() the one route
list, since the dev rig wires its own mux and a second copy is a copy that stops
including the newest game.
Driven in a browser, win and loss: 200 at 2.34x paid 455 and the bar landed on
it; six wrong took the stake and no more; a reload mid-phrase brought back the
board, the limbs, the multiple, the spent keys and the chips on the spot. The
browser found the two bugs a Go test can't — a lives counter under the house
rack, and a word wrapping early because the rack's clearance was on the whole
column instead of the one row beside it.
The table dealt cards but settled money by editing a number. So the felt got
the two things it was missing: a bet spot in front of you, and the house's rack
beside the shoe. Every chip is now always travelling between one of those and
the other.
You build a bet by throwing chips onto the spot — the chip you clicked is the
chip that flies. The stake sits there through the hand. The house pays out of
its rack into the spot, and the pile is then swept back to your stack. A loss
goes to the rack and does not come back.
Two rules hold it together. The number under the pile is a readout of the pile,
never the other way round: the bet starts at nothing rather than at a default
nobody put down, and a settled hand leaves your stake back up on the spot,
because otherwise the panel prints "your bet: 300" over an empty circle. And
the chip bar does not move until the chips that justify it have landed — a
counter that pays you before the dealer turns over is a counter that has told
you the ending.
casino-fx.js is the engine underneath: chips fly on an arc, out of a fixed
overlay so no container clips one crossing from a button to the felt. It knows
nothing about blackjack.
Also: cards land with weight and a degree or two of tilt, so a hand looks dealt
rather than typeset; the dealer takes a beat before drawing out; and a natural
gets confetti, which is the only thing in the room that does.
Driven in a real browser, which is the only way to review an animation — and
which is what caught the verdict pill rendering white on white in a dark room,
a chip rack sitting on top of the dealer, and Hit being offered over a table
that was still being paid out. devcasino_test.go is that harness, kept.
The last attempt built a card face out of text: a "♠" in a span for every
pip. At the size a card actually is, a suit character renders as a speck —
the shape is whatever font answered, it doesn't scale, and it can't be put
on the half-row a real pip layout needs. The result read worse than the
plain rank it replaced.
So each face is one SVG on a 100×140 field, suits as vector shapes, pips at
the coordinates a printed deck puts them. Courts get a framed panel with the
suit above the letter and again below it upside down — mirroring a letter,
which is what the first pass did, just stacks two of them into a blob; a real
court mirrors a figure.
Also restores .pete-card-back, which went out with the text rules it was
sitting among: without it a face-down card had no back at all, so the
dealer's hole card was invisible on the felt. Caught by driving a hand.
The tables were living in the news app's shell: Pete's face in the header
and the footer, the channel nav, search, the reader, the weather canvas,
the PWA. A casino is not a news page with a felt on it.
So it gets its own layout. What carries over is the design language — the
four palette vars, Fredoka/Nunito, the fat rounded cards, the dropped
shadow. What doesn't is every control it has no use for. gamesPage stops
embedding the news pageData, which is what keeps the furniture from
drifting back one convenient field at a time.
It keeps a clock, but tells a different joke with it: Casinopolis by day,
Casino Night Zone from six, palette and felt and the sign over the door all
changing together. The rule lives in roomAt() for the first paint and again
in the browser, so a player abroad gets their own evening.
And the cards are cards now — corner indices in both corners, the bottom
one upside down as printed, pips on the three-by-seven grid every real deck
has used for four hundred years, courts as a letter with the suit over each
shoulder. Driven in a real browser, both rooms, dealt through to a payout.
The engine, the escrow and the wire were all in place; nothing had a browser on
the end of it. This is that end: a lobby, a table, and the five endpoints between
them.
