# The Machine The design of the boot chain, and the backlog for it. `ROADMAP.md` is about servers and services. `README.md` is about how the code works. This file is about the fiction: what xanders.page is pretending to be, why that lets it be five different decades at once, and everything still to build. --- ## The problem this file solves The honest version of the question: I love Windows 95, 98, XP, Vista and 7, AND I love the Fallout terminal, and right now the boot chain is NTLDR from 2001, a Vista loading bar from 2006, a Windows 98 dialog, an XP Command Prompt and a green phosphor terminal from a game set in 2281. Do I have to pick one? **No. But they cannot all stand in the same room arguing.** The mishmash reads as a pile of references because every era is competing for the same screen. Give each one its own room and a reason to be there and the same pile reads as a machine with history. --- ## The frame: one machine, several partitions Real computers accumulate eras. Your firmware is older than your operating system. There is a recovery partition you have never booted. There is an old install you kept because you could not face moving the files. Nobody looks at that and calls it inconsistent, because it is what every long lived machine looks like. So xanders.page is **one beige box** that has been rebuilt several times. ``` FIRMWARE (always older than everything it boots) POST, the boot device menu, PXE over IPv4 | BOOT MENU | +-----------------+------------------+ | | | Pages Web Pages 98 Pages Server Core the front door the toy box the private wing Frutiger Aero Win95/98 no GUI, green phosphor 2006-2009 1995-1998 a serial console ``` ### Why each one has earned its place **The firmware is era neutral and that is the point.** POST screens looked much the same from 1995 to 2010. It is the one part that never changes, which makes it the frame the rest hangs off. Firmware being older than the OS it boots is correct, not sloppy. **Pages Web** is the public creative hub, the thing you show people. Aero glass, the wallpaper cycler, the buttons, the ads. It is already built. **Pages 98** is where the fun lives. A real desktop: draggable windows, a Start menu, icons. This is the room that does not exist yet and it is the biggest single thing on the backlog. **Pages Server Core** is the private wing, and this is the one that fixes the whole problem. Windows Server Core genuinely ships with no graphical shell. You get a console. A green phosphor terminal is not a random Fallout reference bolted onto a Windows site, it is **what you get when there is no GUI**. The RobCo styling is the flourish on top of a justification that already holds. That is the trick. The Fallout terminal stops being a fifth competing theme and becomes the correct answer to "what does this machine look like with the desktop switched off". ### What this buys - Every era you like gets a room instead of a fight. - The private wing feels like a different machine, because it is a different boot. - New ideas have an obvious home. Minesweeper goes in Pages 98. A new gallery goes in Pages Web. A new command goes in Server Core. - Nothing already built has to be thrown away. --- ## Reaching the private wing **Done, and then done again, because the first answer was half of one.** The first version moved the link out of the sidebar and into the boot menu, the way you would reach a recovery partition. That was better than a sidebar link and it was still not right, because the boot menu entry and the login screen were the same page: Pages Server Core, the green terminal that is one of the four partitions and is meant to be shown to people, opened by asking for a password. So the terminal is two rooms now, and the flag in the URL decides which: | | | |---|---| | `pages/private.html?booted` | **Pages Server Core.** Public. Drops straight into the guest shell. Linked from the boot menu, the sidebar and the Pages 98 Start menu. | | `pages/private.html?booted&login` | **The private wing.** The Authelia login, exactly as it was. Linked from nowhere. | You get to the second by booting **MS-DOS** behind F12 at the POST and typing `TELNET PAGES`. Which is period correct rather than a puzzle: NCSA Telnet for DOS was real, and connecting out of a DOS box to a machine that then asks who you are is precisely what you did in 1996. Everything that names the wing is behind the same flag. The public room does not print the Authelia banner, does not say the words "private wing", and `/etc/motd` and `~/README` inside the invented filesystem say which room they are in. `exit` in the public room closes the session and goes to the boot menu rather than dropping you at a login prompt, which is how the split would otherwise have been undone by the most obvious command in a shell. **Be clear with yourself about what that is and is not.** It is flavour. It is not security, and it must never be treated as security. The flag is in the source and anybody can read it; all it changes is how findable the door is. The only thing actually guarding anything is Authelia on the server, checking a password it alone knows, in front of services that are not on this site at all. Obscurity makes it feel like a secret. Authelia makes it be one. --- ## Phase 0: three things that are not features None of these are fun. All three are worth more than everything below them. - [x] **Put this in version control.** 444MB and months of work with no history and no undo. One bad `tools/pages.py` run and it is gone. - `img/` is 436MB of that, and images do not belong in git. Track the code, `.gitignore` the wallpapers and buttons, and back those up separately. Losing the code is a catastrophe; losing the wallpapers is an afternoon of re-downloading. - [ ] **Open the site on the 2010 and the 2012 MacBook.** STILL NOT DONE, and still the most valuable thing on this list. Everything here is written to Safari 7 spec and has only ever run in Chrome, which is theory rather than testing. - **`TESTING.md` is now the actual job**, written 21 August 2026. The reason this kept not happening is that "test it" is not a task, it is an obligation with no edges: you cannot tell when you are finished, so you do not start. That file is the same work with edges. Eight sections, about half an hour a machine, what to look at and what correct looks like, and a place to write down what you saw. - It names the two places where a Safari 7 viewport unit bug would actually show, `.tty` and `.osmenu`, both full screen takeovers, so a failure is loud rather than subtle. - It also says which differences are CORRECT rather than broken: the hardware probe legitimately prints fewer lines on Safari 7, because two of the APIs it reads arrived in Safari 10. A shorter list is right; a line saying `undefined` is not. - `tools/serve.py --lan` now binds to the network and prints the addresses to type into the MacBooks. Nothing to deploy, nothing to buy. Off by default, because binding to 0.0.0.0 lets anything on the same wifi read this folder. - If nothing loads it is almost always the Windows firewall. Allow Python on Private networks, or add an inbound rule for TCP 8080. - Two predicted bugs are FIXED already, without a Mac, because both were certain enough to act on: - `caret-color` is Safari 16. It was hiding the real text caret over the drawn cursor in `css/tty.css` and `css/shell.css`, so both Macs would have shown two cursors. The inputs are parked off screen now. - `localStorage` was unguarded in `js/main.js`, `js/wallpaper.js` and `js/gallerypage.js`. Safari 7 in a Private window throws on `setItem`, killing the rest of the file. Everything goes through `storeGet` and `storeSet` in `js/store.js` now. - Still unknown, and only a real machine can answer it: 11 viewport unit values in `css/tty.css`. Safari 7 has vh and vw but with known bugs inside `position: fixed`. Look at the terminal first. - [x] **The documentation is served, on purpose.** DECIDED, 21 August 2026. It suits this site: nothing here is a secret, the whole of `js/` already ships to every visitor, and a site whose subject is partly itself is a strange place to start hiding the workings. - The change is not that they became readable, they always were. It is that they are now FINDABLE. A file that is served but that nothing points at is neither published nor hidden, it is just sitting there. The site map has a "How it is built" section linking all five, and `sitemap.xml` lists them, so the list of what this site publishes is complete. - **THE RULE THAT FOLLOWS: nothing goes in a markdown file at this root that would not go on a page.** Not a note to self, not a password, not a path into anything. That rule has already been broken once, by a paragraph naming the remote access on this machine while explaining that it was being kept off the site. - And do not leave scratch files at the root. One got left there during the audit that produced this item, and it was served too. - [x] **Decide how the site gets from G: to xanders.page.** It is built and there is no deploy path. Worth settling now because it shapes the private wing: if the VPS serves the public site too, Caddy handles both, and the real client IP endpoint in `js/sysinfo.js` comes free. - SETTLED on 2026-08-25, by building it. Caddy at home serves the repo directly, the VPS stream proxies 443 over WireGuard, and `https://xanders.page` returns the real site. The thing actually in the way was never the deploy path: it was a Windows Firewall with no inbound rule on the WireGuard interface, so the tunnel carried packets that died one hop short of Caddy. Also worth doing at some point, in the boring-but-real category: - [x] **Resize the wallpapers.** RUN, on 2026-08-20. `tools/shrinkwall.py` turned 350 images totalling 389MB into 76MB at 1600px wide, in `img/wallpaper-small`. The originals in `img/wallpaper` are untouched. - The measurement in the script's own docstring was WRONG by half: it said "roughly 560KB each" and every file is actually 1920x1200 at a mean of 1124KB, so the saving was double what it advertised. - THE SWAP TURNED OUT TO BE UNNECESSARY, and that is the better answer. Nothing had to move: the site now SERVES `img/wallpaper-small` because six path strings in `js/wallpaper.js`, `js/gallerypage.js`, `js/featured.js` and `js/desktop.js` were changed to point at it. The originals stay exactly where they are, `MANIFEST` is still valid because the filenames are identical, and the front page went from about 1.2MB of wallpaper to 125KB. - So there is no 389MB move to approve, and no folder rename to get wrong at two in the morning. `img/wallpaper` is now the archive and the source `tools/shrinkwall.py` reads; `img/wallpaper-small` is what the web server hands out. - STILL YOURS: getting both folders onto something that is not this disk. Neither is in git, and `img/wallpaper-small` is now load bearing, so it has to reach the web server too. - [x] **The guestbook cannot work on a static site.** Already handled: the form says plainly that nothing is stored, rather than appearing to accept a message and dropping it. --- ## Backlog Ordered so each phase is worth shipping on its own. ## Round nine: the BIOS became a control panel - [x] **The Setup screen was broken, and the bug was one word.** Every row is a `` joined with NOTHING, relying on those spans being `display: block`. The `.is-setup` rule overrode them to `inline-block` to centre the panel, so the rows flowed inline and wrapped wherever they fit: the tab bar ended up in the middle of the screen and the two columns interleaved. One wrapper element is the inline-block that centres now, and the rows inside it are blocks again. Two elements, two jobs, instead of one element asked to do both. - [x] **It fills the monitor.** It was drawn at the POST's font size, which is deliberately small because a real POST used the top left third of the screen. Setup is the opposite: it was a full screen application that used all 80x25. It has its own size now, `2.8vmin`, which puts 80 columns across 79 percent of the width at 1440x900 with no overflow. Every row is exactly 80 characters, which it was not: an off by one put three of them at 81 and pushed the panel 95px off the right. - [x] **It is a real settings panel.** Up and down walk the settings, left and right change tab, ENTER or plus or minus changes a value, F10 saves, F1 jumps to Help. Monitor Effect, System Sounds, Animation, Quick Boot and 1st Boot Device are all LIVE: each one reads and writes the same value the sidebar switch already owns, so there is one setting with two windows onto it rather than two settings. - A setting whose backing file is not loaded on that page reads `N/A` in dim ink and refuses, rather than pretending. `js/crt.js` is not on the private wing, so Monitor Effect is genuinely unavailable there. - [x] **A Help tab**, which no real BIOS has. This one does, because the site is a machine somebody has to work out how to use and the alternative is a visitor who never finds three of the four partitions. F1 from anywhere in Setup. - [x] **No mouse before an operating system is running.** The POST, Setup, the boot device menu, the network boot, the NTLDR menu and MS-DOS all hide the cursor and ignore both buttons. A BIOS had no mouse driver, DOS had none unless you loaded one, and a console never had one. - THE COST IS REAL AND IS WRITTEN DOWN BESIDE THE RULE. The clickable POST footer and boot menu rows were added deliberately, because F12 on a Mac laptop usually needs Fn held. They are inert now. Deleting one `pointer-events` line brings them back. - [x] **The POST is shown once per machine.** Seen it, and it is skipped entirely on every later arrival, landing at the boot menu instead, which is the part with a decision in it. Three things bring it back: a different browser, which has its own storage and so has never seen it; a changed machine, caught by a fingerprint of screen size, colour depth, pixel ratio, adapter, platform and timezone; and `?post`, for anybody who wants to watch it again. Quick Boot on the Advanced tab turns the whole behaviour off. - Quick boot used to be ONE stored flag meaning three different things, which is why it could never be switched off: the only thing that cleared it was never having booted. ## Round ten: one list of settings, three windows, and a disk - [x] **Pages Server Core split from the private wing.** See the section above. The terminal is public and the login is behind `TELNET` on the MS-DOS partition, which is the least sign posted thing on the site. - `js/tty.js` gained `LOGIN`, and the Fallout banner's last four lines, the ones that set a protection on `PRIVATE.WING` and run `AUTHELIA/FIRSTFACTOR`, are now behind it. Not because naming Authelia gives anything away, since a blog post on this site names it on purpose, but because they are the login's preamble and the public room has no login: they were a caption for something that does not happen there. - `js/shell.js` decides which room it is in once, at `shellStart`, and dresses `/etc/motd` and `~/README` to match. Both of those used to open by naming the wing, and `cat` is one of four commands this shell has. - `exit` in the public room closes the session and goes to the boot menu. It used to go to a login prompt, which handed the front door to anybody who typed the most obvious command in a shell. - [x] **The Termlink minigame came back.** Splitting the terminal took it away: it was offered after three failed passwords, and the public room has no password. It is a `hack` command now, still not in the help listing, with a manual page so that `man hack` does not read as a bug. - [x] **One list of settings, in `js/settings.js`.