fix(content): detect explicitly-rendered Turnstile managed challenge (TPS /InternalCaptcha)

plan/expurge-progress.md

@@ -99,6 +99,7 @@ superseded by the sidebar architecture — see `sidebar-nav.md` for the current
 - ~25 verified people-search brokers in brokers.json (all channels personally verified, trust bits stamped)
 - Pre-launch verify: CCPA template legal language; DROP registry cross-reference (Q-010)
 - CI schema validator: rejects malformed records, enforces trust-bit hygiene (contributed records must be `trust: unverified`)
+- **Per-broker challenge-resolve gate** (onboarding checklist item): for each new broker, confirm its bot-gate **navigates away on solve** (like TPS `/InternalCaptcha`). `detectChallenge()`'s Turnstile-script signal is generic and manifest-bounded (it only runs where `content_scripts.matches` injects), but resolve-safety is proven on TPS **only (n=1)**. If a broker instead resolves **inline** — results swap in place, URL unchanged, the `challenges.cloudflare.com/turnstile` `<script>` persists — the detector would strand the challenge view over real results → forced Skip → missed hit. Before enabling such a broker, add a resolve signal (URL-path check or solved-token) or an option-2 per-broker `challenge` hint, plus a challenge fixture in `classify.test.ts`. (Origin: `fix/challenge-detection-managed`, 2026-07-03; the human-in-the-loop + Skip is the backstop that keeps a miss recoverable, not silent.)
 - Optional stamp helper: `verify <broker-id> <channel>` CLI sets last_checked / verified_by / trust
 - Full run on real brokers, bugs fixed
 - AMO submission prep: screenshots, description, privacy notice, data-practices declaration
@@ -125,6 +126,7 @@ superseded by the sidebar architecture — see `sidebar-nav.md` for the current
 | `src/background/index.ts` | AKA parsing splits on first space only — "Mary Jane Smith" gives first="Mary", last="Jane Smith"; smarter parsing (e.g. last-space split) deferred |
 | `src/options/index.ts` | Settings section has no import JSON (export only); import deferred to M8 alongside persistence opt-ins |
 | `src/background/index.ts` | `webNavigation` is declared in manifest permissions but `browser.webNavigation.onErrorOccurred` is not yet wired — add when M9 broker set makes load-error detection meaningful |
+| `src/content/classify.ts` | Turnstile-script detection assumes solve **navigates away** (proven on TPS only). A broker that resolves the gate **inline** would strand the challenge view — see the M9 per-broker challenge-resolve gate before onboarding one |
 
 ---
 

src/content/classify.test.ts

@@ -60,4 +60,31 @@ describe('detectChallenge', () => {
     addEl(tag, attr, value);
     expect(detectChallenge()).toBe(true);
   });
+
+  // Explicitly-rendered Turnstile (turnstile.render(...)) — the TruePeopleSearch /InternalCaptcha
+  // managed-challenge shape: no .cf-turnstile container, widget iframes are about:blank, but the
+  // Turnstile API script is in the top document. That script is the only reliable signal.
+  it('explicitly-rendered Turnstile managed challenge (API script, no container) → true', () => {
+    addEl('script', 'src', 'https://challenges.cloudflare.com/turnstile/v0/api.js');
+    expect(detectChallenge()).toBe(true);
+  });
+
+  // A real solved Turnstile page carries the api.js script too. The `!turnstile` guard must route
+  // the container case to the solved/unsolved logic above, so the new script branch does NOT
+  // re-detect a solved container-Turnstile just because the script is present.
+  it('solved container-Turnstile with the API script present → false (!turnstile guard)', () => {
+    document.body.innerHTML =
+      '<div class="cf-turnstile"></div><input name="cf-turnstile-response" value="tok">';
+    addEl('script', 'src', 'https://challenges.cloudflare.com/turnstile/v0/api.js');
+    expect(detectChallenge()).toBe(false);
+  });
+
+  // Trap: the /InternalCaptcha page also hosts a cf.clym-widget.net iframe (Clym consent widget,
+  // not Cloudflare). Keying off a bare "cf"/"cloudflare" substring would false-positive on it — the
+  // selector must match the specific challenges.cloudflare.com host only.
+  it('Clym consent widget only (cf.clym-widget.net, no Cloudflare challenge script) → false', () => {
+    addEl('iframe', 'src', 'https://cf.clym-widget.net/latest/api-bridge/?instance=us6.clym.io');
+    addEl('script', 'src', 'https://cf.clym-widget.net/latest/loader.js');
+    expect(detectChallenge()).toBe(false);
+  });
 });

src/content/classify.ts

@@ -24,6 +24,16 @@ export function detectChallenge(): boolean {
     return true; // standalone CF iframe (non-Turnstile challenge)
   }
 
+  // Explicitly-rendered Turnstile (turnstile.render(...)): no .cf-turnstile container and the widget
+  // iframes are about:blank, so neither branch above matches — but the Turnstile API script is in the
+  // top document. This is the TruePeopleSearch /InternalCaptcha managed-challenge shape. Match the
+  // specific host ONLY: a co-resident cf.clym-widget.net iframe (Clym consent, not Cloudflare) must
+  // not match a bare "cf"/"cloudflare" substring. The `!turnstile` guard defers the container case to
+  // the block above — a solved container-Turnstile still carries this script and must stay resolved.
+  if (!turnstile && document.querySelector('script[src*="challenges.cloudflare.com/turnstile"]')) {
+    return true;
+  }
+
   // Other embedded CAPTCHA widgets.
   return [
     'iframe[src*="hcaptcha.com"]',