Rust + WASM typing game. https://srdusr.com/git/typerpunk/atom?h=main 2026-02-27T23:17:00+00:00 2026-02-27T23:17:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2026-02-27T23:17:00+00:00 urn:sha1:5af72f55bb7fb620af740c9da394e093d7f1280a Merchandise. Shirts, mugs and deskmats, sold through the same hosted checkout. Physical goods differ from cosmetics in three ways the schema had to carry: they have a size to choose, they need an address, and payment unlocks nothing. What a paid order produces is an obligation to pack and post something. The address is collected by Stripe on its own page and arrives here on the webhook, so no postal detail is ever entered on this site. Postage is a separate line item rather than folded into the price. A size is checked against the sizes the item actually comes in, on the server as well as in the browser, so a request naming anything else is refused rather than quietly posted as a medium. /api/admin/orders lists what has been paid for and not yet sent, and the Contribute screen shows an administrator the same list with the address and a button to mark each one posted. Without that the list of what is owed lives only in the processor's dashboard. Shipping is limited to 21 countries, which is a list of places somebody is willing to post to, not a technical limit. The race view. The sprite sat still at the left of a 6px bar, which made the one thing a player owns and can see the least visible part of the race. It rides the track now, moving with progress, on a dashed line that reads as road ahead with the trail behind it. The field is 1040px wide rather than 680, the sprites are 30px rather than 18, and each row carries its own percentage as well as its speed. Advertising. Two 160x600 rails in the page margins, shown only above 1424px, which is the width at which they fit beside the widest content column without crowding it. Below that they do not exist. They follow the same rules as the banner: never while typing, never for a supporter. --border-color was used in five places and defined in none. An undefined custom property invalidates the whole declaration at computed-value time, so every one of those borders fell back to currentColor: bundle cards, merch cards, the size buttons and the settings checkbox all had borders that were either invisible or faint text-coloured lines. It is defined now, derived from each theme's neutral so it tracks the palette. .settings-hint caps itself at 340px, which is right under a heading and wrong for a line introducing a full-width grid. Verified against a real database: the catalogue endpoint, a shirt refused without a size, a shirt refused with a size it does not come in, a valid size and a no-variant mug both reaching checkout, 404 on an unknown item, 401 unauthenticated, and 401 on the admin orders list without the role. 24 Rust tests, 3 browser tests, and all 21 screen and viewport combinations clean. 2026-01-14T20:23:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2026-01-14T20:23:00+00:00 urn:sha1:4933f669f8338ca5e4c64cf0294818c5d15ed122 The dataset's median passage was 69 characters, about twelve words, which is over in twelve seconds at an ordinary speed. Some packs were worse: shell had a median of 46 and a shortest entry of 17. That is the reason packs felt small. It was not the number of entries, which was 15 to 37 per pack, but the length of each one. Multiplayer had already hit this and worked around it: load_race_texts filters to 120 characters or more. That filter left only 61 of 350 passages eligible, 38 of them from two packs, so races repeated constantly and four packs could never come up at all. Both halves are fixed. Content: 104 longer passages added across every pack. The command packs get whole pipelines rather than single flags, which is how the tools are actually used and what the explanations were always for. Prose packs get passages that run 150 to 320 characters. Selection: a floor of 120 characters. Prose packs draw from their long entries where a pack holds at least five, so no pack is reduced to the same few passages. The command packs chain consecutive entries into one drill, which is the natural shape for them, and the explanations are collected so each line is still described. Measured over 400 draws per pack, every category now runs a median of 146 to 218 characters with a shortest draw of 120. The race pool went from 61 of 350 to 130 of 454. Attribution: 24 passages had none and displayed nothing at all under the text. They now say Unknown, which is the honest answer for a fact written for the pack, and the pangram is credited as one. The results and typing screens both fall back to Unknown rather than rendering an empty line, and the pack name is shown beside the source. Multiplayer results, three fixes: - PB never appeared. A race has no mode of its own, so modeKey was undefined, so recordResult never ran. Races share one key, because the passage is whatever the