** The five settings were written inside the AMIBIOS Setup screen's closure, which was correct while Setup was the only screen with settings on it. It is not any more. - `SITE_SETTINGS` is the list. Each entry is a live view onto a real switch: `get()` reads the real value, `toggle()` goes through the real toggle, and a setting whose backing file is not on the page reads `N/A` and refuses rather than pretending. - The two storage keys live there too. `SYS_QUICK_PREF` was declared in `js/sysinfo.js` and `OS_KEY` in `js/osmenu.js`, and neither of those files is on every page, so a panel that wanted the value without them had to write the key out a second time. - [x] **A Control Panel in Pages 98.** All five, drawn as Windows would have drawn them, on the Start menu and the desktop. Display Properties is still there and still has the three display switches, because that is where a Windows user looks for those, but Quick Boot and 1st Boot Device are firmware settings with no business in a display applet. - [x] **A Settings panel in Pages Web**, in the sidebar, under a heading that says Settings. Three of the five used to be appended to the bottom of the *wallpaper* widget by `js/crt.js`, with no heading, because that was the only panel on the page and they had to go somewhere. Ninety lines of settings UI came out of the file that draws scanlines. - [x] **The three panels cannot disagree.** `js/crt.js`, `js/sfx.js` and `js/motion.js` each expose one change hook, and whoever assigns it last wins. Display Properties assigned all three when it opened, so opening the Control Panel and then throwing a switch in Display Properties left the Control Panel showing a state that was no longer true. There is a real list of listeners now, and the single slot hooks are chained rather than replaced. - `sfxSet` announces its own change. The notification used to be bolted onto the three callers that happened to have a switch on screen: the Ctrl+M handler and the `sound` and `mute` commands in the shell. Three copies of one line, and any new caller silently did not notify. - [x] **Help under the boot menu, on F1.** The first thing anybody sees is a question with a clock on it and three names they have never met. It now says what the three are, that each can reach the other two, that Restart comes back here, which keys do what during the boot, and where the settings are. `js/osmenu.js` never sets `is-help` and `is-adv` together, so there is no third combination for the CSS to have an opinion about. - [x] **The boot menu is sized in `vmin`, not `vw`.** The same fault the POST had: a screen with a fixed number of ROWS on it was being scaled by the window's WIDTH, so an ultrawide monitor pushed the bottom of it past the bottom of the window. `2.8vmin` is 22.4px at 1280x800, which is exactly what `1.75vw` was on the two machines this is built for. - [x] **MS-DOS got the rest of its disk.** The root listing named `GAMES` and `DOS`, and `CD` answered "There is nothing in there. That is the whole point of this partition", which is a joke that costs more than it earns: `DIR` promises two directories and the next command takes them away. - `CD` works, the prompt says where you are, and both directories have contents. `HIMEM.SYS`, `SCANDISK`, `DEFRAG` and `EDIT` moved into `C:\DOS`, where `CONFIG.SYS` and `AUTOEXEC.BAT` were already saying they were. - **`FDISK`** is the one worth the trouble. You are standing in the smallest partition on the machine, running the 1993 program for looking at partitions, and it can see the other three and cannot read a byte of any of them. - **`MSD`** reads the same real browser values `js/sysinfo.js` gives to System Properties on Pages 98 and to the Main page of the BIOS. Same browser, three period costumes. Its Mouse line is where the site explains its own no pointer rule. - `TREE`, `CHKDSK`, `VOL`, `PATH` and `ECHO`, and the up arrow walks what you typed, because `DOSKEY` is loaded in `AUTOEXEC.BAT` and that is the only thing anybody ever loaded `DOSKEY` for. ## Round eleven: the machine turns on and off - [x] **The order is the order a machine actually boots in.** POST, then the splash, then a boot device. It used to be POST, then the splash INSIDE the network boot, then the network boot, with no way to reach the firmware on any visit after the first. - **The splash is the firmware now**, which is what an OEM logo always was: the same firmware doing the same work with a badge over the top of it, and the same footer along the bottom offering DEL and F12. It is shown on EVERY boot. - **That is what fixed the real bug.** DEL and F12 were only ever armed while the verbose POST was on screen, and the verbose POST is shown once per machine. So Setup, the boot device menu and the whole MS-DOS partition behind it were reachable on a visitor's first arrival and never again. - **ESC on the badge shows the POST text**, the way ESC did on a real machine. It is the first way of asking for the verbose POST that does not involve editing the URL. - The boot menu is sized in `vmin` now, not `vw`. Same fault the POST had: a screen with a fixed number of rows was being scaled by the window's width. - [x] **A real boot device order, and the BIOS can change it.** The Boot page said this: ``` Boot Device Priority 1st Boot Device [Pages Web] 2nd Boot Device [Hard Disk] 3rd Boot Device [CD/DVD] ``` Only the first row did anything, and what it did was not what it said: its value was an operating system, not a device. The two rows that did name devices were painted text. - Three slots, three devices, stored as one value so the machine can never have two 1st Boot Devices. Changing one swaps it with wherever that device was, which is what pressing + on that row did. - **Put Hard Disk first and ENTER at the POST goes straight to MS-DOS.** - CD/DVD is empty, says so, and falls through to the next device. - The stored order is validated rather than trusted: a corrupt value cannot leave the machine unable to boot, and the machine is the site. - `1st Boot Device` is called `Default OS` now and has moved up into Boot Settings Configuration, where its name is true. - **BIOS only.** A Windows Control Panel cannot reorder the boot devices of the machine it is running on. `setFor()` in `js/settings.js` is how a setting says which screens may draw it. - [x] **Shut Down and Restart, on both graphical partitions.** Pages Web had neither: its only exit was a sidebar link that replayed the whole network boot. - **Restart** blacks the screen for about a second, the way a machine does between stopping and its firmware coming back, and lands on the splash. It does not replay the POST, because you have just watched this machine run. - **Shut Down** asks the browser to close the tab, which it will usually refuse, and then puts up the screen a stopped machine leaves behind. On Pages 98 that is the amber `It's now safe to turn off your computer`, which is the correct screen rather than a consolation prize. - Both mark the disk clean, so a deliberate shutdown is not greeted by `chkdsk` next time for having done the one thing right. - Pages 98's dialog is the real one: Stand by, Shut down, Restart, and **Restart in MS-DOS mode**, which is a one shot override of the boot order and does not change it. - Server Core gained `shutdown`, `poweroff` and `halt`. MS-DOS gained `SHUTDOWN`, which 6.22 never had, because turning a machine off was a switch on the front of it. - [x] **THE EDITOR WAS BEING ERASED HALF A SECOND AFTER IT OPENED.** `EDIT` painted correctly and then the cursor blink wiped it: `draw()` repainted the command prompt screen unconditionally into the same element, and the blink calls `draw()` every 500ms. - It survived because of how it failed. The blink interval is only started when motion is ALLOWED, so anybody testing with reduced motion on, which halves every animation on this site and is therefore what you test with, got a perfectly working editor. - One early return at the choke point fixes every caller at once. The editor's block cursor blinks now, which the real one did. - [x] **MS-DOS is the admin partition, and G: is the website.** `LOGIN` mounts the site's own files as a second drive: the scripts, the stylesheets, the page fragments and the documentation, on the machine serving them. `CD` and `DIR` walk it, `TYPE` fetches a file and prints it, `VIEW` puts a real photograph on the text screen, and `EDIT` opens a real file in the MS-DOS Editor. - **It is not the G:\ drive on the owner's computer and it cannot be.** The web root is `G:\Websites\xanderspage`. Above it sit an archive, a recycle bin and a folder called `KeePass Database`. A server rooted at the drive would publish a password vault, and `tools/serve.py` has a `--lan` flag that would then hand it to the network. The web root is the boundary, deliberately. - **F2 saves to this browser and says so in those words.** A static site cannot write to its own disk: there is no application on the other end of it. The draft survives a reload, does not survive a different computer, and is not on the website until it is copied into the file and the build is run. `js/paint.js` set that precedent and this follows it. - **`admin` is a mode, not a permission.** Everything on G: is public and fetchable by anybody who knows the URL, which is why `LOGIN` asks for nothing. A password prompt in front of a public directory teaches somebody the wrong thing about which of their locks are real. - `MANIFEST.site` is how it knows the names. A browser cannot list a directory over HTTP, which is the whole reason `js/manifest.js` exists. - [x] **The DOS prompt could not type a colon or an underscore.** There were two character maps and only the editor's was finished. Typing `G:` needs a colon and `baby_rabbits.jpg` needs an underscore, so the new drive was unreachable and so was most of what is on it. One map now, used by both. - [x] **A checker for the fault that broke four features silently.** `tools/manifest.py` emitted `.split("` followed by a real line break, which is an unterminated string literal, so `js/manifest.js` stopped parsing and `MANIFEST` went undefined. That took the wallpapers, the gallery, the Pages 98 desktop background and the home page featured panel with it, and nothing caught it: the manifest lane reads the names with a regex and the regex still matched. - `check_strings` walks 19,961 lines and reports any that end inside a quote. It knows a regex literal from a division, because the first version flagged six lines of `.replace(/"/g, """)` and a checker that cries wolf gets ignored. ## Round eleven, the part that was not done yet Twenty nine defects, found by six reviewers reading the round eleven changes and then an independent skeptic trying to refute each one. These are what survived refutation. Nothing here is speculative: every one names a file, a line, and the exact state that produces the wrong outcome. **ALL TWENTY FOUR FIXED on 2026-08-25**, one agent per file so that no two edits could land in the same file at once, each handed to a separate reviewer that reread the file rather than the report. **The reviewers were the useful half.** Five of the eleven fixes shipped with a new false claim inside the very comment written to remove a false claim, and those were caught and corrected before this was written down: - `TESTING.md` told the tester to look for **green text** on the POST screen. `css/boot.css` paints it `#c8c8c8`. The one screen you cannot check by reading the code is the one a wrong instruction wastes a trip to Perth on. - `js/sysinfo.js` said "three exits carry it" and then named four. A miscount inside the fix for a miscount. - `js/firmware.js` swapped the false claim about `?boot` for a smaller false claim about `?boot`. - `js/firmware.js` invented "Setup opens on DEL and on nothing else", which the boot menu's own Enter Setup row disproves. - `js/dos.js` told you F2 "was not pressed" in a case where it was pressed and the save silently failed, which is the same lie defect 516 was filed against. Two fixes went further than the ticket asked, and both were verified: - **571 was made true rather than deleted.** The comment claimed the off screen put the page behind it out of reach, "Tab included". Instead of correcting the comment down to what the code did, `powerSeal()` now actually traps focus, so the sentence is true. That is 96 lines of the change. - **502 was verified by execution, not by reading.** There is no `node` on this machine, but `cscript //E:JScript` is a real ES5 engine, so the boot chain was driven through a virtual clock harness: 22 assertions including the exact ESC-within-1.6s repro, and five deadlock probes for the risk the new guard introduces. It boots exactly once in all of them. Checked mechanically afterwards, across all eleven files: every file parses clean under that same ES5 engine, zero arrow functions, zero `let` or `const` declarations, and every file is 100 percent ASCII, so no em dash or en dash got in. **Still open, and none of them are code:** testing on the two MacBooks needs the two MacBooks, the `SERVICES` inventory is a decision rather than a defect, and the feature ideas below are features. One thing found and deliberately not fixed: `js/sysinfo.js` declares `var SYS_QUICK_KEY` twice, at lines 302 and 314, with identical values. It predates this round, `var` redeclaration is a no-op, and it was not on the list, so it is recorded here rather than quietly changed. ### The six that break something - [x] **`bootFrom()` can run twice and boot the machine twice.** `js/firmware.js`. `handOver()` guards on `chosen`, and `chosen` is set only by the network branch, so the CD/DVD and Removable Devices fallbacks sit for 1.2 to 1.6 seconds with the keyboard still attached and no guard set. Press ESC in that window and a second boot chain starts on top of the first: `opts.done()` fires twice, `typePost()` interleaves two PXE scrolls into one array, and the spinner slots point at the wrong rows. - Repro: F12 at the splash, arrow to Removable Devices, ENTER, then ESC within 1.6s. Also reachable with no F12 at all by putting CD/DVD first on the Boot page. - Fix: a re-entry guard in `bootFrom` itself, and `clearAll()` before the fallback timers, not only inside the hard disk branch. - [x] **The DOS editor tells you nothing could have been saved, when it just lost your typing.** `js/dos.js`. `editExit()` prints the C: wording for a G: file when nothing was saved with F2. That message is printed at exactly the moment it is most wrong. - [x] **`TYPE` on G: leaves "Reading ..." on screen.** `js/dos.js`. `printed.pop()` removes the blank line `submit()` pushed, not the progress line, so every successful read keeps a stale message above it. Fix: remember the index rather than popping. - [x] **A failed `CD` moves you to another drive.** `js/dos.js`. `cmdCd()` switches drive before validating the rest of the path, so `CD G:\NOSUCH` leaves you on G: root having lost where you were. The captured `var here = drive` is never read, which is the tell. - [x] **An in-flight `dosGet()` can repaint the DOS screen after DOS has exited.** `js/dos.js`. The callbacks never check `done`, so a slow response lands on top of the firmware's boot screen. - [x] **Three places tell the visitor the logo screen is on every boot.** It is not: a first boot and any boot with Quick Boot disabled show the POST text instead. `js/firmware.js`, and the two help screens. ### The Pages 98 Shut Down dialog - [x] **The accelerator handler destroys Escape.