feat: Cloudflare Turnstile on auth, CSP fixes, admin/SEO/analytics additions

Turnstile bot protection (sign-in, sign-up, password-reset):
- Register Better Auth's captcha plugin with the cloudflare-turnstile
  provider; endpoints listed explicitly rather than relying on defaults.
  /reset-password is intentionally excluded — it is reached only via a
  single-use emailed token.
- Add an explicit-render Turnstile widget component. Tokens are single-use,
  so each form resets the challenge after a failed submit; submit stays
  disabled until a token is held.
- Read the site key server-side and pass it down as a prop, so rotating it
  does not require a rebuild.
- Fail fast in production when TURNSTILE_SECRET_KEY is missing, and when a
  secret is set without a site key (that combination would demand a token
  no form can produce, locking every user out).
- Pass a throwaway secret during `next build` in the Dockerfile, mirroring
  the existing BETTER_AUTH_SECRET treatment, so image builds don't need it.

CSP fixes in middleware (these blocked Turnstile entirely):
- Add frame-src for challenges.cloudflare.com. Without it the widget's
  iframe fell back to default-src 'self' and was blocked outright.
- Allow 'unsafe-eval' and websockets in DEVELOPMENT only. `next dev`
  compiles with eval(), so the strict policy threw EvalError and killed
  hydration — no client JS ran at all, which also meant form submit
  handlers never fired. Production policy is unchanged and still strict.

Also included (concurrent work in the tree):
- Admin organizations pages and lib/admin/orgs.
- Episode moderation migration, SEO metadata (sitemap, robots, JSON-LD,
  OG/Twitter images, manifest), Umami analytics, not-found page.

Local dev database: docker-compose.dev.yml provisions Postgres 18 on port
5443 (5432-5442 are in use by other local projects).

Note: `npx tsc --noEmit` currently fails in app/(app)/team/page.tsx — an
`invitations` prop the component does not accept. This predates the commit
and will fail `next build` until fixed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

lib/episodes/status.ts

@@ -41,6 +41,36 @@ export async function finishJob(
   });
 }
 
+/**
+ * Generic, user-safe copy for a failed generation.
+ *
+ * Episode.errorMessage is rendered straight to the end user (see
+ * app/(app)/episodes/[id]/page.tsx) and streamed over SSE, so it must never
+ * carry raw pipeline text: ffmpeg stderr leaks internal binary paths, and an
+ * ElevenLabs/OpenAI response body leaks our upstream quota and billing state.
+ */
+export const GENERIC_FAILURE =
+  "Generation failed. Our team has been notified — please try again shortly.";
+
+/**
+ * Record a terminal pipeline failure.
+ *
+ * The RAW error is preserved on GenerationJob.error, which is only ever surfaced
+ * in the admin UI (lib/admin/ops.ts), so debuggability is unchanged. The episode
+ * itself gets the generic message above.
+ */
+export async function failEpisode(
+  episodeId: string,
+  rawError: string,
+  userMessage: string = GENERIC_FAILURE
+): Promise<void> {
+  await prisma.generationJob.updateMany({
+    where: { episodeId, status: { in: ["queued", "running"] } },
+    data: { status: "failed", error: rawError, finishedAt: new Date() },
+  });
+  await setEpisodeStatus(episodeId, "FAILED", { errorMessage: userMessage });
+}
+
 /** Terminal episode states — used by the UI/SSE to stop polling. */
 export function isTerminal(status: EpisodeStatus): boolean {
   return status === "READY" || status === "FAILED";