fix: repair what a full verification pass found

.gitignore

@@ -47,3 +47,11 @@ Thumbs.db
 # typescript
 *.tsbuildinfo
 next-env.d.ts
+
+#ai agent
+.claude/
+.cursor/
+.codex/
+
+# MCP tool artefacts — console logs and page snapshots from browser automation
+.playwright-mcp/

CLAUDE.md

@@ -26,15 +26,19 @@ than none. Whoever asked you to make this change will tell you your scope.
 
 ## `TODO(name), Day n:` markers are assignments
 
-Every stub names the person who owned it and the day it was due during the original build. Implement
-the ones you were asked to implement; leaving the rest alone is correct behaviour, not incompleteness.
-Deleting a marker while implementing around it destroys the only record of who owes what.
-
-## Some tests are red on purpose
-
-`compress-dom`, `bisect`, and `evaluatePredicate` ship with failing tests. They were written first, as
-specifications. **Never** make them pass by weakening an assertion, adding `.skip`, or deleting a case:
-that converts a specification into a lie, silently.
+There are none left — every stub from the original build has been implemented. The convention is
+documented here because the rule outlives the markers: a marker names the person who owned the work and
+the day it was due, so **deleting one while implementing around it destroys the only record of who owes
+what.** If you add a marker, name yourself in it. If you find one, either implement it or leave it
+exactly where it is.
+
+## The tests are green, and two suites must stay honest
+
+All 285 tests pass. That is worth stating because of how some of them got there: the `compress-dom`,
+`bisect` and `evaluatePredicate` suites were written first, as specifications, and were red for as long
+as it took the implementations to satisfy them. **Never** make a test in those suites pass by weakening
+an assertion, adding `.skip`, or deleting a case: that converts a specification into a lie, silently. A
+red test there means the implementation is wrong.
 
 `no-eval.test.ts` greps the source and fails if `eval(` or `new Function` appears anywhere. It is a
 security boundary, not a lint rule.

CONTRIBUTING.md

@@ -28,6 +28,20 @@ order of fidelity:
 You don't need a real agent for most work. You do need one before claiming a tool works — see
 [Testing a tool](#testing-a-tool).
 
+`traces/` and `bugbait/` install separately — neither is a workspace member, and there is no lockfile at
+the repo root. `npm ci` works in both, and `npm audit` reports zero vulnerabilities in both; if either
+stops being true, that's a change to fix rather than to note. Two entries in those manifests exist for
+reasons JSON has no room to state:
+
+- **`overrides: { "postcss": "^8.5.26" }`** in both. The advisory is against a copy of postcss nested
+  under `next`, not the one either app depends on directly, so bumping the direct dependency does
+  nothing. The alternative `npm audit fix --force` offers is a Next major, which is a much larger change
+  than the one-line fix the advisory actually needs.
+- **`vitest` on 4.x**, not the 2.x this was built against. Two advisories, one of them critical, are
+  fixed only in 4.x. The suite needed no changes — but the config had to become `vitest.config.mts`,
+  since `.ts` in a package without `"type": "module"` is loaded as CommonJS and Vite's native config
+  loader is about to stop tolerating that.
+
 ---
 
 ## Three rules that aren't negotiable

README.md

@@ -57,11 +57,11 @@ Agent: find_element({ text: "Pay" })
 Agent: bisect({ selector: "#pay-submit",
                 predicate: { kind: "propertyEquals", property: "disabled", equals: true },
                 from: 0, to: 47200 })
-     → firstTrue: 28_450ms   (6 iterations, 1.4s, ±250ms)
+     → firstTrue: 28_577ms   (10 iterations, 1.4s, ±250ms)
 
-        ← the playhead visibly jumps six times while this runs
+        ← the playhead visibly jumps ten times while this runs
 
-Agent: read_dom_at({ timestamp: 28450, scope: "#checkout" })
+Agent: read_dom_at({ timestamp: 28577, scope: "#checkout" })
      → form#checkout [visible]
          input[name=email] value="ana@..." [valid]
          select[name=province] [empty options: 0]
@@ -97,12 +97,12 @@ to an agent through `document.modelContext.registerTool()`. Lots of things *coul
 that would work just as well as a REST API. Traces is not one of them, and the reasons are worth
 stating precisely:
 
-**1. The answer doesn't exist until the page computes it.** `read_dom_at(28450)` isn't a lookup. The
+**1. The answer doesn't exist until the page computes it.** `read_dom_at(28577)` isn't a lookup. The
 page must replay the mutation stream to that instant and reconstruct the DOM. A server holding the
 same recording file cannot answer the question without becoming a replay engine itself.
 