** `js/desktop.js`. It assigns `win.onkeydown`, overwriting the handler `win98Open` installed, so the dialog can no longer be dismissed with Escape. - [x] **Stand by does not close the dialog.** The comment says the `return` closes it. `win98Open` ignores an action's return value. ### Comments and documentation that are now false This project treats a lying comment as a bug, and a lot of behaviour moved. - [x] `README.md` still documents the pre-change boot order, cites a line number that does not do what it says, and denies behaviour that exists. - [x] `TESTING.md` section 1 describes a screen order that cannot happen. - [x] `js/boot.js` has a stale `SPLASH_MS` comment block that survived above its replacement, and a comment above the firmware handover still saying the chain is POST then F12 then the network boot. - [x] `js/firmware.js`: a comment claims Restart uses `?boot` and therefore always gives the full sequence. Restart uses `?restart`. - [x] `js/firmware.js`: two comments claim MS-DOS is reachable only from the F12 menu. The boot order reaches it, and so does Restart in MS-DOS mode. - [x] `js/firmware.js`: a block still says nothing in Setup changes anything. - [x] `js/sysinfo.js`: the paragraph naming which exits force a full boot names the wrong flag for three of the four, and one of them no longer reboots at all. - [x] `js/dos.js`: the header argues the partition must not have a filesystem or a text editor. It has both. The command count is wrong. - [x] `js/dos.js`: the comment justifying ESC without a confirmation says there is nothing to save, which stopped being true with G: editing. - [x] `js/main.js`: two comments still describe a setting called "1st Boot Device" whose value is an operating system. - [x] `js/power.js`: the comment claims the off screen puts the page behind it out of reach, "Tab included". A z-indexed overlay does not remove anything from the tab order. Either make it true or stop saying it. - [x] `js/settings.js`: the column key documents `help` as the BIOS help pane, but no screen renders a setting's own help array. ### Small - [x] `README.md` spare-icon paragraph is two spliced lists, and four of the icons it calls spare are now used. - [x] `css/chrome.css` focus-ring comment counts three `outline:none` rules and names a stylesheet that no longer exists. - [x] `js/boot.js` splash timing comment quotes four seconds and 1.8 seconds under reduced motion. Both numbers are wrong. - [x] `closeSetup()` and `closeMenu()` restore the badge on `SPLASH` (the mode) rather than `splashOn` (what is actually on screen), so leaving Setup after ESC has revealed the POST hides the POST again behind the badge for a beat. ### Still open from before this round - [ ] Testing on the two MacBooks. `TESTING.md` is the list. - [ ] The `SERVICES` inventory decision. - [ ] The 22 feature ideas below. ## Ideas, chosen from 53 Generated on 21 August 2026 by seven lenses proposing independently, then scored out of 30 by three judges: one for whether it can really be built under the constraints, one for delight per hour, and one for whether it belongs in this machine rather than just being a feature. 53 proposed, and these are what came out on top. - [ ] **THESE ARE ON THE LIST** Chosen on 21 August 2026. Not ordered by priority beyond the grouping, so take whichever one you fancy: they are deliberately independent of each other. Every entry names the files it would touch, and those file claims were spot checked against the tree rather than taken on trust. One number in the list as generated was wrong and has been corrected here: the button archive is 6,781 images, not 6,784. Build **Lost clusters** first. It is one evening in two files you already own, it needs no new page, no stylesheet and no change to tools/pages.py, and it is the only thing on this list where coming back next month is different because of what you did last month. SCANDISK currently finds nothing, which is the one thing on that partition that is inert. Make it find the machine's own past instead. #### An evening - [ ] **Lost clusters** SCANDISK's first run reports lost chains and asks "Convert lost chains to files?". Say yes and FILE0000.CHK appears in DIR, dated today, and is still there on every future visit; TYPE FILE0000.CHK gives a screen of garbage with one readable strip through the middle, a line from an AUTOEXEC that is not the current AUTOEXEC, a sentence from a NOTES.TXT written by somebody who does not think that way any more. Later scans turn up one more each, up to about six, numbered the way DOS numbered them. The fragments are a fixed array in the file, so every visitor's disk accumulates the same history rather than random noise, and the partition stays almost empty, it just stops being frozen. Files: js/dos.js (cmdScandisk, cmdDir, cmdType, DOS_DIR, DOS_FILES), js/store.js. - [ ] **Restart in MS-DOS mode** Replace the Shut Down box's two buttons with the real Windows 98 radio group: Shut down, Restart, Restart in MS-DOS mode. Shut down blacks the screen and prints "It's now safe to turn off your computer." in amber in the VGA face, alone on the tube, with a quiet press any key arriving three seconds later so it is not a trap. The third option prints the handover line and gives the whole screen to the DOS partition, which is already written, using the same detach, dosRun, done callback dance js/firmware.js performs around line 358. Files: js/desktop.js (appShutdown), css/desktop.css, tools/pages.py (add js/dos.js and css/boot.css to the desktop tuple near line 277), fonts/vga-9x16.woff. - [ ] **Abort, Retry, Fail?** At the C:\> prompt, A: or DIR A: prints "Not ready reading drive A" then "Abort, Retry, Fail?" and waits for a single key. R grinds and fails again, forever, exactly as it did. On Pages 98 the same absent floppy is visible from the other side: My Computer gains a drives view, and double clicking A: makes the seek noise and raises "A:\ is not accessible." over "The device is not ready." NOTES.TXT already says the backup is a box of floppies in the cupboard, so the drive is written into the fiction and is simply not there yet. Files: js/dos.js (dispatch near line 782, plus the single-key mode the editor's handler at line 641 already models), js/sfx.js (new sfxSeek built from the cached sfxNoise buffer through the sfxTick filter chain), js/desktop.js (appDrives, on appRecycle's .w98field markup). - [ ] **You are not alone on this box** `who` and `w` show two sessions: you on pts/0, and root on pts/1 from a LAN address, idle four minutes, running top. `finger root` prints the real layout with a Plan file, closing on a line admitting this user is invented, in the voice /home/xander/README already uses. Then once per session, unannounced, somewhere between forty and a hundred and twenty seconds in, a broadcast from root cuts across whatever you were typing, with your half typed command still sitting underneath because js/tty.js keeps input in the field rather than the buffer. The motd on that partition says nobody else is looking, and that sentence is waiting to be contradicted exactly once. Files: js/shell.js (shellDefine for who, w, finger, write; guard on tty.stage() === "shell"), js/sfx.js (sfxBell), js/tty.js (typeOut for root's reply). - [ ] **The battery is flat** Once the machine's own recorded first boot is more than a set number of days old, the POST starts printing "CMOS Battery Low", "CMOS Checksum Bad", "Press F1 to Run SETUP", and Setup's Main tab shows System Date as 01/01/2003, which is the date XANDERS.TXT says the disk was last touched. Load Optimal Defaults clears it for that visit and it is back the next time, because the battery is still flat. The trigger is real elapsed time, not a random draw: the machine gets old because it is old. Make it mutually exclusive with the keyboard gag in js/boot.js so two halt screens never stack. Files: js/sysinfo.js (new cmosFlat() next to sysFirstBoot, over xanders_firstboot), js/firmware.js (setupPanels, the POST step chain), js/boot.js. - [ ] **dmesg, reporting the visitor's real hardware** A kernel ring buffer with correct bracketed monotonic timestamps, opening on the init lines the Server Core boot already prints, then getting specific: CPU count, framebuffer resolution and pixel ratio, GPU string, platform, timezone, all read live and phrased as kernel messages, closing on mounting the wallpaper volume with a file count and Authelia binding its port. Push a line only when the value is actually present, so nothing ever reads undefined on Safari 7, where a shorter list is correct rather than broken. `dmesg -w` follows the buffer on the same takeover plus setInterval shape as top. The POST reporting real hardware is the best trick in the chain; this is that trick at the prompt, where somebody has chosen to keep typing. Files: js/apps.js (beside top and ping), js/sysinfo.js (sysPlatform, sysBrowser, sysRenderer, sysMachine, sysZone, sysClock), js/tty.js. - [ ] **Aero Peek, and a right click on the wallpaper** The whole site is glass over 350 wallpapers and there is currently no way to move the glass. Add a ridged Show Desktop bar at the right end of the footer: click it and every panel drops to a transparent wireframe with a white outline, exactly like Windows 7, and the picture underneath is finally visible. Second, a right click anywhere on the background gives a small Aero menu: Next background, Hold, Personalize, Show desktop, Restart, every entry calling something that already exists. Bind the context menu at document level and check the target, since the wallpaper layers sit at z-index -3 and -4. Files: css/style.css (new html.peek block, honouring html.motion-off the way section 24 does), tools/pages.py (SHELL and SHELL_HOME footer), new js/aeromenu.js modelled on dtDesk.oncontextmenu at js/desktop.js:1228, js/wallpaper.js (step, startTimer, wallPaused). - [ ] **Hardware configuration has changed** On the first boot the machine stores a small fingerprint of the browser reading it: screen size, colour depth, renderer, platform, timezone. Every later boot compares, and if anything moved the POST prints "Warning! Hardware configuration has changed since last boot." with the diff under it, "Adapter : was Intel HD Graphics 3000, now Apple M2", then the F1 prompt. It never fires on a first visit, so it is never noise. Keep it to one diff line: THE-MACHINE.md records the POST was cut from 42 lines to 38 to keep the American Megatrends badge on screen, and the warning plus prompt already spends most of the headroom. Files: js/sysinfo.js (new sysFingerprint and sysConfigChanged over xanders_hwcfg), js/firmware.js (one step before the footer), js/store.js. - [ ] **Start, Run, and "Cannot find the file"** A Run item above the Shut Down separator opens the real dialog, running figure icon, the full "Type the name of a program, folder, document, or Internet resource" sentence, an Open combo, and OK, Cancel, Browse. notepad, mspaint, pbrush, winmine, control, sysedit and telnet open what already exists; everything else gets the exact message box, "Cannot find the file 'sol' (or one of its components)." The combo remembers what you typed, most recent first, because the MRU list is the half of Run everybody used and nobody rebuilds. Browse points at a drive that is not there. Take the Run icon from the existing 98 set: tools/icons.py builds the Vista icons from an external archive and cannot supply it. Files: js/desktop.js (appRun, RUN_MAP, the APPS registry, dtRun), js/win98.js (win98Open, dialog mode with icon "!"), js/store.js (xanders_98_run), tools/content/desktop.html, css/desktop.css (.w98combo). - [ ] **The POST code display** A small amber pair of seven segment characters, bottom right of the boot layer, deliberately outside the 80 column text field because on a real board it is on the board rather than on the screen. D0 while the CPU initialises, C1 for the memory count, 2A for PCI, 03 for the keyboard controller, A0 at handoff. It never resets, so it rests on whatever screen you are sitting on, and if the machine halts on the keyboard gag or a bad checksum it stops on the code for that fault, which is free storytelling. Drawn with box shadows and the existing monospace face, no webfont and no image. Note that the element has to go into both tools/content/bootlayer.html and tools/content/private.html, which carries its own postLines div. Files: js/firmware.js (one postCode call per existing step), js/boot.js, css/boot.css, tools/content/bootlayer.html, tools/content/private.html. - [ ] **Ctrl+Alt+Del and the Close Program box** The real grey list of everything running, each open window by title, then Explorer and Systray, with End Task, Shut Down and Cancel. End Task twice on the same window gets the This program is not responding follow up first. End Task on Explorer takes the taskbar and every desktop icon away, leaving bare wallpaper for about eight seconds before it returns with the startup sound, which exposes that the desktop is a shell running on something. Pressing the combination again while the box is open reboots to the boot menu. It works on the two target MacBooks, where Ctrl+Option+Delete means nothing to macOS, but Windows swallows it, so give the box a second door: a Start menu entry or a taskbar right click. Files: js/desktop.js (dtCloseProgram, the existing document keydown reading ctrlKey, altKey, keyCode 8 or 46, appShutdown at line 888), js/win98.js (win98Windows at line 36, win98Open with raw:true), css/desktop.css (hiding .dt__bar and .dt__icons), js/sfx.js. - [ ] **Desktop themes, Hot Dog Stand included** appDisplay already prints an Appearance heading with nothing under it. Fill it with the real scheme names, Windows Standard, Brick, Desert, Eggplant, Maple, Plum, Rainy Day, Rose, Slate, Storm, Wheat and Hot Dog Stand, each repainting title bars, window faces, bevels, taskbar, Start menu and menu bars instantly and remembered between visits. Body class scoped overrides are exactly the technique the era used, so no custom properties are involved. Budget more than a dozen selectors per scheme: there are around 136 colour declarations across css/win98.css and css/desktop.css before paint98 and saver98. Add the class rather than setting body.className, which would clobber the pages98 class that scopes the whole font. Files: new css/themes98.css, tools/pages.py (desktop stylesheet list near line 288), js/desktop.js (appDisplay at line 742, a key beside DT_WALL_KEY), js/store.js. #### A weekend - [ ] **The 88x31 Archive** There are 6,781 rescued buttons on this disk and the site shows ninety of them. Build a browsable wall of the whole archive, two hundred at a time, with a filename filter, a shuffle and a permalink for any single button. The part that gets forwarded: most filenames still carry the GeoCities neighbourhood they were scraped from, like Area51_1269_images_xvtnow1.gif, so every button can offer a Wayback search for the dead page it used to live on. Do not flip LIST_ALL_BUTTONS, the names really are about 237KB; write them to their own generated js/archive.js with the same blob() trick and load it on one page only, the way js/gallerypage.js is loaded on wallpapers.html. pages/lostmedia.html already writes the essay about this archive and then shows none of it. Files: tools/manifest.py (new js/archive.js output), tools/pages.py (PAGES row, NAV row, DESCRIPTIONS, extra scripts), tools/content/archive.html, new js/archivepage.js (galLazy