The browser holds no game. It sends intents and gets back a view — the cards it
is entitled to see, and the script of how they arrived, one event per card off
the shoe. The dealer's hole card is not in the payload at all until the reveal,
because a field the client is told to ignore is a field somebody reads in
devtools. The shoe lives in game_live_hands, which also means a redeploy
mid-hand no longer costs a player their stake: the hand is still there when they
come back.
The money is ordered so nothing can be spent twice. The stake leaves the stack in
the same statement that checks it exists, before a card is dealt. Every new hand
is seated with a plain INSERT, so a double-clicked Deal is decided by the primary
key rather than by a read that raced — it loses, gets its chips back, and the
hand in progress is untouched. A double takes its raise up front and hands it
straight back if the engine refuses the move.
Cards are dealt rather than swapped in — they fly out of the shoe and turn over,
which was a requirement and not a flourish. The faces and the chips are still
plain; that's next.
The baked atlas sprites (blurred-circle clouds, gradient moon with flat
maria blobs) read as cheap. Clouds are now built per-instance in the
particle fragment shader from fbm density with a flattened base and lit
tops; the moon renders in the sky shader with sphere shading, emboss
relief, maria and a halo. Stars skip the moon disc since the sky pass
draws first.
The background weather is now GPU-rendered: one instanced-quad draw call
over a baked sprite atlas plus a fullscreen sky shader (fog, Saharan
haze, aurora, sun rays, storm gloom and lightning flash). The old
Canvas2D renderer stays as weather-2d.js and kicks in automatically
when WebGL2 is missing; weather.js is now a thin controller that owns
the toggle, prefs and the PeteWeather API.
Effect upgrades: shared wind with gust pulses leans the whole scene
together, rain gets depth, splash pops and velocity-aligned streaks,
storms grow procedural branched lightning bolts, snow mixes soft motes
with spinning six-arm crystals, clouds drift in two parallax layers,
clear nights get a moon with maria, twinkling stars and the occasional
shooting star, blossoms and leaves tumble with a faked 3D flip.
New variants: haze (Saharan calima), wind (autumn gusts with streak
lines), hail (bouncing stones with drizzle) and aurora. The /weather
demo page switches variant, intensity and phase in place without a
reload and shows the active renderer plus an FPS meter.
The reader overlay opened without moving focus, so Arrow-Up/Down and
PageUp/Down scrolled the page behind it until the user clicked into the
text. Make the scroll container focusable (tabindex=-1) and focus it on
open, and drop its focus outline.
Reader read-aloud now streams neural WAV audio from a new POST /api/tts
endpoint that shells out to Piper, instead of the browser's Web Speech
voice. Each paragraph is synthesized on demand with the next one
prefetched during playback, keeping the existing highlight/scroll sync.
Voices are configured under [web.tts] (piper binary + voices_dir + a
labelled voice list) and exposed to the client as window.PETE_TTS; the
reader gets a Voice selector in the Aa menu, persisted per-device. Still
a signed-in-only perk and gated on auth.
- reader.js: guard read-aloud against the loading placeholder (bodyReady)
and invalidate stale speechSynthesis callbacks with a generation token
- storage: check rows.Err() after iterating story-view/content-length reads
- metrics: reuse placeholders() instead of duplicating the IN-clause builder
Surface read counts and sharpen the reader:
- story_views table + RecordStoryView on /api/article (background,
filter-guarded); "Popular this week" home rail via TrendingStories;
read-count badge and reading-time chip decorated onto every listing
- reader: signed-in-only read-aloud (TTS), native share/copy, and an
Aa typography popover (size/serif/sepia) persisted per device
- real alt text on card/reader/related/search images; time-of-day Pete
greeting on the home hero
- harden exec() to skip (not panic) on a nil DB so background writes
can't crash on a closed handle
Tests: story_views_test.go, trending_test.go. Suite green, CSS rebuilt.