server dealt and a per-passage best would never be beaten. - Play Again started a solo test. A race carries standings and a solo run does not, so the results screen can tell them apart and now queues for another race. The end screen's own cleanup leaves the old room first. - CONSISTENCY was the longest label on the screen and made the accuracy column wider than the WPM column opposite it. It reads CON. RaceText carries the pack name so the results can show it for a race the same way single player does. 24 Rust tests and 3 browser tests pass. 2026-01-13T19:14:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2026-01-13T19:14:00+00:00 urn:sha1:695bf0df2ec48e67d716dad25711cdc2e96bf7f4 The catalogue served every price as $0.00. The Cosmetic struct typed price_cents as i64 against an INTEGER column, so sqlx refused the decode and unwrap_or_default turned the failure into a zero. Nothing was logged. The whole store read as free while the database held the real prices. Decode errors are now returned rather than defaulted away, and create_session refuses any amount at or below zero, so a price that fails to decode cannot become a session that grants an item without charging. Layout, all measured in a real browser at 1280, 820 and 390 pixels: - The close button sat underneath the fixed top rail. The rail reaches into the content column on any viewport under about 1350px, so the overlap was there for nearly every visitor. The button starts below the rail now. - The wordmark and the top rail overlapped by 75px at 390px. Both are fixed to the top of the viewport and neither knew about the other. Both give ground on narrow screens, and Sign Up drops out of the rail because Sign In reaches the same screen. - Store rows were 93px tall for one line of content, because .menu-button carries a vertical margin meant for a stacked menu. Twenty six items came to 2400px of scrolling. Rows are 47px and the page is 2467px rather than 3784px. - .stats-screen centres its children, so any child without a declared width shrink-wraps. That left the sign-in box at 415px, the bundle grid at 488px (which collapsed it to a single column) and the leaderboard table at 498px, all inside a 636px column. Every screen was swept for the same defect. .account-panel is the one narrow child that is deliberate: it declares max-width 360px because a sign-in form should not be 636px wide. - Leaderboard rows had a hard 461px minimum from fixed column widths, so at 390px the row ran from x=-36 to x=426 and the date column was cut off the side of the screen. The date is hidden on narrow viewports. - The profile links in the leaderboard were 21px tall, under the 24x24 minimum target size, and they are the only route to a player's profile. Sprites never showed as equipped: the store compared the equipped caret and flair but not the sprite. The store test bought items by clicking Buy, which used to grant them for nothing. It now asserts that Buy does not grant, then grants the items the way a signature-verified webhook would, and goes on to check that an equipped caret colour reaches the typing screen. All 21 screen and viewport combinations are clean for overlap, overflow, clipped content and target size. 3/3 browser tests pass. 2026-01-12T19:19:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2026-01-12T19:19:00+00:00 urn:sha1:a19c03bc6394ab08cc5a1a9fbabf5eb46b6791f7 The store had 26 items, a price on each and a working equip flow, but the purchase endpoint granted ownership without taking any money. Anyone signed in could take the whole catalogue for nothing. That endpoint is now gone. Payment goes through Stripe Checkout, which is hosted by Stripe. The buyer is redirected there and comes back, so no card details reach this server and it stays outside PCI scope. Three rules hold the money path together: - The price comes from the server's own catalogue row. The client sends an item id and never an amount. - Nothing is granted at checkout. The item appears only when a webhook arrives with a valid HMAC-SHA256 signature, checked in constant time against a 5 minute timestamp window. - Fulfilment keys off the processor's session id, which is UNIQUE, so a webhook delivered twice cannot grant the same item twice. Prices now vary by item. Every caret cost the same as every other because they are the same thing in a different colour, which left nothing to save for. Carets run 149 to 349, flair 129 to 299, and sprites 249 to 399, since a sprite is the one cosmetic every other racer sees. Four bundles sit above the catalogue, each priced below the sum of its parts: Starter Kit, Neon Set, Racer Set and The Lot. The saving is computed from the current item prices rather than asserted, so it cannot drift. The Lot is defined as every cosmetic rather than a fixed list, so it stays complete as items are added. Three bugs found while wiring this up: - Sprites never showed as equipped. The store compared the equipped caret and flair but not the sprite. - Supporter status was a stored boolean that was set on payment and never cleared, so a 30 day subscription lasted forever. Both read paths now derive it from the expiry. - sqlx::migrate! reads the migrations directory at compile time, but cargo watches source files only. Adding a migration did not trigger a rebuild, so the binary shipped the old migration set and the schema change never ran. A build.rs now declares the dependency. Verified against a real database: bundle maths, the grant statement and its replay, 401 unauthenticated, 404 on unknown ids, 501 with no Stripe keys, and both already-owned refusals. 