copied from js/gallerypage.js, shuffle from js/wallpaper.js), css/style.css section 19. - [ ] **The machine remembers what you did to it** One capped ring buffer of events written by things that already happen: an unclean shutdown, which OS you picked at NTLDR, Shut Down from the Start menu, a Minesweeper win, a picture saved in Paint, a guestbook entry composed, a failed login, a completed boot. Surface it as an Event Viewer window on Pages 98 with the real three column layout and the i, ! and x icons, and as one line on the POST: "Last shutdown : Pages 98, unclean, 00:03:11 ago", which is the sentence that makes somebody realise the machine was running while they were gone. That POST line must be exactly one row. Every writer is a single line dropped into code that already runs, and logAll must always return an array even when the stored value is half written. Files: new js/eventlog.js, js/store.js, js/chkdsk.js (CHK_KEY), js/osmenu.js (OS_KEY), js/desktop.js (appShutdown, a new APPS entry on dtWindow), js/minesweeper.js, js/paint.js (ptFilesWrite), js/guestbook.js, js/tty.js, js/sysinfo.js (sysBootSeen), js/firmware.js, tools/pages.py (three script blocks). - [ ] **view: 350 wallpapers as phosphor ASCII** `view` draws a random wallpaper into the tube as ASCII at the measured column width, using a ramp from space through .:-=+*# to @, with the darkest run in tty__dim and the brightest in tty__hi so the phosphor has depth. `view landscapes/moraine_lake` names one, `ls /var/wallpaper` lists the seven categories, `slideshow` holds the screen and changes every eight seconds until q. Rendering is an offscreen canvas, drawImage down to tty.columns() wide by about half that tall because a character cell is twice as tall as it is wide, then getImageData, same origin so nothing is tainted. The green wing currently cannot see a single pixel of the largest thing on this machine. Files: new js/ttyart.js, tools/pages.py (add js/manifest.js to the private wing script list near line 297, it is genuinely missing), js/tty.js, js/apps.js (the run coalescing close() helper in matrix at line 418), img/wallpaper-small. - [ ] **CMOS that actually holds** Make the Advanced and Boot tabs real: Quick Boot, Halt On, Sound, Monitor effect, Animation, and Boot Device Priority, which reorders the NTLDR menu itself because osDefault() already reads xanders_lastos. Every one of those settings exists today with its own switch in its own room, so the machine can disagree with itself about its own settings depending on which door you came in. Then give the Exit tab Load Setup Defaults, which really does clear every xanders_ key and reload, and Discard Changes: clearing the CMOS resetting your whole computer is a joke that is also completely accurate. Two things are additions, not reuse: drawSetup has no row cursor at all, only tab inversion and heading brightening, and Setup rows are not clickable today, so the data attribute click path has to be built the way the DOS boot menu was. Files: js/firmware.js (setupPanels, drawSetup, the setupOpen key branch near line 777), js/store.js, and the existing keys in js/sysinfo.js, js/sfx.js, js/motion.js, js/crt.js, js/osmenu.js. - [ ] **The 88x31 forge** Type your site name, pick a palette, and a canvas draws you a real Aero glass 88x31 button on the spot: vertical gradient, glass highlight across the top half, one pixel white border, your text in the middle, with a 1x and 4x preview and the paste snippet underneath. Port the recipe out of tools/mybuttons.py, VARIANTS and all, so the four buttons you ship and the ones a visitor makes are drawn by one recipe and can never drift. Saving is right click, Save Image As, and the page says so plainly, because Safari 7 has no download attribute. This would be the first script pages/buttons.html has ever loaded, so guard every element lookup the way js/main.js does. Files: new js/buttonforge.js (from tools/mybuttons.py), tools/content/buttons.html, tools/pages.py (buttons row of PAGES), js/guestbook.js (gbMailto for the Send me yours handoff), css/style.css. - [ ] **Dial-Up Networking, with the handshake** A Connect To box with a connection called xanders.page, a user name filled in, a number and Save password ticked. Connect gives Dialing with the touch tone pair of each digit played in turn, then the carrier handshake, then Verifying user name and password, then Connected at 49,333 bps, and the dialog collapses to two flashing monitors in the tray with Status and Disconnect on a right click. The site has a Web Audio synth and has never once made the noise that defined being online in 1998, and every ingredient is already in js/sfx.js. Do not store the connection state: a modem still connected when you come back tomorrow is not a modem. Under reduced motion, hold the tray lights still and keep the sound. Files: new js/dialup.js, js/desktop.js (appDialup), js/sfx.js (sfxTone at 121, sfxNoise at 194, sfxStack at 289), tools/content/desktop.html (.dt__tray), css/desktop.css line 482, tools/pages.py (desktop script block near line 268). #### A big one - [ ] **lynx, the text mode browser** `lynx about` or `lynx /pages/gallery.html` fetches that real page off the same server and renders it as text: headings underlined, paragraphs wrapped to the measured column width, links numbered [1] [2] [3] in bright phosphor, type a number to follow, b to go back, q to quit. It is browsing the actual public site, so the 20 pages tools/pages.py already builds are the content and nothing new gets written. Server Core has no graphical shell, so a text browser is not a novelty, it is the only way that room can see the web, and what it sees is this machine's own other partition. Strip script, style, link and img by regex before assigning into a detached div, or WebKit fires the image requests anyway; on status 0, which is what a file:// open gives, print the honest "lynx: Unable to connect to remote host." Files: new js/lynx.js, tools/pages.py (private entry script list, lines 296 to 313, after apps.js), js/tty.js (print, esc, columns, takeOver, release, setPrompt), js/shell.js (shellDefine, shellPromptText), css/tty.css (tty__hi). - [ ] **C: is one disk, seen through four windows** There are three unrelated invented filesystems today, DOS_DIR and DOS_FILES at js/dos.js:49 and :63, FS at js/shell.js:62, and a single Notepad buffer under xanders_98note at js/desktop.js:1342, and none of them knows the others exist. Replace the overlap with one volume: NOTES.TXT in the DOS listing, ~/notes/todo.txt in Server Core and a My Documents window on Pages 98 all read and write the same bytes. Type a line in Notepad, shut down, boot MS-DOS behind F12, TYPE NOTES.TXT, and your line is there with DIR reporting the new byte count. Canonical form is a lowercase POSIX path holding a string, shipped defaults stay as literals so an empty machine still has content, and only the differences go to storage. Two consequences to price in: everything shared lands in one key that Safari 7 in a private window silently refuses, which store.js is built to survive, and giving the MS-DOS Editor a real File > Save turns its honest "it does not save" notice into a write that DIR must resize. Files: new js/disk.js (diskRead, diskWrite, diskList, diskStat over js/store.js), js/dos.js (DOS_FILES becomes a lookup, the EDIT exit prompt), js/shell.js (shellResolve serving /home/xander), js/desktop.js (appDocuments on dtWindow, appNotepad), tools/pages.py (HOME_SCRIPTS near 835, SHELL_98 near 269, SHELL_BARE near 297). - [ ] **The floppy somebody left in the drive** Booting Removable gives an A: prompt with four files on a 1.44MB disk: AFUDOS.EXE, H410MA.ROM, FLASH.BAT and a README from whoever made it. FLASH.BAT plays the real AFUDOS screen, reading the file, erasing the block, the write bar crawling across, verifying, DO NOT TURN OFF THE SYSTEM across the middle, then reboots through ?boot, and from then on the banner, the Setup Main tab and the Boot tab all report the new revision, permanently. Two of the five boot devices are still drawn, selectable, clickable and dead, and this one does the thing only firmware can do: it changes the machine itself. Closing the tab mid flash is already caught by the dirty flag, because the DOS session runs before finish() clears it, so the next boot recovers from the boot block, which is what those boards genuinely did. Files: js/dos.js (dosRun with an options.drive of "A", or a sibling js/floppy.js), js/firmware.js (DEVICES[2], chooseDevice at 337 to 368, SERVER.bios at 98 becoming a cmosGet with "1401" as its default), js/chkdsk.js, tools/check.py. --- ### Round eight: auditing the documentation against the code The backlog audit in round five found eight false claims, and updating the README an hour later found its description of `tools/check.py` listing four lanes when there were nine. So the same treatment was given to the prose: 1,511 lines of README plus ROADMAP and ASSETS, six readers over six slices, every reported error handed to a second reader told to refute it. **347 claims checked, 55 reported, 52 confirmed.** Roughly thirty distinct corrections after merging duplicates. The refuters killed three, which is the same one-in-eighteen false alarm rate as the backlog audit. - [x] **A real leak, and it was mine.** A paragraph in `README.md` named the two tunnel products on this machine, in the course of explaining that they were being kept off the site. Worse, the comment in `js/firmware.js` that removed them said the details lived in THE-MACHINE.md, "which is not deployed". **That was false.** The repo root IS the web root, as the first line of the README says, so every markdown file here answers with a 200, along with `.gitignore` and every script in `tools/`. Checked with curl rather than assumed. The names are gone, the false sentence is corrected, and the decision about whether the docs should be served at all is now an open item above rather than an accident. - [x] **The machine was still documented as having two operating systems.** The front door diagram and the NTLDR transcript both listed Pages Web and Pages Server Core, with Pages 98 missing from between them, and the banner comment in `js/osmenu.js` said "Two entries" in as many words. The MS-DOS partition was not in the diagram either. - [x] **`js/buttonpage.js` said it builds `pages/buttons.html`.** It builds `pages/collection.html`. The error was in the file's banner, which `tools/layout.py` copies verbatim into the README, so the README was regenerating it every run. Fixed at the source. - The same confusion ran through a whole README section, which described one page in five parts when it was really the first part of `buttons.html` and all four parts of `collection.html`. Rewritten as two pages. - [x] **Counts, all measured off the disk rather than taken from the audit.** Fifteen scripts is sixteen. Fifteen pages is twenty. Eighteen site map entries is twenty. Fifty five linked buttons is ninety. Thirteen exo.pet gifs is fourteen. 6,781 archived buttons is 5,565, and 47MB is 48MB. The icon set is 192KB across 48 files, not 7KB. Fifteen sidebar icons is twenty, and three of the four "spares" listed are now in use. - [x] **Wallpaper sizes were wrong by half and then overtaken.** "1920x1200 averaging 560KB" appeared twice; the real mean is 1124KB. It is moot either way now, because the site serves `img/wallpaper-small`, which is 1600x1000 at a mean of 220KB. - [x] **Things last night invalidated.** The POST transcript still showed an Array Configuration line it no longer prints, and two paragraphs described a Terminal Configuration block that was removed. The terminal was documented with five text shadow layers when it has three, and three focusable inputs when it has four. - [x] **`cmd` survived in the command list and in the quoted motd**, weeks after the Command Prompt was deleted. A command list is exactly where a deleted command lives longest. The list is now complete and correct against `shellDefine` in both files: twenty six commands and seven programs. - [x] **`css/login.css` does not exist and never has.** The reset it credits is in `css/tty.css`. - [x] **The Fonts section and three ASSETS sections were instructions for work already done**, describing gradient squares in a sidebar that has twenty real icons, a marquee that was removed, and a `site/img/` prefix that has never existed. And the note explaining why `woff` is listed before `woff2` describes the opposite of what the file now does, for a reason that was wrong: an unknown `format()` is SKIPPED, not fatal. - [x] **The 404 page does not share `SHELL_HOME`.** It has `SHELL_404`, which carries no boot placeholders rather than empty ones. - [x] **The workflow itself had a bug worth keeping.** Its first run hit a usage limit, all six readers died, and it returned "the documentation checks out" because the findings list was empty. It was empty because nothing had been read. Silence from a checker has to mean "I looked and found nothing", never "I did not look", so it reports the failure now. - [x] **The keyboard section claimed "exactly one interception"** and there are four: select all, extend the selection, the running program hook that lets Ctrl+C interrupt `sl`, and Ctrl+M for the sound. All four are deliberate, but four is not one. It also said arrows are taken only during 2FA; the password game takes them too. - [x] **The two phase network boot was listed in the wrong order.** LOAD comes first: `sequence()` shows the splash and only calls `typePost()` four seconds later. The list had PXE first, which is what you would guess from the name and not what happens on screen. ### Round seven: walking the boot chain end to end - [x] **The guestbook form overflowed the 980px column**, found by measuring the page at 1280x800 rather than looking at it. Two bugs from one cause: the 100px label offset was on the FIELD'S MARGIN. - The textarea carried `width: 100%`, and a percentage resolves against the parent, so 100% of the row plus a 100px margin reached 100px past the row. It ended at 1087 where the column ends at 1122. - The two text inputs had `width: auto`, which fills the line for an ordinary block box but NOT for an input: an input is a replaced element, so auto means its intrinsic size. They sat at the browser default of 143px beside a 738px textarea. - The offset is the ROW'S PADDING now, so `width: 100%` means the space that is actually left. All three fields are 638px and identical, and the form still validates and signs. And then the thing that actually unblocks the top of this file: - [x] **`TESTING.md`**, a written pass for the two MacBooks. See the Phase 0 item above for why the shape of it matters more than the content: an obligation with no edges never gets started. - [x] The README now points at all four root documents, and its description of `check.py` was rewritten, because it still listed four lanes when there are nine. - [x] Confirmed by reading `fonts/vga-9x16.woff` rather than guessing: the face carries all eight box drawing and block glyphs that DEFRAG and the editor draw with. So if those come out wrong on a Mac it is the font failing to LOAD, not the font lacking the character, which is a different fix and worth knowing in advance. Then a resolution pass at 1280x800 and 1440x900, the two target MacBooks, which caught two things in the editor built an hour earlier: - [x] **The editor was double spaced and its field ran twice as wide as its frame.