- push digest queries now exclude _duplicate channel like every other
visibility query (bookmarks list/count and NewClassifiedSince)
- advance push watermark to newest scanned story, not pass-start now, so
stories arriving during a long send pass aren't re-counted next pass
- replace hand-rolled escapeHTMLText with stdlib html.EscapeString
- drop em-dashes from user-facing copy; bump PWA CACHE_VERSION so clients
pick up the changed shell assets
Code review of the personalization/feeds/PWA/push work surfaced ten
confirmed issues, now fixed:
- Web Push delivery bypassed the SSRF guard (unguarded default client);
now routes through safehttp.NewClient with a hard timeout, and the
subscribe handler validates the endpoint URL.
- Push unsubscribe deleted by endpoint with no owner check; added
RemovePushSubscriptionForUser scoped to the signed-in user.
- Byte-slice body/content truncation could split a UTF-8 rune and break
the RSS content:encoded XML; added a rune-safe truncateUTF8 helper.
- Digest sender could permanently starve a user who hid a high-volume
source; step the watermark past a full hidden-source scan window.
- Service worker cached personalized HTML navigations into a shared
cache (identity leak across PWA users); navigations are now
network-only, CACHE_VERSION bumped to v2 to purge stale pages.
- Public /api/article leaked discarded/unclassified bodies; filter to
classified, non-sentinel stories.
- runLocal never started the push sender; digests now fire in -local.
- Push client had no timeout, so one hung endpoint stalled all sends.
- Reader migration resurrected cross-device-cleared reads; gate it
behind a one-time flag so the server stays authoritative.
- Bookmarks count didn't match the classified list filter.
A multi-session build turning Pete's read-only web UI into something people
return to. Five phases, signed-in features keyed off the OIDC subject; anonymous
visitors keep the reverse-chron feed and localStorage-only state.
Phase 1 — per-user read + bookmark state: user_story_state table +
storage/userstate.go; auth-gated /api/read, /api/bookmark, /api/state and a
/bookmarks page; reader.js syncs state server-side for signed-in users. Also
hides the Matrix-posting UI when posting.enabled=false (web-only mode).
Phase 2 — outbound feeds: storage.ListForFeed + web/feed.go hand-build RSS 2.0
(content:encoded) and JSON Feed 1.1 (no new dep); /feed.xml, /feed.json and
per-channel variants; <link rel=alternate> discovery tags.
Phase 3 — "For you" + related: storage/rank.go scores recent unread candidates
by channel/source affinity + recency decay; RelatedStories via FTS5. ForYou rail
+ /for-you page; public /api/related feeds the reader's "You might also like".
Phase 4 — source-health dashboard: source_health table + storage/sourcehealth.go
(RecordPollResult, ListSourceHealth, SourceContentStats), written by the poller;
admin-gated /status page behind web.admin_subs.
Phase 5 — PWA + offline reader + Web Push: root-scoped manifest.webmanifest and
sw.js (app-shell precache, /api/article runtime cache for offline reading,
offline fallback, push/notificationclick handlers); PNG icons from pete.avif;
pwa.js registers the SW and drives a notifications toggle. Web Push adds
webpush-go, a [web.push] config block (pete -genvapid mints VAPID keys), a
push_subscriptions table, auth-gated subscribe/unsubscribe endpoints, and a
digest sender that pings each subscriber "N new stories" past their watermark,
honoring disabled-sources and pruning gone endpoints.
Tests added beside each new storage/web file; go test ./... and go vet clean.
Reader mode presents the stories on a page one at a time in a focused
overlay, marking each read as it's shown. Left/right arrows (or the header
book button / `f`) page through them; read stories dim on the grid. Read
state is device-local in localStorage.
Backing this required actually capturing article bodies, which Pete wasn't
doing — it kept only the RSS <description> lede and discarded content:encoded:
- stories.content column (idempotent migration; old rows fall back to lede)
- parser keeps content:encoded as paragraph-preserving text
- article fetch already done for paywall detection now also returns its body,
so ingest stores the richer of feed-content vs scraped body with no extra
request (prefers the archive snapshot body for paywalled stories)
- GET /api/article?id= serves the stored text; card queries now select id and
expose it as data-id for the reader
Tests cover content extraction, the storage round-trip, and the article
endpoint + card rendering end to end.
