feat(server): Bun/TypeScript re-implementation of PR review engine

Reimplements the pr_agent review pipeline (previously Python package) in
Bun/TypeScript, verified end-to-end against a real GitHub App + 9router:

- src/diff.ts: git patch processing (extend_patch, line-numbered hunks,
  token-budget generate_full_patch, generated/invalid file filters)
- src/review.ts: orchestrator (fetch PR/diff -> render prompts -> LLM ->
  YAML parse -> markdown render -> publish), fallback models
- src/github.ts: GitHub App auth via @octokit/auth-app (JWT -> installation
  token, custom authStrategy for Bun), PR/diff/languages/comments REST
- src/prompts.ts: verbatim pr_reviewer_prompts.toml port (nunjucks templates)
- src/yaml.ts: resilient load_yaml with 9 fallback fixers
- src/markdown.ts: convert_to_markdown_v2 port ('PR Reviewer Guide' comment)
- src/llm.ts: 9router OpenAI-compatible chat completions (non-stream,
  temp omitted for claude-opus-5), timeout 600s
- src/token.ts: js-tiktoken o200k counting
- src/index.ts: webhook server (HMAC verify, segment analytics, /health,
  /api/metrics) + cli.ts one-shot review
- e2e.ts: real-world harness (bun e2e --repo o/r --pr N [--publish])

Verified: bun test 13/13, tsc clean, E2E published a real review comment
(## PR Reviewer Guide) on asepharyana/nextjs-template#19 via GitHub App
MythEclipseBotReview + claude-opus-5 through 9router.

Python run_server.py + pr_agent remain for the live queue worker; server/
is the replacement path.

server/src/token.ts

@@ -0,0 +1,50 @@
+// Token counting — port of pr_agent.algo.token_handler using js-tiktoken
+// (o200k_base, same as Python tiktoken). The Python version tried an accurate
+// Anthropic count_tokens API call when anthropic.key was set; with 9router that
+// call is not an Anthropic API, so we always use the local tokenizer estimate
+// (which is what the Python server effectively falls back to on failure).
+
+import { getEncoding, type Tiktoken } from "js-tiktoken";
+
+let _encoder: Tiktoken | null = null;
+
+export function getTokenEncoder(): Tiktoken {
+  if (!_encoder) {
+    try {
+      _encoder = getEncoding("o200k_base");
+    } catch {
+      // cl100k_base as a safe fallback if o200k unavailable on old bundles
+      _encoder = getEncoding("cl100k_base");
+    }
+  }
+  return _encoder;
+}
+
+export function countTokens(text: string): number {
+  if (!text) return 0;
+  try {
+    return getTokenEncoder().encode(text, "all").length;
+  } catch {
+    // approximate: ~4 chars/token
+    return Math.ceil(text.length / 4);
+  }
+}
+
+/**
+ * Approximation of pr_agent's TokenHandler.prompt_tokens: renders the system
+ * and user prompt strings with the given vars (nunjucks), then counts tokens.
+ */
+export function countPromptTokens(
+  systemTemplate: string,
+  userTemplate: string,
+  vars: Record<string, unknown>,
+  render: (tmpl: string, data: Record<string, unknown>) => string,
+): number {
+  try {
+    const system = render(systemTemplate, vars);
+    const user = render(userTemplate, vars);
+    return countTokens(system) + countTokens(user);
+  } catch {
+    return 0;
+  }
+}