** Both from one wrong `display: block`. `js/dos.js` joins rows with `
` like every screen in that file, and a block breaks the line too, so every row broke twice: 658 pixels of screen where 22 single rows should be 337. A block also fills its container, so the blue ran 1219px against a 78 character frame that stops at 590px. `display: inline-block` fixes both. Worth recording that I had already LOOKED at a screenshot of this and read the black gaps as part of the frame. Measuring it is what caught it. - [x] **The skip link was `position: absolute`.** It is the first tab stop, so you normally reach it at the top of the page and never notice. Shift Tab from halfway down a long page is the case that was broken: absolute positions against the document, so it would have appeared at the top of the PAGE, off screen. Fixed, and lifted above the CRT overlay, because a control whose whole job is making the page usable should not be the one thing showing through a filter. - [x] **`tools/check.py` checks the 350 wallpapers and their thumbnails.** This was the one class of broken path nothing could see: no wallpaper path exists in any file, because they are assembled at runtime in JavaScript from a folder name and a filename in `js/manifest.js`. So the largest thing on the site and its best page had no check at all. - It matters more since the site started serving `img/wallpaper-small`. That folder is gitignored, correctly, and it is now LOAD BEARING: a deploy that copies only what git tracks would put up a site whose every wallpaper is a broken image, and nothing would have caught it before a visitor did. - It reads the folder name out of the four files that build the paths rather than assuming it, so if those paths move again the check follows them. - Verified by hiding one wallpaper and one thumbnail: both reported by name, exit 1, count down from 700 to 698, and clean again once they were put back. The whole chain was walked from a cleared machine state, which had not been done since round three changed most of it: cold boot, POST, ENTER, the PXE failure and retry, DHCP, TFTP, iPXE, the hardware probe, the operating system menu with its countdown, and a boot into Pages 98. No errors anywhere, and the POST held its badge the whole way. Two things came out of actually walking it rather than reading it: - [x] **The MS-DOS partition could not be reached with a mouse.** The POST footer was made clickable in an earlier round precisely so that DEL and F12 would not depend on a function key. That gets you as far as opening the boot device menu, and the menu then wanted arrow keys and ENTER, so the door led into a corridor with no handle. It matters more here than anywhere else on the site, because the MS-DOS partition is reachable through that menu AND NOWHERE ELSE: a visitor who cannot work the arrow keys could not reach a whole operating system. - Each row is clickable now. Click to select, click the highlighted row again to boot it, which is the convention `js/osmenu.js` already uses one screen later, and which means a single stray click can never start an operating system. - The ENTER branch was 73 lines inline, including the whole DOS handover with its `detachKeys` dance. It is `chooseDevice()` now, so the key and the click run the same code. Two copies would drift, and the half that drifted would be the mouse half, because somebody testing with the keyboard never takes that path. - Verified by mouse alone through every branch: Hard Disk boots MS-DOS, CD/DVD prints the real "Reboot and Select proper Boot device" failure and falls back to the network, and the keyboard still works. - [x] **The boot said the same thing twice.** Round four added a "Your Terminal" block to the POST, and the network boot ends with "Probing host hardware" printing the same facts and more of them about twenty seconds later. The POST block went, because the probe block belongs to a screen that scrolls and the POST is a fixed screen that pins to the bottom, where four lines is four lines of headroom the badge needs. The POST is 38 lines now, and `js/firmware.js` no longer calls `sysReal()` at all. ### Round six: the keyboard, and the switches nobody could reach - [x] **Every external link on the site was checked, and all 78 hosts are alive.** `python tools/check.py --links`, run 21 August 2026. The README asked for this to be re-run occasionally and it had not been since the pass that caught Angelfire and Tripod dying in April. - It found two failures and BOTH were faults in the checker, not on the site. `www.adobe.com` hangs for the full timeout on every attempt while `get.adobe.com` answers instantly and every other host replies in under a second, which is a bot filter rather than a dead host, so timeouts are a note now instead of a fault. - The second was better: a host called `xanders.page `, reported dead every run. `js/guestbook.js` contains `"...at https://xanders.page "` and in a SOURCE file that ` ` is two characters, neither of which the URL pattern excluded, so the trailing escape became part of the hostname and sailed past the rule that skips this site's own domain. A URL never ends in a backslash, so cutting at the first one fixes the whole class. - [x] **Pages 98 had the CRT filter and no way to turn it off.** `js/crt.js` hangs the monitor, sound and motion switches off `.wallpanel`, the sidebar panel every other page carries. The 98 desktop has no sidebar, so it took the scanlines over eleven pixel text, plus the sounds and the animation, with no control over any of it. The one partition most in need of a monitor switch was the only one without one. - They live in **Display Properties** now, which is where they belong on a Windows desktop and which was, until tonight, a window called Display Properties containing no display properties. - `crtToggle` and `crtOn` had to be exported from `js/crt.js` for this; `sfxToggle` and `motionToggle` were already global. The sidebar switch now calls the same exported toggle, so the two cannot drift. - The window claims `sfxOnChange`, `motionOnChange` and `crtOnChange` while it is open and HANDS THEM BACK on close. Each hook holds only one handler, so a window that took one and kept it would leave the sidebar switches on every other page unable to relabel themselves for the rest of the session. - [x] **A skip link and landmarks**, on the seventeen pages that have a sidebar. Measured on the About page: **35 tab stops** sit between the top of the document and the first word of the content. On a responsive site you would put the content first in source order; this layout is deliberately fixed at 980px forever, so source order IS the order and a skip link is the whole fix. - Off screen and clipped until focused, NOT `display:none`, which would remove it from the tab order and produce a skip link that can never be reached. - Not on Pages 98 or the private wing: both are full screen takeovers with no navigation, so there is nothing to skip and nowhere to go. - [x] **The terminal no longer traps the keyboard.** Two separate bugs. `css/tty.css` claimed `z-index:-1` and `pointer-events:none` kept the inactive fields out of the tab order; neither property does that, so Tab at the login prompt moved the real caret into the hidden PASSWORD field while the drawn caret stayed put. `focusField` sets `tabIndex` now, which is the thing that actually does it, and not `disabled`, which would break password manager support. Separately `js/shell.js` swallowed Tab on a DOCUMENT level keydown whenever the stage was "shell", which is the whole time a dialog is open over the tube, so Tab was dead page wide with nothing on screen explaining why. Scoped to `tty.atPrompt()`. - [x] **Pages 98 is usable from the keyboard.** All nine desktop icons were tab stops that announced themselves as actionable and did nothing: opening was on `ondblclick`, and there is no double click on a keyboard. Enter opens now, which is what Windows did. `js/win98.js` had no key handling at all and never moved focus, so an opened window left the caret at the top of the page and the Shut Down dialog could not be dismissed; Escape closes the window and focus moves into it on open. The white ink focus list in `css/chrome.css` was written for the Aero side and never grew, so the blue screen's four exit links carried a black dotted ring on `#0000aa`. - [x] **A Safari 7 lane in `tools/check.py`.** Constraint 1 of this project is its hard floor and was the only rule with no mechanical check, while the rule about dashes had a thorough one. It scans `js/` AND the inline scripts of the generated pages, because the head scripts live as Python strings in `pages.py` and are the first things that run. Verified by planting seventeen real violations and confirming all seventeen are caught with exact line numbers. - Its first version had two bugs of its own, both found by reading the output rather than trusting it. Stripping comments replaced them with a single space, which collapsed the newlines inside them, and this project is mostly forty line banners, so every reported line was tens of lines out. And the `-webkit-` twin check compared per BLOCK, which is wrong for keyframes, where the prefixed version is a whole separate at-rule by design: fifteen confident reports, every one wrong. - NOTES ARE NOT FAULTS. Some things cannot be decided by grep, so they print under their own heading and do not touch the exit code. A command that always exits 1 is a command whose output stops being read, which is the same failure the fifteen wrong reports would have caused. - [x] **Pages 98 has its own sound scheme.** It was the partition a visitor spends longest in and the silent one, on a machine whose other three all beep. Windows open, close, minimise, restore, a menu tick, a Critical Stop and a shutdown chord, in sine and triangle rather than the square waves everything above them uses, because those are a PC SPEAKER and by 1998 the sounds came off a sound card. The error boxes in Paint and the guestbook window used `sfxBeep`, the PC speaker, which was exactly the inconsistency this fixes. - [x] **chkdsk's arithmetic works and describes this machine.** It claimed a 232 GiB volume matching neither the POST nor `df`, and its three subtotals came to 269,760 KB MORE than the disk they were on. It is the 931 GiB mirror now, the same volume `df` reports, with the parts summing exactly to the total and the allocation units dividing evenly. - [x] Smaller: the gallery's meta description promised "art and images" on a page whose first sentence says there is none; the front page printed `tools/content/index.html` at a visitor, a source path that is not on the web server, so anybody who followed it would get the blue screen; and the home page description still mentioned the scrolling buttons that were removed in round four. ### Round five: auditing this file against the code The survey in round four turned up three `[x]` items whose claims the code did not support. Three, found by three different readers who were looking for something else, is a pattern rather than bad luck, so all 208 completed claims in this file were checked against the code, and every reported failure was then given to a second reader told to REFUTE it. That second pass earned its place: it killed three of the eleven reports as misreadings, including one that objected to the word "keeps" in a line about persistence. Eight stood up. **What the audit found, and what has been done about it:** - [x] **The `portal` flag was only honoured in one of two places.** This was a gap in the round four fix itself, made a few hours earlier. Both no fetch branches in `js/authelia.js` return `portal: true`, but only `submitPass` in `js/tty.js` checked it. `submitCode`, the second factor path, still read `ok:true` as "correct code" and would print ACCESS GRANTED for any six digits on Safari 7. Unreachable today, because the first factor now hands off before you can reach that stage, which is EXACTLY the reasoning that let the original bug live. Fixed, and the two halves now match. - [x] **The wallpaper `aria-label` reached sixteen pages and skipped the front page.** Also from round four. `WALLPANEL` in `tools/pages.py` is substituted into `SHELL`, which builds `pages/*.html` only. The home page is built from `SHELL_HOME`, which has no `{wallpanel}`, and the nav rewrite stops at a lookahead for that very block. So the page with the most visitors was the one page without the label. All seventeen have it now. - [x] **The updates page still advertised "an NT console".** That page opens by inviting you to check whether the site is really being kept current, which makes a line selling a deliberately deleted feature the worst stale sentence available anywhere on the site. - [~] **`color`, `tree` and `shutdown -s` were marked done and do not exist.** All three were built inside the XP Command Prompt and deleted with it in round three. Re-marked `[~]`, built then removed, rather than deleted from this file, so the record survives without the claim. - [~] **The system configuration table was marked done and is not built.** Something good exists in its place, but there is no box, no second column, and none of the seven fields the item names. See the entry itself for the full correction, including that the "Probing host hardware" block it promised to replace is still there. - [x] **A stale comment in `css/boot.css`** justified a white-space rule by "the whole drawn box of the system configuration table", a box nothing draws. It now names the frames that really do depend on it, drawn by `drawSetup` and `drawMenu`. - [x] **The wallpaper swap note** described a 389MB folder move as still pending. It is not pending, it is unnecessary: the paths were pointed at the small set instead and nothing has to move. **The lesson, which is the reason to write this down.** Every false claim here was written at the moment the work was done, when it was true or nearly true, and went stale later when something else changed. Two of the eight went stale within hours, in the same session that wrote them. A backlog is not a record of what happened, it is a claim about the present, and it decays like any other. `tools/check.py` cannot check prose, so the only defence is reading it against the code occasionally. ### Round four: from a survey of the whole site An eight lens survey of the repository, with three skeptics arguing against everything it produced. 58 proposals, most rejected. What survived and has been done, on 21 August 2026, all verified in a browser: - [x] **The POST fits on one screen again.** Round three took it to 52 lines against a screen that holds about 50, and `.boot__post` pins itself to the bottom on every write, so the overflow came off the TOP and took the American Megatrends badge with it. It is 42 lines now. **And the font is `vmin`, not `vw`,** which is the deeper fix. Sizing text off the width while the height is what runs out made the row count depend on the shape of the window: about 50 on a 16:10 laptop, 45 on 16:9, and 33 on a 2560x1080 ultrawide, where the old rule gave a 27px font. It was already clipping before round three added a line to it, and there was no way to see that without owning the other monitor. Verified at 1280x800, 1440x900, 1920x1080 and 2560x1080. - [x] **Safari 7 no longer prints ACCESS GRANTED for any password.