Reuses the saved weather location's lat/lon to fetch us_aqi from
Open-Meteo's air-quality API (no key), folded into the existing 2h
forecast cache (bumped v1->v2). AQI is best-effort: a failed fetch
never sinks the forecast. Shows as a header chip and a colored row
on the forecast card; both self-hide when there is no reading.
Signed-in users get their preferences (hidden feeds, weather location,
weather toggle) stored server-side keyed by their OIDC subject and synced
across devices. Anonymous visitors keep using browser localStorage, so the
site stays public. First sign-in migrates existing localStorage prefs up.
- config: [web.auth] section (issuer, client_id/secret, redirect, session_secret)
- storage: user_preferences table + Get/PutUserPrefs
- web/auth: OIDC code flow, HMAC-signed session cookie, CSRF state + nonce
- web/prefs_api: GET/PUT /api/preferences (auth-gated, 64KB cap)
- frontend: prefs.js sync layer seeds localStorage from server, pushes on write
- header: sign-in / account control
OIDC discovery is non-fatal at boot: if Authentik is down, Pete serves
anonymously rather than refusing to start.
Visitors can save a postal code (international, via Zippopotam) to get the
local forecast from Open-Meteo — a header chip + a 5-day home-page card —
and the canvas background switches from the seasonal effect to live
conditions. Entirely client-side: no keys, no server logic. Geocode cached
permanently, forecast cached 2h. Celsius by default, Fahrenheit opt-in.
New canvas effects: clear (sun by day, shaded moon + stars at night),
clouds (blurred drifting sprites), snow, fog, and storm (rain + lightning).
Seasonal effects remain the no-location fallback.
btoa() throws InvalidCharacterError on non-Latin1 input, so feed
names with em-dashes ("The Guardian — World") killed render()
before the dialog's hidden class was removed and nothing visibly
happened on click. Use the loop index for the row id instead.
Visitors can hide individual feeds via a gear-icon panel in the header.
Preferences live in localStorage; the server ships the full source list
(name + channel) as window.PETE_SOURCES so the panel lists every feed,
grouped by channel, regardless of what's on the current page.
- Add internal/safehttp: hardened HTTP client (DNS-resolved dial guard
blocking loopback/RFC1918/CGNAT/link-local, redirect re-validation,
body-size cap) and rewire article/feed/thumb clients through it
- Cap goquery body at 5 MiB so a hostile origin can't OOM the process
- search.js: reject non-http(s) hrefs to block stored XSS via javascript:
- dedup: tracking-param key "CMP" was unreachable (lookup lowercases);
fixed to "cmp" so CMP= is actually stripped from canonical URLs
- ForcePost: postItem now returns bool; on dedup-skip ForcePost returns
false so !post falls back to DB lookup instead of silently consuming
- Bound reaction callbacks behind an 8-slot semaphore; drop overflow
- Add stories indexes on (channel, classified, seen_at DESC),
(classified, seen_at DESC), and partial image_url to kill full scans
in IsKnownImageURL and ORDER BY seen_at hot paths
- Surface FTS5 probe Scan error instead of swallowing it
Search: new FTS5-backed SearchStories query, /search JSON endpoint,
client-side overlay (search.js) wired into the layout header. EU
channel gets its own theme color (#003399) across bg/text/border/glow
classes. Sources can now route to non-Matrix channels without
validation error (web-only mode); a warning still flags typos.
Server picks variant from Lisbon-local date (rain/petals/jacaranda/motes/leaves)
and renders behind content via a canvas particle layer. Each variant has a
hand-drawn silhouette so shapes are recognizable. /weather demo route exposes
variant + intensity + phase pickers, locking the time-of-day phase override.
