Initial build: typo, a typing test built from your own code

No backend, no build step -- static HTML/CSS/JS, themed identically to
flit/wisp (same CSS vars, same "> " title prefix, same dark/mono/green
aesthetic). Snippets are real, shipped code pulled from six repos in
this family (flit, wisp, keep, stash, delve-term, goonk) rather than
generic pangrams.

Stats: WPM = (correct chars / 5) / minutes elapsed, timing from first
keystroke not page load. Accuracy = correct keystrokes / total
keystrokes tracked at the keydown level (comparing the pressed key
against the exact next expected character before it's inserted), so
backspaced-over mistakes still count against it rather than only
diffing the final text. Keys/sec updates live via a 200ms tick.
Tabs in source are rendered as two spaces so players never need to
press Tab -- sidesteps a browser default-Tab-moves-focus fight
entirely rather than fighting it with preventDefault.

Verified via simulated typing sessions (puppeteer), not just static
review -- both an artificially fast pass (confirms the math is
internally consistent: 308 correct/310 total keystrokes -> 99%,
matches displayed) and a realistic ~60wpm-paced pass (confirms no
unit-conversion bug that only a slow, deliberate typist would hit
tests-at-lightspeed wouldn't surface). That process caught two real
bugs before they'd have hit a real player:

1. A stray `tabindex="0"` on the code display let the browser's
   native click-to-focus behavior steal focus away from the hidden
   textarea before our own focus() call ran -- typing silently did
   nothing. Fixed by removing the tabindex and switching to a
   mousedown+preventDefault handler instead of click.
2. The help modal's CSS set `display: flex` unconditionally, which
   (author CSS always beats the UA stylesheet) overrode the `hidden`
   attribute's default `display: none` -- the modal was invisibly
   covering the entire viewport and intercepting every click, at all
   times, whether or not it had been opened. Fixed with an explicit
   `.modal-overlay:not([hidden])` rule.

Also verified the actual Docker image builds and serves correctly
(nginx, no build stage needed since there's nothing to compile).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

README.md

@@ -0,0 +1,54 @@
+# typo
+
+A typing test built from your own code. Every snippet is real, shipped
+code pulled from real repos — flit, wisp, keep, stash, delve-term, goonk
+itself — not generic pangrams or Lorem Ipsum.
+
+No backend, no build step — plain HTML/CSS/JS served statically, with
+snippets baked into `snippets.json`.
+
+## Stats
+
+- **WPM** — (correct characters / 5) / minutes elapsed, the standard
+  typing-test formula. Timing starts on the first keystroke, not page load.
+- **Accuracy** — correct keystrokes / total keystrokes, including ones you
+  backspaced over. Tracked at the keydown level (comparing the pressed key
+  against the exact next expected character), not just diffed after the
+  fact.
+- **Keys/sec** — total keystrokes / elapsed seconds, live.
+
+Mistakes are highlighted in place rather than blocking progress — you can
+backspace to fix them or keep going, same as most real typing tests.
+Tabs in the original source are shown as two spaces so you never need to
+press Tab.
+
+## Adding snippets
+
+Edit `snippets.json` — each entry is `{ repo, file, lang, code }`. Keep
+snippets short (10-20 lines), self-contained, and free of exotic
+Unicode in comments (em dashes etc. are typeable but annoying) if you
+want them pleasant to type.
+
+## Local development
+
+Any static file server works, e.g.:
+
+```bash
+python3 -m http.server 8850
+```
+
+## Status
+
+Built and verified via simulated typing sessions (not just static
+review): focus handling, live character-by-character highlighting,
+newline handling for multi-line snippets, and the WPM/accuracy/KPS math
+all confirmed against both artificially fast and realistic human-speed
+input. Caught and fixed two real bugs in the process — a focus-stealing
+issue caused by a stray `tabindex` on the code display, and a modal
+overlay that was invisibly covering the whole page and intercepting
+every click because its CSS unconditionally overrode the `hidden`
+attribute.
+
+## License
+
+MIT.