feat(preview): SSE-streamed generation, no CF 100s edge cap

Architectural fix for "spec_too_large" / preview_timeout — the sync
endpoint had to fit the whole model run into Cloudflare's ~100s edge
window, which made the system fragile against any prompt that produced
a verbose spec. The new streaming path pipes Anthropic's token deltas
as Server-Sent Events; every chunk resets CF's idle timer and a 15s
keepalive comment guarantees activity even during slow first-token
windows.

@bmm/llm: new streamSpecFromAnthropic() exposes the SDK's .stream()
flow with the same typed-error contract as generateSpec — same
SpecTruncatedError / SpecValidationError / SpecTimeoutError raised from
the relevant moment.

API: POST /v1/servers/preview/stream returns text/event-stream with
events 'text' (deltas), 'spec' (final success payload, same shape as
the sync endpoint), 'error' (typed). Anthropic-only — GLM/hobby falls
back to the sync route via 409 streaming_unavailable.

Frontend: apiSseStream() handles the POST + ReadableStream + SSE
parser. The wizard's analyze() prefers the stream and only uses the
sync endpoint on the explicit 409 fallback.

nginx (api.buildmymcpserver.com): the /v1/builds/ location block (which
already had proxy_buffering off + 600s read timeout for the WS build
stream) now also matches /v1/servers/preview/stream so the SSE
response isn't buffered.

packages/llm/src/index.ts

@@ -295,6 +295,79 @@ async function generateWithAnthropic(
   return { spec: parsed.data, source: 'claude' };
 }
 
+// ──────────────────────────────────────────────────────────────────────────
+// Streaming generation (Anthropic only)
+// ──────────────────────────────────────────────────────────────────────────
+
+export interface StreamHandlers {
+  /** Called for each text delta emitted by the model. */
+  onText: (text: string) => void;
+  /** Called once when the stream completes successfully with the final spec. */
+  onSpec: (result: GenerationResult) => void;
+  /** Called once on any terminal error (timeout, truncation, validation). */
+  onError: (err: Error) => void;
+}
+
+/**
+ * Stream a spec from Anthropic, piping text deltas to a handler and finally
+ * surfacing the parsed/validated spec or the relevant typed error.
+ *
+ * Why streaming for /preview: Cloudflare's edge timeout is ~100s on the Free
+ * tier and our previous sync call could approach that for ambitious prompts.
+ * With streaming the TCP connection writes bytes from the first model token,
+ * which keeps CF (and nginx) from cutting us off — runtime is bounded only by
+ * the model itself and our own AbortController, not by CF.
+ */
+export async function streamSpecFromAnthropic(
+  prompt: string,
+  opts: { apiKey: string; model: string; maxTokens: number; signal?: AbortSignal },
+  handlers: StreamHandlers,
+): Promise<void> {
+  const client = new Anthropic({ apiKey: opts.apiKey });
+  let accumulated = '';
+
+  try {
+    const stream = client.messages.stream({
+      model: opts.model,
+      max_tokens: opts.maxTokens,
+      system: SYSTEM_PROMPT,
+      messages: [{ role: 'user', content: prompt }],
+    });
+
+    if (opts.signal) {
+      opts.signal.addEventListener('abort', () => stream.abort(), { once: true });
+    }
+
+    stream.on('text', (delta) => {
+      accumulated += delta;
+      handlers.onText(delta);
+    });
+
+    const final = await stream.finalMessage();
+
+    if (final.stop_reason === 'max_tokens') {
+      throw new SpecTruncatedError(
+        `model hit max_tokens (${opts.maxTokens}) before finishing the spec`,
+      );
+    }
+
+    const json = extractJson(accumulated);
+    const parsed = GeneratorSpec.safeParse(json);
+    if (!parsed.success) {
+      const preview = accumulated.slice(0, 400).replace(/\s+/g, ' ');
+      throw new SpecValidationError(`${parsed.error.message} :: raw="${preview}"`);
+    }
+    scanForInjection(parsed.data);
+    handlers.onSpec({ spec: parsed.data, source: 'claude' });
+  } catch (err) {
+    if (err instanceof Anthropic.APIConnectionTimeoutError) {
+      handlers.onError(new SpecTimeoutError('spec generation exceeded the time budget'));
+      return;
+    }
+    handlers.onError(err instanceof Error ? err : new Error(String(err)));
+  }
+}
+
 const GLM_ENDPOINT = 'https://open.bigmodel.cn/api/paas/v4/chat/completions';
 
 async function generateWithGlm(