 **2. `bisect` doesn't fetch — it *programs* the page.** The agent sends a predicate and the page runs
-a binary search across the replay timeline, six probes deep. The agent isn't retrieving data; it's
+a binary search across the replay timeline, ten probes deep. The agent isn't retrieving data; it's
 handing the page an algorithm to execute over time. There is no request/response shape for that.
 
 **3. The human is inside the agent's loop, not watching it.** `ask_human_visual` doesn't resolve until
@@ -140,13 +140,16 @@ bisect({
   predicate: { kind: "optionCount", equals: 0 },
   from: 0, to: 47200
 })
-// → { firstTrue: 12_600, iterations: 6, elapsedMs: 1_180, precisionMs: 250,
-//     trace: [ { atMs: 23600, result: true }, { atMs: 11800, result: false }, ... ] }
+// → { firstTrue: 12_721, lastFalse: 12_537, iterations: 10, elapsedMs: 1_180, precisionMs: 250,
+//     trace: [ { atMs: 0, result: false }, { atMs: 47200, result: true },
+//              { atMs: 23600, result: true }, { atMs: 11800, result: false }, ... ] }
 ```
 
-Six probes instead of forty-seven calls. The `trace` isn't decoration either — the UI animates it, so
-a human watches the playhead jump six times and *sees* the agent's reasoning as motion on the
-timeline.
+Ten probes instead of forty-seven calls: the two boundary probes that bracket the window, then eight
+halvings to close it to 250 ms. The count is the same wherever the transition sits — that is what a
+binary search buys. (`elapsedMs` is the one number above that depends on the machine; the probe count
+does not.) The `trace` isn't decoration either — the UI animates it, so a human watches the playhead
+jump ten times and *sees* the agent's reasoning as motion on the timeline.
 
 Predicates are a **closed set of seven structured shapes**, never a string, and never evaluated. That's
 a security property (see [threat model](docs/threat-model.md)) but also a usability one: models fill in
@@ -368,13 +371,18 @@ Full analysis, including what is deliberately **out of scope**:
 | Dependency | Licence | Why |
 |---|---|---|
 | [rrweb](https://github.com/rrweb-io/rrweb) + rrweb-player | MIT | records and replays DOM mutations; the only mature option that reconstructs a real DOM rather than pixels |
-| [Next.js](https://nextjs.org) | MIT | static export, trivial Vercel deploy, response headers for the origin trial token |
+| [Next.js](https://nextjs.org) | MIT | prerendered pages, trivial Vercel deploy, response headers for the origin trial token |
 | TypeScript | Apache-2.0 | strict mode; tool schemas and domain types stay honest |
 | [Tailwind CSS](https://tailwindcss.com) | MIT | dense instrument UI without a component library |
 | [zustand](https://github.com/pmndrs/zustand) | MIT | state readable and writable *from outside React* — see below |
+| [IBM Plex Sans + Plex Mono](https://github.com/IBM/plex) | OFL-1.1 | one superfamily, so a mono timestamp and a sans label share a line without a step in it — `src/app/fonts.ts` |
 | [vitest](https://vitest.dev) | MIT | the pure modules are tested; the UI is not |
 
-Every dependency is permissively licensed and compatible with MIT redistribution. We deliberately
+Every dependency is permissively licensed and compatible with MIT redistribution. The fonts are the one
+thing not in `package.json`: `next/font` fetches them during `next build` and serves them from this
+origin, so the deployed bundle carries the woff2 files under OFL-1.1 — the licence text and the reserved
+font names live with [the upstream project](https://github.com/IBM/plex/blob/master/LICENSE.txt), and
+nothing here modifies or renames a face. We deliberately