2025-12-24T22:24:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2025-12-24T22:24:00+00:00 urn:sha1:b97beebfd527a12b182a12a556381768a728d9dc Design tokens - The stylesheet had five corner radii in use with no rule for which applied where, so a button was square while the input beside it was rounded and the icon button next to that was something else again. Two tokens now: --radius for controls, --radius-panel for the surfaces they sit on. - One border weight. Buttons were 2px and inputs 1px, which made a field and its own button read as different weights of the same idea. The two start buttons keep 2px, where it is doing work. - Letter-spacing was 1px, 2px, 3px and 4px with no rule behind it. Now 1px for small caps labels and 2px for button text. - Removed the .ghost modifier. After the hierarchy pass it rendered identically to .quiet, so the two were one control under two names. Its uses were toggles, which .quiet plus .active already expresses. Racer sprites - Six sprites in the same angular language as the rest of the icon set, one per racer in the lobby and the race. A race reads as characters moving rather than as coloured lines with names attached. Assigned from the server's player order, so every client draws the same person as the same character without needing to agree on anything. Store - Six items to twenty-six, across three slots. Race sprites join carets and flair as a third slot, with the equipped sprite stored per user. Nothing is required to race: an unequipped player keeps the sprite their position in the room assigns. Advertising space - Reserved, empty, and loading nothing. The slot exists so that adding a network later cannot push the page around when the tag loads. Shown to signed-out visitors and to accounts without the supporter flag, on the end screen only: interrupting somebody mid-test is the one placement that would cost more than it earns. - No ad code is included, and none can be added without three deliberate changes recorded in adSlot.js: the CSP forbids third-party script, frame and image sources; the privacy page states there is no advertising and no third-party script; and an ad network brings consent requirements that this app has no mechanism for. Verified after this change that the only external request the page makes is still fonts.googleapis.com. 2025-12-18T17:21:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2025-12-18T17:21:00+00:00 urn:sha1:b371f21989bfca13fe8cc9bd59e95f5614f63b0d End screen - There was an auto-fit routine that forced this screen into one viewport: it shrank the graph to a 120px floor, trimmed the Play Again margin, then capped the passage box at 80px with its own scrollbar. On a 650px window that left the passage 80px tall and clipped, and the graph 120px, which are the two things the screen exists to show. Removed. The page scrolls instead, which is the right trade for a screen that is read rather than acted on under time pressure. - The screen is a flex column, so its children also shrank by default once the content was taller than the viewport. The passage, graph, stat row and standings no longer shrink, and the chart has a floor below which it stops carrying information. Bottom chrome - The keyboard hint and the footer links were both fixed at bottom centre and overlapped at every window size. The hint now sits above the footer. - Normal-flow content could end up underneath the fixed footer and the corner rails. One --bottom-chrome variable reserves that space on every screen. - On a narrow screen the footer grows to the full width once its links wrap, so at 375px it ran through both corner rails and covered Play Again, which made the button unclickable. The chrome stacks there instead: rails on the bottom line, footer above them, hint above that. Mode picker - It ran to the last pixel of the window at every size. It now keeps clear of the bottom edge, and opens upward when a short window leaves more room above than below. Responsive - Checked at nine viewports from 1920x1080 down to 375x667: no horizontal overflow and nothing off-screen on the menu, the typing screen or the end screen. Configuration - dotenvy searches upward from the working directory, so crates/server/.env was only found when starting the server from that directory. The repository root is tried as well, which is where it is usually started from. - .env.example and the README explain where secrets belong: the environment, a gitignored .env for local work, and EnvironmentFile or a platform secret store in production. Also what to do if one is exposed. - TYPERPUNK_ADMIN_USERNAME and TYPERPUNK_ENV are documented rather than left to be discovered in the source. 