** `js/authelia.js` returned `{ok:true, redirect}` when `window.fetch` is missing, meaning to hand off to Authelia's own portal. `js/tty.js` reads `ok && redirect` as "authenticated, no second factor owed", prints ACCESS GRANTED and calls `window.shellStart`, which takes over so the `location.href` handoff never runs. On the PRIMARY TARGET MACHINE, any password opened the private wing shell. Latent rather than live: `autheliaWhere()` returns "local" on localhost and the site is not deployed, so this would have appeared on the first day it went up. The no-fetch branches now carry `portal: true`, and tty.js routes those to a real navigation. - [x] **The guestbook accepted `javascript:` URLs.** `gbTidySite` returned any string containing `://` untouched, which `javascript://x%0Aalert(1)` satisfies, and it went straight into an href. It persisted in localStorage and re-rendered on every later visit, on the page and in the Pages 98 window, on the origin that draws the Authelia prompt. Now an http and https whitelist. Everything else falls through to the `http://` branch, which is a safe result rather than an error. - [x] **Paint's saved images are checked and escaped.** The data URL was the one field in that markup not passed through `ptEsc`, and it comes back out of localStorage rather than out of memory. It is now verified to start `data:image/` and escaped, in Paint and in My Pictures. `dtEsc` escapes double quotes now too, which it never did, while most of its calls are inside attributes. - [x] **The head paint script validates what it reads.** It puts a stored value straight into a CSS `url()`. It now requires the path to sit under the folder actually served, and rejects quotes, brackets and backslashes. That also caught a real regression: after the wallpaper switch below, anyone with an old path stored would have pulled the 1.2MB original on the first paint of every page. - [x] **`tools/serve.py --lan` starts.** `PORT = int(sys.argv[1])` ran at module level with no flag filtering, so the one command that opens the site to the two MacBooks died on `int('--lan')` before binding anything, with a traceback that never mentioned `--lan`. It had never worked. All five argument forms tested. - [x] **`tools/thumbs.py` runs.** Line 21 read `ROOT` before anything bound it, so the module died on import with a NameError. It is the only tool that was missing the standard root line, and the name it bound then shadowed that root, which is how it hid. - [x] **`tools/cursors.py` cannot destroy `css/cursors.css` any more.** It reads from a Downloads folder that is in neither the repo nor the backup zip, printed MISSING for each absent file, then called `build_css` anyway, overwriting a tracked stylesheet with a valid file containing no rules. Every custom cursor on the site, gone, and `check.py` passes a rule-less stylesheet because it has no `url()` lines to fail. It refuses twice now: no pack, and nothing converted. - [x] **A failed `pages.py` stops `blog.py` before it publishes.** pages.py never called `sys.exit`, so it returned 0 after printing MISSING FRAGMENTS; blog.py judged it by that code and had a "FAILED" branch that could never fire; and `build_feed()` sat outside the guard, so the feed went out anyway. A feed item pointing at an anchor that was never built is cached by every reader that fetches it, and is the one thing here that cannot be taken back. blog.py also stopped swallowing pages.py's output, which was hiding the reason. - [x] **The site serves `img/wallpaper-small`.** `tools/shrinkwall.py` was written to produce it and says why, and nothing had ever pointed at its output: six path strings still named the 1920x1200 originals. The home page went from about 1.2MB of wallpaper to 125KB. The folder is gitignored now, since it was untracked only by luck and one `git add -A` would have put 76MB into history forever, with a note that it IS load bearing for deployment even though it is not tracked. - [x] **The wallpapers page lazy loads its thumbnails.** 350 cells all carried a real `src`, so opening it asked for 2.9MB in one go, on a page most people never scroll to the bottom of. It is 63KB on load now, filling in on scroll. Not IntersectionObserver, which is Safari 12.1, and not `loading="lazy"`, which is Safari 15.4 and would be ignored rather than degrade on exactly the machines with the problem. - [x] **F8 does what the POST says it does.** The screen printed "Press F8 for BBS POPUP" and `isKey` had no case for F8, so it was dead on every machine. Both body lines are clickable now, like the footer. - [x] Smaller: the crash screen's Restart link carries `?boot`, without which the boot gate silently drops you on the front page with no restart; `id` no longer claims group 27, which is exactly what puts an account in the sudoers file `sudo` says it is not in; the `man` page for `cmd` is gone, a confident manual for a command removed in round three; the equaliser on the front page no longer writes to fourteen spans seven times a second forever while nothing is playing; `js/sfx.js` wraps its whole setup in the try, since `js/bsod.js` owns `window.onerror` and a throw there would paint a crash screen over a half drawn POST; `js/buttons.js` is off the home page, where it was 12.8KB fetched to look for an element that is no longer in the page; the wallpaper select has an `aria-label`; the guestbook form is `method="post"`, so a script failure cannot put somebody's message in the address bar; and `store.js` is no longer loaded twice. Declined, with the reasoning recorded in the file itself rather than only here: the splash logo stays at 256px, because it displays at 96 CSS pixels and that is 192 physical ones on the Retina MacBook, and `srcset` is not available on Safari 7, so the older machine is the one that would break. ### Round three: eight asked for by name All eight done and verified in a browser on 21 August 2026. These came from using the thing rather than from an audit, which is why every one of them is about a rough edge no checklist would have found. - [x] **The navigation flash.** Clicking through the site left up to two seconds with no CRT filter and no wallpaper. Two separate causes. `js/crt.js` BUILT the scanline overlay, so it could not exist until a script at the bottom of the page had run; the overlay is now plain markup in the shells and `crt.js` adopts the one it finds rather than appending a second. And the wallpaper layers ship empty at opacity 0 and fade up over 1.4s once the manifest parses, which is the intended effect on a first visit and just a wait on the fifteenth navigation. `PAINT_HEAD` in `tools/pages.py` now paints the last wallpaper during head parsing with the transition off, and `js/wallpaper.js` resumes on that same picture and drops the rule once its own crossfade settles. Resuming also fixed a bug nobody had reported: the rotation restarted at `wallList[0]` on every page load, so it could never get past the first picture in a category. - [x] **The guestbook signs**, on both Pages Web and Pages 98, sharing `js/guestbook.js`. It fills in your mail program rather than pretending to post. See the README for why this is the third design of it. - [x] **Server Core scrolls back.** `.tty__screen` was `overflow: hidden` while `draw()` had always ended with `scrollTop = scrollHeight`, so output scrolled in and could never be scrolled back to. `help` prints more commands than fit and the top of the list was simply lost. Now `overflow-y: auto` with a phosphor green scrollbar, Shift+PageUp and Shift+PageDown the way a Linux console does it, and `draw()` only pins to the bottom when you were already at the bottom, so scrolling up during output no longer yanks you back down. - [x] **Pages 98 uses its own logo.** The Start button carried the site orb, `img/logo-32.png`, while the splash two seconds earlier showed a four pane flag. An operating system uses one logo. `.flag98` in `css/desktop.css` is the splash flag at button size, built from the same four colours rather than exported as a PNG, so they cannot drift and so it stays crisp at sixteen pixels. - [x] **The Command Prompt is gone**, and `css/shell.css` with it. Wrong joke, wrong room. Server Core is one tube and one prompt. - [x] **The POST reports the real machine.** `SERVER` in `js/firmware.js`. The board really does have an American Megatrends BIOS. - [x] **The badge marquee is off the home page**, with every line of the code kept and commented as deliberately unused. - [x] **Paint saves.** `File`, `Save As` puts a picture in My Pictures and it is still there next visit. `File`, `Open` browses them with thumbnails and can delete one. Quota exhaustion puts up a disk full box, which is a real condition here rather than set dressing. - [x] **`tools/check.py` now reads stylesheets.** Added because this round introduced a reference to `img/cursors/pointer.png`, which has never existed: the file is `link.png`. Nothing caught it, because the link check reads `href` and `src` out of HTML and never opens a CSS file. Cursors are the worst case of the three: a missing background is a blank area you notice, but a missing cursor image falls back to the keyword after the comma and looks completely normal, so the site quietly loses a custom pointer and nothing ever tells you. Verified that the check fires by breaking a path on purpose. ### Phase 1: the three that matter most - [x] **Sound.** The single biggest upgrade available, and it is not close. Per character typing tick in the terminal, a short POST beep, a low CRT hum under the green screen. All synthesised with Web Audio, no asset files. `webkitAudioContext` exists in Safari 6, so both Macs are fine. - Modern browsers block audio until you interact with the page, so the POST beep on a cold load gets silently swallowed in Chrome. The 2010 and 2012 Macs have no such restriction and will just play it. - Unlock the audio context on the first click or keypress. By the time you reach the terminal there has always been one. - Needs a mute control that is remembered. Sound you cannot turn off is a reason to close a tab. - [x] **A real BIOS footer instead of the skip hint.** The bottom right currently says `click or press any key to skip` in the site font, which is the one thing on that screen admitting it is a web page. Real firmware puts its own line there. Same job, no illusion break: ``` Press DEL to enter SETUP, F12 for Boot Menu, ESC to skip ``` - [x] **Make F12 work: the boot device menu.** This is a genuine gap, not a nice to have. The machine currently network boots for no stated reason. ``` Please select boot device: Hard Disk CD/DVD Drive Removable Devices > Network (IPv4) Enter Setup ``` Explains why PXE is happening, is a real screen almost nobody recreates, and reuses the menu widget that already exists. Auto selects Network after a beat if F12 is never pressed. ### Phase 2: the boot sequence - [~] **The system configuration table.** The iconic end of POST artifact: the boxed two column table listing CPU Type, Base Memory, Extended Memory, Cache, Display Type, Serial and Parallel Ports. Also the right home for the real machine details, which currently sit in a slightly invented "Probing host hardware" block. - MARKED `[~]`, NOT DONE, after a check of this file against the code. Something was built and it is good, but it is not this. The POST ends with two plain unboxed blocks, "Server Configuration" and "Your Terminal", in `js/firmware.js`. There is no box, no second column, and not one of the seven named fields exists anywhere in the tree. - The box was deliberately dropped rather than never attempted: the POST was 52 lines against a screen that holds about 50, and the overflow was taking the American Megatrends badge off the top. The packing comment in `js/firmware.js` records the trade. - The second half was genuinely undone and IS NOW RESOLVED, the other way round from how the item imagined it. The audit was right that the details had been duplicated rather than moved: "Probing host hardware" in `js/boot.js` prints Display, Renderer, Platform, Agent, Locale, Processors, Memory, Link and Clock, and the POST had grown a "Your Terminal" block saying the same thing twenty seconds earlier. - The POST block is the one that went. The probe block belongs to the network boot, which scrolls and has room for nine lines; the POST is a fixed screen that pins to the bottom, where four lines is four lines of headroom the badge needs. So each fact is said once, in the place that can afford it, and the POST is 38 lines now rather than 42. `sysReal()` is no longer called from `js/firmware.js` at all. - There is a stale comment in `css/boot.css` explaining a white-space rule as needed for "the whole drawn box of the system configuration table", describing a box nothing draws. - If it is ever built for real it costs rows the POST does not have, so it means taking lines out of the probe block first. - [x] **A memory count that races.** `Memory Test : 262144K OK` ticking upward fast. Iconic, and about ten lines. - [x] **A PXE failure that recovers.** `PXE-E51: No DHCP or proxyDHCP offers were received`, a pause, a retry, then success. Adds tension and length without implying anything is broken. - [x] **Dots during the TFTP transfer** instead of cutting straight from "Downloading NBP file..." to success. - [x] **DEL enters SETUP.** A blue and grey BIOS setup screen with menus that do nothing. Pure decoration, enormously satisfying. Optional but it is the natural partner to F12. ### Phase 3: the boot menu and loading - [x] **Remember the last choice as the default.** Real `boot.ini` has a persistent `default=`. If you always pick Server Core it should become the highlighted entry. Authentic and genuinely useful. - [x] **A different loading screen per operating system.** Right now both paths share one Windows loading bar, which is the least defensible bit of the whole chain. - Pages Web keeps the Vista bar it already has. - Pages Server Core gets a Linux init scroll instead: ``` [ OK ] Mounted /srv/photos [ OK ] Started Authelia [ OK ] Reached target Multi-User System ``` - Pages 98 gets the Windows 98 splash. Done in Phase 6. - [x] **Add Pages 98 to the menu** once Phase 6 exists. It is the middle entry, and the menu still opens on whichever one you booted last, the way `boot.ini` did. ### Phase 4: the terminal - [x] **A CRT power on animation.** The bright horizontal line that snaps open vertically when a tube wakes up. Pure CSS transform, no library, and it is the perfect transition into the terminal. - [x] **The Fallout hacking minigame.** The hex address grid with words hidden in the garbage, `likeness=3`, four attempts, duds and allowance resets in the brackets. - Offer it after three failed logins, or on a `hack` command. - It must NOT touch real authentication and must not pretend to. On a win it prints something like "impressive. Authelia is unmoved. Password?" The joke only works if it is honest. - [x] **A terminal bell on error.** Needs Phase 1. - [x] **Screen burn in on idle.** After a few minutes, the current text slowly ghosts and dims. Cheaper and more unsettling than a screensaver. ### Phase 5: the two shells Linux side: - [x] **Tab completion.** The most missed feature in any fake shell. - [x] `top` with a live updating process table - [x] `ping` with animated per packet output - [x] `man` for the commands that exist - [x] `sl` for when you typo `ls`. A steam locomotive crosses the screen. - [x] `matrix` as a screensaver you can quit - [x] `fortune` and `cowsay` Command Prompt side: **ALL OF THESE WENT WITH THE COMMAND PROMPT.** They were built, they worked, and they were deleted in round three along with the window they lived in and `css/shell.css`. Nothing in the current code registers any of them: typing `color`, `tree` or `shutdown` in Server Core answers "command not found". They stayed marked `[x]` for a while afterwards, which is how a backlog starts lying. Marked `[~]` now, meaning built and then deliberately removed, because deleting the lines would lose the record that they ever existed and leaving them as done would claim the site has three commands it does not. - [~] **`color`** that actually recolours the console. `color 0a` for green on black, `color 1f` for the classic blue. Authentic and cheap. REMOVED with the Command Prompt. - [~] `tree` with the real ASCII branch characters. REMOVED with the Command Prompt. - [~] `shutdown -s` that produces a Windows shutdown dialog. REMOVED with the Command Prompt. Pages 98 has its own Shut Down, which does work. - [x] `edit` opening the MS-DOS Editor blue screen. Big, but glorious. BUILT, on 21 August 2026, in `js/dos.js`, which is where a DOS editor actually lived rather than in the Windows console that no longer exists. Type `EDIT`, or `EDIT XANDERS.TXT`, at the DOS prompt. - The grey menu bar, the blue field, the single line frame with the filename sitting in the top edge, and the status line counting the cursor. The three colours are the real ones out of the sixteen VGA text mode had: blue is colour 1, the bars are colour 7. - It EDITS. Arrows move, typing inserts with its case intact, Enter splits a line, Backspace joins one at the left margin, Delete works forwards, Home and End and the page keys work, and the buffer scrolls when it outgrows the eighteen visible rows. - It does not save, and says so on the way out rather than pretending. The buffer is a copy, so editing `AUTOEXEC.BAT` and closing it leaves the file exactly as `TYPE` will read it back, which was verified. - No word wrap, because EDIT scrolled sideways instead and sideways scrolling is a lot of machinery for a case nobody will hit. The line stops accepting characters at 75 and the status bar says why. - The prompt uppercases what you type, because DOS did. The editor does not, because being SHOUTED BACK AT is the most obvious way a text editor could feel wrong. - TWO BUGS FOUND BY MEASURING IT RATHER THAN LOOKING AT IT, both from the same wrong `display`. The rows were styled `display: block`, and js/dos.js joins its rows with `