 avoided three tempting libraries: **DuckDB-Wasm** (it's the flagship example in the challenge's own
 materials — using it would weaken the originality of the entry), **ffmpeg.wasm** (LGPL/GPL build
 ambiguity we didn't want in an MIT repo), and **HyperFormula** (AGPL-3.0).

bugbait/next.config.mjs

@@ -1,3 +1,5 @@
+import { fileURLToPath } from 'node:url'
+
 /**
  * bugbait — the app whose bugs become the recordings Traces investigates.
  *
@@ -8,6 +10,10 @@
  */
 const nextConfig = {
   reactStrictMode: true,
+
+  // Same reason as in traces/next.config.mjs: this folder installs on its own, so the tracing root is
+  // this folder, not whatever ancestor happens to contain a lockfile.
+  outputFileTracingRoot: fileURLToPath(new URL('.', import.meta.url)),
 }
 
 export default nextConfig

bugbait/package.json

@@ -27,8 +27,11 @@
     "eslint": "^9.0.0",
     "eslint-config-next": "^15.5.0",
     "playwright-core": "^1.62.1",
-    "postcss": "^8.4.47",
+    "postcss": "^8.5.26",
     "tailwindcss": "^3.4.14",
     "typescript": "^5.6.0"
+  },
+  "overrides": {
+    "postcss": "^8.5.26"
   }
 }

bugbait/scripts/record-fixtures.mjs

@@ -1,8 +1,6 @@
 /**
  * Regenerates the sample recordings in `traces/public/recordings/` by driving a real browser.
  *
- * Owner: Vicko.
- *
  * **These are scripted fixtures, not human sessions,** and the difference is worth being precise about.
  * Everything in the resulting file is real: a real Chrome, the real bugbait build, the real
  * `rrweb.record()`, real HTTP to the real stub endpoints, real layout and real CSS. What is synthetic is

bugbait/src/app/api/cities/route.ts

@@ -4,8 +4,6 @@ import { LATENCY_MS } from '@/lib/bugs'
 /**
  * `GET /api/cities` — always succeeds, always returns cities.
  *
- * Owner: Vicko.
- *
  * There is no scenario branch here, and that is the interesting part. In the `race` scenario this
  * endpoint behaves perfectly: 200, a full list, in under a second. The bug is entirely on the client,
  * which reads the list once during its first render — before this response exists — and never looks

bugbait/src/app/api/provinces/route.ts

@@ -4,8 +4,6 @@ import { isBug, LATENCY_MS, SCENARIO_HEADER } from '@/lib/bugs'
 /**
  * `GET /api/provinces` — the request at the centre of the primary demo bug.
  *
- * Owner: Vicko.
- *
  * When the `empty-province` scenario is armed this returns **200 with `[]`**, and the distinction from
  * a 500 is the whole point. An error status shows up red in a network tab and the investigation takes
  * four seconds; a success that happens to be empty looks like nothing at all, and only the *shape* of

bugbait/src/app/checkout/page.tsx

@@ -15,8 +15,6 @@ import { RecorderPanel } from '@/components/recorder-panel'
 /**
  * The checkout form. Every bug lives here.
  *
- * Owner: Vicko.
- *
  * The form has to be genuinely ordinary — name, address, province, city, postcode, card, pay — because
  * the investigation is only interesting if the page looks like something that would ship. A page with
  * one dropdown and a broken button demonstrates nothing; the agent has to actually find the failing
@@ -34,7 +32,7 @@ import { RecorderPanel } from '@/components/recorder-panel'
  *      rendered markup. The recording captures the DOM; anything written there hands the agent the
  *      answer and makes the demo a re-enactment.
  *
- * Day 5 (vicko) — implemented. Three notes on how each bug is kept invisible in the finished DOM,
+ * Three notes on how each bug is kept invisible in the finished DOM,
  * because that property is the argument of the whole project and the easiest thing to break by accident:
  *
  *   - **empty-province.** `/api/provinces` is requested the moment this page mounts and resolves 1.5s