2025-12-15T18:44:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2025-12-15T18:44:00+00:00 urn:sha1:1740327557074df0c8b99635ee949e2540ac94d0 Dependencies - The server build carried 37 known advisories, including RUSTSEC-2024-0363 in sqlx 0.7, which is the database layer. sqlx moved to 0.8 with default-features off, which also drops the MySQL and SQLite drivers and with them rsa and RUSTSEC-2023-0071. reqwest moved to 0.12, which brings hyper 1.x and was the sole source of every remaining advisory: h2 0.3, rustls-webpki 0.101, rustls-pemfile 1.0 and idna 0.3. - The server build now reports no known vulnerabilities against OSV. cargo audit itself would not compile, so the check queries OSV with the crate versions cargo tree reports for the server binary. - Cargo.lock is committed. This workspace produces binaries, so the lockfile is what makes a deployed build reproducible and the audit above meaningful. Headers - The application sent no security headers at all. The static server now sends a Content-Security-Policy, nosniff, frame options, a referrer policy and a permissions policy; the API sends a policy of its own, since it serves JSON and should load and frame nothing. - The one inline script in index.html moved to a file so script-src needs no unsafe-inline. WebAssembly needs wasm-unsafe-eval, without which nothing types at all, so that is present and explained. - Five style attributes moved to the CSSOM rather than adding unsafe-inline for styles. A style attribute in markup is refused by the policy; the same property set through element.style is not. Production configuration - With TYPERPUNK_ENV=production the server refuses to start if COOKIE_SECURE is off, if DATABASE_URL is still the development default, or if FRONTEND_ORIGIN is http on a non-local host. These were warnings, and a warning in a log nobody reads is not a safeguard. Administration - Moderators were appointed with psql. There is now an admin role, bootstrapped from TYPERPUNK_ADMIN_USERNAME at startup, and a UI to appoint and remove moderators. An administrator's own role cannot be changed through the API, so a mistake cannot lock everyone out of moderation. Corpus - scripts/export_approved.js writes approved submissions back into data/packs/community-*.json. Approved passages are served from the database and merged at startup, so without this the repository dataset and the live corpus drift apart, and a fresh checkout or the TUI sees only what shipped. Documentation - README rewritten for the repository: what it does, how to run it, the pack format, the server variables, deployment, and what the security posture actually is. Plain English, no em dashes, no emoji. Checked and found already correct: every private endpoint refuses anonymous callers, session cookies are HttpOnly and SameSite=Lax, CORS names a single origin, internal errors are logged rather than returned, and every query is parameterised. 2025-12-14T07:24:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2025-12-14T07:24:00+00:00 urn:sha1:3dbebfbc9345d2603908f32c0dabebc0ff21feb3 Submissions - POST /api/texts proposes a passage; nothing reaches players until a moderator approves it. GET /api/texts serves the approved set, which the client merges on top of its bundled packs at startup. - Validation the server enforces rather than trusts: category from a fixed list, 40 to 600 characters, no control characters (a newline makes a passage untypeable in a single-line input), attribution length, and a unique index on md5(lower(btrim(content))) so the same passage cannot be submitted twice under different whitespace or casing. - Moderation is a flag on users. The queue and the review endpoint both refuse a non-moderator, and reviewing an already-reviewed submission is a 404 rather than a silent second write. - Submissions are rate limited per user: enough for a real contributor, not enough to fill the queue from a script. - A Contribute screen carries the form, your own submissions with their status, and - for moderators only - the review queue. This is the half of TypeRacer's model the packs could not reach by authoring: their corpus is roughly twelve thousand passages, grown by submission. Custom text as a study tool - Imported documents are kept between visits, with how far through each one you are. Custom text lived only in memory, so importing a set of notes and reloading the page lost them - fine for pasting a paragraph to race, useless for working through a file over several sittings. - Position is recorded when a segment is finished, not when the next is started, so closing the tab after a segment does not lose it. - Markdown is chunked as markdown: fenced code blocks are kept whole and typed line by line, and the decoration - hashes, asterisks, backticks, link brackets, table pipes - is stripped so what you retype is the material rather than the punctuation around it. - Everything stays on the device. Notes are not uploaded anywhere. Fixed while doing it: a chunk could contain a newline, which cannot be typed in a single-line input at all. Any paragraph with a line break inside it -- ordinary in notes and in wrapped prose - produced an unfinishable segment. Whitespace inside a chunk is now flattened. 