` the way every screen in that file does, so each row broke twice and the whole editor was double spaced: 658 pixels of screen where 22 single rows should be 337. A block also fills its container, so the blue field ran 1219px wide against a frame that stops at 78 characters, leaving the box floating in a much wider field. `display: inline-block` fixes both: the `
` supplies the single break and the element is exactly as wide as its 78 characters. I had already looked at a screenshot of this and read the black gaps as part of the frame. The numbers are what caught it. - Verified at 1280x800 and 1440x900, the two target MacBooks: 22 rows at exactly one line height each, the blue matching the frame width, and no overflow in either direction. ### Phase 6: Pages 98, the toy box The biggest item here and the one that makes the site genuinely interactive. Most of the hard part already exists: `js/win98.js` draws and drags windows. - [x] A desktop: wallpaper picked from the same manifest the Aero side cycles, icons on a grid, click to select - [x] **The selection rectangle.** Press on empty desktop and drag and a dotted band follows the mouse, selecting by intersection rather than by containment, so clipping a corner counts. Four pixels of slop before it appears, or every ordinary click flashed a one pixel rectangle. The move and up handlers live on `document`, because the mouse always ends up leaving the window mid drag and a band bound to the desktop sticks down forever the first time it does. - [x] **The icon grid wraps into a second column** when it runs out of screen, which is what Windows did. Done in `dtLayoutIcons` rather than in CSS: multi-column expresses it exactly, but only with `column-fill: auto`, and that is the one part Safari 7 does not have. The CSS version laid all nine icons out in a single row across the top. - [x] A taskbar with a Start button, window buttons and a system tray clock - [x] A Start menu, drawn from `tools/content/desktop.html` - both menus close on a click anywhere outside them, which is the part that took the fixing: bound to the desktop alone, the Start menu hung open over whatever you clicked next - [x] **Right click opens a real context menu.** Note this reverses the current behaviour: right click is suppressed on the terminal, which is correct there and wrong here. A fake context menu on a fake desktop is both more authentic and much more fun than blocking the mouse. - [x] Put existing content in windows: My Buttons reads the same table the Aero side uses, and My Pictures reads `MANIFEST.gifs`, so there is one list of each and two operating systems showing it. Webring, Guestbook and the Recycle Bin are in too. - [ ] still to move in: the media player. The WMP11 recreation is already built, but `MANIFEST.music` is empty, so a player window today would be an empty player. Worth doing once there is music. - [x] Real toys: Minesweeper first, it is the smallest and the most loved - first click is never a mine, flood fill, flags that block clicks, a counter that goes negative when you over flag, and the fixed number colours everybody knows without knowing they know them - [x] Notepad that actually types - [x] **Paint that actually draws.** Ten tools, rubber banded shapes on a scratch canvas over the committed one, a real iterative flood fill, and right button drawing in the background colour. The stroke handlers move to `document` on mousedown, so a drag that ends off the canvas ends cleanly instead of leaving the tool stuck down. - [x] **A screensaver on idle.** Starfield or flying windows, chosen when it fires, after four minutes. The hard part is not the animation: it is that the event which dismisses a saver must not also reach the desktop underneath, and killing the mousedown does not kill the click the browser builds from it. `js/saver.js` swallows the whole gesture. Preview it from Display Properties. - [x] Window minimise and maximise, not just close - maximise measures its own chrome rather than trusting `width: 100%`, which overhung the screen by the width of the border - [x] A shutdown sequence that returns you to the boot menu - [x] **The Pages 98 splash**, moved down from Phase 3. Set by an inline script in the head, not by `desktop.js`: that file is the last of ten and by the time it ran you had already seen the desktop, so the splash arrived second and the sequence read desktop, splash, desktop. - [x] **Its own font and its own icons.** MS Sans Serif from `fonts/msw98ui-*` and 32x32 icons in `img/icons98/`, both scoped to this page alone. The rest of the site is Selawik and Vista icons and the two cannot share a screen. The font scoping needs a class on `body`, because `css/chrome.css` sets the face with a universal selector and a universal selector beats an inherited value no matter where it was set. - [x] **Content no longer leaks out of its window.** The bare pages never load `css/style.css`, so they never got its `* { box-sizing: border-box }`, and every app body asking for `height: 100%` plus 8px of padding came out 16px taller than the window holding it. `.w98__raw` was also styled only in `css/shell.css`, which this page does not load either, so the window body had no clipping at all. Both fixed, and the Notepad text area is now pinned to the bottom edge rather than guessing at 100%. - [x] **Menu titles carry their own bevel.** File, Edit, Search and Help in Notepad, and the same in Paint. A departure from the original, which drew them flat until you touched them: on a page full of things that are not pressable, these should read as things that are. ### Fixed after an audit, not from the backlog Seven read-only audits went over the whole tree. What they found and what was done about it: - [x] **The boot curtain failed closed.** `css/osmenu.css` says "fail open, never fail closed", and the CSS default really is safe, but the inline head script always runs and hides the page the moment it does. The only thing that ever unhid it was `bootReveal` in `js/boot.js`. That file failing to load left the site invisible forever, and `js/bsod.js` could not catch it because `window.onerror` does not fire for a script tag that 404s. Both curtains now watch for a heartbeat. NOT a fixed deadline: the POST waits indefinitely for ENTER by design, so a timer would have snatched the BIOS screen away from somebody reading it. - [x] **Safari 7 could not sign in to the private wing at all.** `AUTH_MODE` is "api", both factors guard on `window.fetch`, and fetch arrived in Safari 10.1. On the 2010 MacBook, the stated primary target, a visitor typed a password and was told to go and edit a variable called AUTH_MODE. It now falls back to the Authelia portal, which needs no fetch because it does not collect anything itself. - [x] **The fake filesystem said where the password vault lives.** No credential, and the vault is correctly nowhere near the web, but `services.txt` named the host, said it was a file rather than a service, and named the client that opens it. The principle was worth keeping and the address was not. - [x] **Two of the three Paint fill styles were unreachable.** The options panel showed the width picker and the fill styles stacked, 90px of content in a 54px box with `overflow: hidden`, and a clipped element is not hit testable. `style` could never leave 0, so every rectangle and ellipse drew as a bare outline and two branches of the shape code were dead. Paint swapped the box contents rather than stacking them. - [x] **A lost mouseup left the airbrush spraying forever.** Measured going from 16 inked pixels to 168 with nobody touching the mouse, and nothing in the interface stopped it. Strokes now abort on window blur, and picking a tool ends whatever was still running. - [x] **The screen saver deleted the nearest flying windows on resize.** The paint loop sorts far to near, so `pop()` took the closest and largest one first, which is exactly the "crashed and recovered" look the trim was written to avoid. It now removes the furthest. - [x] **The saver went deaf after a wheel notch.** The swallow exists to eat the click that follows a mousedown. Wheel and mousemove have no tail, so opening it for them just ate the next real click. - [x] **woff was listed before woff2.** On the reasoning that Safari 7 needs the old format first. Wrong: a browser skips a format it does not know rather than giving up, so this only made every modern browser fetch the larger file. Still open from that audit, and deliberately left for a decision rather than changed unilaterally: - [x] **`background-attachment: fixed` removed from `body`** in `css/style.css`. In WebKit any element carrying it registers as a slow repaint object on the FrameView, which disables threaded scrolling for the whole document: scrolling becomes a main thread repaint of the visible area on every frame. What it bought was the sky holding still, which is invisible whenever a wallpaper is on, because `.wallpaper` is `position: fixed` and opaque and covers it. - [ ] If you want the gradient anchored again WITHOUT the cost, move the four gradients onto a `position: fixed` div at `z-index: -5`, which is the trick `.wallpaper-scrim` already uses. Only worth doing if you browse with the wallpaper switched off. - [ ] **The service inventory is public.** `SERVICES` in `js/shell.js` lists six real subdomains. They are discoverable through certificate transparency anyway once they exist, and showing off a homelab is half the point of a site like this, but it should be a decision rather than an accident. ### Phase 7: the machine misbehaving Screens you only see when something has gone wrong. These are the ones that sell the fiction hardest, because a machine that only ever succeeds is a demo, and a machine that occasionally complains is a machine. - [x] **A blue screen for the 404.** The one I would do first. The 404 page is currently the only screen on the site that breaks character: it shares the home shell and quietly appears with no boot. Give it the real thing instead, with the file you asked for as the fault: ``` A problem has been detected and Pages has been shut down to prevent damage to your computer. PAGE_NOT_FOUND Technical information: *** STOP: 0x00000404 (0x00000000, 0x00000000, 0x00000000) ``` Keep a real link back to the site on it. A joke that traps you is not a joke. - [x] **chkdsk after an unclean shutdown.** If you close the tab part way through the boot, drop a flag. If the next visit finds that flag still set, run a disk scan before the boot sequence: ``` Checking file system on C: The type of the file system is NTFS. One of your disks needs to be checked for consistency. ``` Almost nobody does this and it is exactly the detail that makes a machine feel like it has been running the whole time you were away. Clear the flag when the boot completes normally. - [x] **"Keyboard not detected. Press F1 to continue."** The oldest joke in the BIOS. Rare, random, and it must accept any key rather than literally requiring F1, or somebody on a laptop is stuck forever. - [x] **A JavaScript error should look like a crash**, not like nothing. A `window.onerror` handler that draws the blue screen would turn the worst case into the best gag on the site. Only worth it once the blue screen exists. - [ ] **A status page for the services**, once the VPS is real. Green and red lights for Nextcloud, Immich, Navidrome, mail. Fits naturally in Server Core as a `status` command and in Pages 98 as a control panel. ### The sweep, done in one pass Everything the eight lens audit turned up, worked through in order. The worklist that tracked it is gone; this is what it left behind. **Things that were misleading a visitor** - [x] Eight public pages were addressing the OWNER, not the reader. About opened with "Replace all of this with your own words." Contact said the address below was a placeholder, directly above the address. All eight rewritten in the site's own voice. - [x] The mailing list said "not actually stored yet" AFTER you handed over an address. It says so above the field now. - [x] The guestbook made a live link out of whatever you typed in the website field, `javascript:` included. Only http and https now. - [x] The 404's "Restart the machine" went to `/index.html`, which the head script suppresses for a same origin referrer, so the one control named after restarting the machine did not restart it. **Routing, so the machine is circular** - [x] Pages 98 and Restart are both in the sidebar and the site map now. Pages 98 was reachable from exactly one place: an arrow key at the boot menu inside a ten second countdown. - [x] **The countdown was booting the wrong operating system.** `osTick` called `osBoot(0)` while the screen above it said "highlighted choice", and then wrote the wrong answer to storage, so it overwrote the preference it had just ignored. Invisible with two entries. - [x] **`guest` is offered at the terminal login.