2025-12-07T19:58:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2025-12-07T19:58:00+00:00 urn:sha1:8fc4440150ce84c9cc3f5ec13d50beb58a5b65bc The packs were not proper. An audit found 86 of 253 items (33%) carrying an attribution that just restated the category - prose *about* a topic with an invented source, which is the same fault the movies pack had. Six of thirteen packs were mostly that: technology had 15 items and one distinct attribution. - science, technology, history, nature and business are now sourced quotes with real attributions: Feynman, Curie, Hopper, Dijkstra, Lincoln, Carson, Drucker, Goodhart. 82 items, all attributed to a person or a work. - general is original factual prose, so it now carries no attribution at all rather than claiming "General knowledge" as a source. merge_packs no longer invents one from the pack's filename. - Four explanatory passages in philosophy lost their "Philosophy" attribution for the same reason. - One duplicated passage removed. Generic attributions: 86/253 before, 0/349 now. New technical packs - shell: 24 awk, sed and pipeline drills, each explaining what the line does - field splitting, associative arrays, !seen[$0]++, process substitution, xargs -0, strict mode. - sysadmin: 24 operational one-liners across systemd, disk, processes, network, permissions, SSH, backup and containers. - programming grew to 37 and hacking to 30, with git bisect, window functions, EXPLAIN ANALYZE, certificate transparency and capability audits. - All 115 technical drills carry an explanation. Race passage length - Multiplayer drew from the same pool as single player, so a race could land on a 22-character quote and be over before anyone had their hands in position. Races now require 120 characters; the filter is applied when the pool is loaded, not to the packs, since a short quote is fine to type alone. For reference, TypeRacer organises by difficulty and language rather than topic: one default English pool of ~11,900 texts plus per-language universes and specials (accuracy, repeat, easytexts, anime). Their scale comes from user submission with moderation, which is still the feature this does not have. 2025-12-07T18:38:00+00:00 srdusr 99972264+srdusr@users.noreply.github.com 2025-12-07T18:38:00+00:00 urn:sha1:5b1ea38522dbf6bf60db5a2270463de0c12d9de3 PostgreSQL - sqlx switched from the sqlite feature to postgres; the server now runs on Postgres 18 and the SQLite file is gone. - 95 placeholders renumbered from ? to $N. - REAL widened to DOUBLE PRECISION: Postgres REAL is float4 and will not decode into the f64 the code reads. - flagged and is_bot are real BOOLEANs rather than 0/1 integers, with the decode side reading bool. - The leaderboard's derived table gained the alias Postgres requires, its flag comparisons became boolean predicates, and INSERT OR IGNORE became ON CONFLICT DO NOTHING. - u32 binds cast to i64; Postgres has no unsigned integer types. - Integration tests run against a real database - Postgres has no in-memory mode - each in a throwaway schema, with search_path set per connection because it is session state and the pool opens more than one. - Timestamps stay TEXT for now and LISTEN/NOTIFY is still unused; both are recorded in TODO-postgres.md rather than left implied. Custom text and lyrics, checked rather than assumed - Custom files never reach the server: they are read in the browser through the File API, so there is no upload, no path handling and no remote file inclusion to have. Verified by driving a hostile file - markup in the body and in the filename - all the way onto the typing screen: it renders as literal characters, no nodes are created, nothing executes, and the filename is escaped in the attribution too. - That test found a real regression: picking Custom from the new mode picker selected it without ever starting it, so the mode was unstartable. - /api/lyrics fixes its upstream host, so it cannot be pointed elsewhere, but it was an unbounded relay: now rate limited per IP, with length caps on artist and track and a ceiling on the response body it will read. Hacking mode - 22 single-line drills across recon, web, memory safety, exploit development, crypto, post-exploitation and defence, each syntax highlighted and each explaining what the line actually does. All 19 modes verified to start, render and be typable.