** `?demo` already walked the entire private wing and was mentioned in one line of README.md. A guest account is period correct and discoverable at the moment you are stuck. **The machine remembering itself** - [x] The MAC and the SMBIOS GUID persist. The POST printed one MAC and `ifconfig` printed a different one ninety seconds later in the same session, because the shell called `sysFake` a second time. - [x] `uptime` is counted from a stored first boot instead of claiming "up 41 days" on a machine you just watched cold boot. - [x] Pages 98 keeps its wallpaper, its Notepad buffer and a Minesweeper best time. It was the only partition that remembered nothing, and `store.js` was already loaded on the page doing nothing. **What a crawler and a link preview see** - [x] `robots.txt` and `sitemap.xml`, both generated from NAV. - [x] meta description, og: and twitter: tags on every page, generated from the DESCRIPTIONS dict that already existed. The only meta tag on the whole site was the charset. - [x] The two takeover pages have the site name back in their titles. **How it feels on a 2010 laptop** - [x] The terminal's 26px and 48px bloom layers are gone. A blurred text-shadow rasterises per glyph on the CPU, and `draw()` rewrites the whole buffer on every keystroke. `.tty__glass` already does the room glow once, statically. - [x] `background-attachment: fixed` off the body. In WebKit it marks the page as a slow repaint object and disables threaded scrolling for the whole document. - [x] Focus styles, in a 98 dotted idiom. There was not one `:focus` rule in fourteen stylesheets. - [!] Capping the terminal scrollback is NOT done, deliberately. `caretSlot`, `menuAt` and the spinner's `slot` all hold indexes into `lines`, so trimming from the front points them at the wrong row. The one line fix would introduce a worse bug than it closes. **Dead weight** - [x] `DT_CELL_W`, `DT_CELL_H`, `ntPad`, `typingUntil` and the `loginLink2` branch, all gone, along with the comments that claimed they mattered. - [x] `.ogg` off the audio whitelist. Safari has never decoded Vorbis on macOS and `player.js` never called `canPlayType`, so an Ogg track appeared in the playlist and silently did nothing. **The fiction** - [x] Pages 98 has the CRT overlay. `crt.js` bailed on `logon-root`, which Pages 98 inherited from the terminal, so the monitor vanished when you booted it. - [x] chkdsk speaks in the firmware's voice. It ran BEFORE the boot menu and announced NTFS and "Windows will now check the disk", telling you the answer to a question it had not asked yet. - [x] The 98 window chrome letters itself in MS Sans Serif everywhere, not just on the desktop. The Command Prompt in the private wing was an XP console in a 98 frame in a Windows 7 font. - [x] **A broken comment in boot.css was deleting the POST font size.** One comment closed early and CSS error recovery ate `font-size: 1.05vw` with the loose prose. The BIOS screen had been rendering at an inherited 12px. - [x] The adoptable pet is on the page. `main.js` has wired it and `style.css` has animated it since they were written, and no element with the class ever existed. - [x] The front page leads with a featured wallpaper you can click to set, instead of an empty box captioned "Coming soon!". **Writing and documentation** - [x] Four blog posts, lifted from the source comments where the best writing on this site was buried: why the 404 has real exits, why the boot menu stops its clock, why the hacking game refuses to let you win, and why the machine notices you closed the tab. - [x] ASSETS.md can actually recover the wallpapers now: which archive, how many, and the seven category folders. `.gitignore` promises it can. - [x] `tools/layout.py` generates the README layout block from the banner comment of every file. It was hand maintained and had drifted to missing 13 of 29 JS files. - [x] ROADMAP.md lost 80 lines describing a different project, called Nexus-7, under a `site/` directory that never existed. **And the fourth partition** - [x] **`js/dos.js`.** Press F12 at the POST and pick Hard Disk. Hard Disk, CD/DVD and Removable Devices were three drawn, arrow selectable, completely dead doors. One of them now holds the old install nobody could face moving off: a C: prompt, eleven files, about a dozen commands, and an AUTOEXEC.BAT that says "Xanders. Do not reformat this one." EXIT hands back to the firmware and the network boot carries on, so it is a detour rather than a dead end. It is deliberately almost empty. The whole value of this partition is that there is nothing on it. ### Round two: the first visit, and a review of the newest code A second set of lenses, deliberately aimed away from what the first audit covered: a stranger's first ten minutes, an adversarial read of the newest files, the empty rooms, what the machinery now makes cheap, and what rots. **What a stranger actually hits** - [x] **The Private Wing link made you watch the whole boot again, in a new tab.** The sidebar linked `private.html` with no query string, so `js/boot.js` ran the entire chain a second time: POST, the indefinite wait for ENTER, the splash, the eleven second network boot, in a `target="_blank"` tab with no Back button. The operating system menu had always linked it as `?booted`. Now the sidebar does too. `boot.js` had already written the rule this broke: two network boots back to back is where a nice thing becomes a toll gate. - [x] **F12 and DEL were the only doors to Setup, the boot menu and the whole MS-DOS partition, and neither key reaches the page on the target machine.** On a MacBook the key marked `delete` is Backspace, and F12 is a volume key unless standard function keys are switched on. On Windows, F12 opens DevTools and `preventDefault` cannot stop it. So the best drawn screen in the chain and the newest partition on the box were both unreachable on the 2010 laptop this site exists for. The three words in the POST footer are clickable now, with the wording untouched. - [x] **The one button on the front page pointed at the emptiest page.** "enter the world" went to About, which opens by admitting it has not been written. Same failure the featured box was fixed for, one click downstream. It boots Pages 98 now. - [x] **The longest screen had no skip hint.** The network boot has always been skippable, but "ESC to skip" only existed on the four second splash, which is hidden the moment the eleven second scroll starts. - [x] **QUICK BOOT.** A bookmark or a typed URL sends no referrer, so the head script cannot tell a returning visitor from a stranger and played the full twenty one seconds every time. The machine now remembers that it has booted all the way through once, and every later direct arrival runs at the brisk pace without stopping for ENTER. Set on COMPLETION, so a visitor who closes the tab halfway still gets the full thing. Never applies to `?boot`, so Restart, Shut Down and `reboot` always give the whole sequence. AMI called this Quick Boot and had it on by default. **Real defects in the newest code** - [x] **firmware.js kept its keyboard while dos.js ran.** Both draw into the same element, and both were listening. Pressing ESC anywhere in the DOS session ran the FIRMWARE's ESC branch: it cleared the screen and started the network boot while dos.js carried on writing its prompt underneath, and the two took turns redrawing. Verified before the fix by picking Hard Disk and pressing ESC 300ms later: the machine network booted. The firmware now hands the keyboard over and keeps the handle so it can stop the partition before taking the screen back. - [x] **Maximise then restore threw a window off centre.** `win98Open` centres with `margin-left` and `margin-top` only, so the CSSOM will not serialise the shorthand and `win.style.margin` reads back as the empty string. That empty string was saved and assigned back on restore, which CLEARS all four longhands rather than leaving them alone, so the centring margin was deleted and the window jumped half its own width. It saves the two longhands it actually sets now. - [x] **The machine still contradicted itself about its own address.** Pinning the MAC fixed the half visible on two screens; the IP, gateway and subnet were still re-invented on every call, from a randomly chosen RFC1918 block, across five call sites in one session. The whole network identity is now held for the life of the page. Deliberately NOT persisted: a card keeps its MAC, a DHCP lease does not, so the address changes between visits and stays put within one. - [x] **The featured wallpaper did not restart the cycler**, so clicking it 44 seconds into the 45 second cycle set your pick and had it replaced a second later. Every other control already called `startTimer()`. - [x] **Two Notepads shared one storage key.** Both the icon and the Start menu open it, each window loaded the note on open and wrote its own box back on close, so closing a stale window overwrote what you had just typed in the other. Notepad is single instance now. One finding was a false positive: the review read `js/firmware.js` while the clickable footer was half written and reported the spans as unwired. They were wired a few minutes later and verified in a browser. ### The rest of round two **The tools, where one of them was a publishing bug** - [x] **`blog.py` published the feed before the page existed.** It wrote the fragment and `feed.xml`, but `pages/blog.html` was not rebuilt until `pages.py` ran afterwards. So one command pushed RSS to subscribers while the post was not on the site, and the guid anchor landed on nothing. It now writes the fragment, runs `pages.py`, and writes the feed last, in one command, because a reminder is a thing you follow four times and forget once. - [x] **The two generated files in `tools/content/` say so now**, emitted by their own generators rather than pasted in, so a rebuild cannot lose the warning. That folder is the one the README calls editable, which is exactly where an unmarked generated file does the most damage. - [x] **Every tool computes the repo root from its own location.** Eleven of them hardcoded the full drive path and three did not, so moving or copying the folder would have left half the toolchain writing to the original. Verified by running them from another drive. - [x] **`tools/check.py`**, the checker the README asked for and never got. It folds in `checkhtml.py` and adds what was missing: the no-dashes rule, internal links, fragments newer than the pages built from them, orphan js and css, and external hosts with `--links`. Proved by planting an em dash, a broken link and an orphan file: it caught all three and exited 1. Every check is there because that exact thing has broken here once. - [x] **The build order is written down in one place**, at the top of README.md, rather than scattered across five sections and four docstrings. **What the machine already made cheap** - [x] **Paint: File, Set As Wallpaper.** Two subsystems that had never met were already on the same page: a canvas with an undo buffer, and a wallpaper with a storage key. The picture is kept whole as a data URL, because there is nowhere to put a file. It survives a reload, and New Wallpaper still goes back to a photograph. - [x] **DEFRAG.EXE.** `DIR` had always listed it at 78,016 bytes and typing it got "Bad command or file name" about a file the machine had just named. The block grid animates by writing back into the printed buffer by index, which is how a console animates in place. - [x] **`pages98` in the terminal.** Phase 8 asks that every operating system reach the others, and this was the missing edge: Pages 98 had icons for both of the others, Server Core had `vista` and `reboot` and no way to the desktop. **The empty rooms, mostly reframed rather than filled** - [x] **Web Games points at the four games that already ship.** The page was an argument that Flash is dead followed by "Games go here", while the machine one partition over had Minesweeper, Paint, the screen saver and DEFRAG. Nothing needed writing, only linking. - [x] **The guestbook takes entries by email.** The form validated your input and then told you it had thrown it away, which was the only interaction on the site a visitor was invited to perform. The page's own third option was the answer, and it is what most people did in 1999. `js/guestbook.js` went with it. - [x] **Lost Media has three entries**, written from material already in ASSETS.md: the GeoCities shutdown and the 31,119 buttons, GifCities and why redundancy beat quality, and Flash having a death with a date on it. It is the only empty room backed by an interest the site states twice. - [x] **The webring lost its dummy widget**, which read "site 0 of 0" with dead links and was the one block making an otherwise finished essay look unfinished. - [x] **Gallery says what it is for** and points at the rooms that are full, instead of making the same promise as Wallpapers with nothing behind it. - [x] **Music is honest about why it is empty**: everything worth playing belongs to somebody else. It was a page of instructions to its owner. - [x] **The under construction gif is on one page, not eight.** At that density it stopped being the 1999 joke about a page being alive and became a status report, usually attached to a caption naming the thing that was missing. - [x] **Two pages stopped claiming to be constantly updated** when neither had ever held an item, on a site that invites you to check. **How the site presents itself** - [x] **The nav leads with the finished rooms.** Wallpapers and Blog now come before Who Am I, Projects and Gallery. The emptiness was fine; the order made a site with several finished rooms read as unfinished. - [x] **The front page explains the machine.** Grepping every public page for "F12", "MS-DOS" or "partition" returned nothing at all, so the one thing a visitor would tell a friend about was invisible. There is a box naming all four partitions, the guest login and the F12 key. - [x] **The updates page is current.** Its first paragraph offers to let you check whether the site really is being updated, and it had broken that promise within a day. ### Phase 8: connective tissue - [x] Every operating system has a way back to the boot menu. The terminal has `reboot`, Pages 98 has Shut Down, and Pages Web now has Restart in the sidebar pointing at `index.html?boot`. The 404's "Restart the machine" carries the flag too. - [ ] A consistent idea of "restart" that replays the whole chain. - [ ] Decide whether the boot plays on a deep link from a search engine. Current rule: only when you arrive at the home page from outside the site. Landing on the Gallery from Google never boots, which is probably right. - [ ] Keep the em dash rule. Still zero across the whole tree. --- ## Constraints that apply to all of it - **Safari 7 on Mavericks is the floor.** No CSS variables, no grid, no flexbox gap, no `inset`, no `clamp()`, no `:has()`, no arrow functions, no `let` or `const`, no template literals, no `padStart`, no `forEach` on a NodeList. Prefixes on every gradient, transform and animation. See `ROADMAP.md`. - **Desktop only, fixed 980px.** No mobile layout, on purpose. - **No em dashes and no en dashes anywhere**, including HTML entities. - **Fail open.** Anything decorative that gates the site must default to off when JavaScript is unavailable. `#bootlayer` already does this: it is `display: none` until a script says otherwise, so a crawler or a browser with JavaScript off sees the site immediately rather than a black rectangle nailed over it forever. - **Reduced motion means less movement, not less content.** Text that types is content arriving. Things that sweep, spin and slide are movement. Stop the second kind and keep the first. - **Nothing in the browser is a security boundary.** Every line of every file here ships to the visitor. The fake shell, the fake filesystem and the hidden route to the private wing are all flavour. Authelia is the only thing actually guarding anything.