feat(review): add token-aware context control with tokenlens

Replace hardcoded char-count context limits with token-based budgets using
tokenlens (data from models.dev). TokenCounter provides 3-tier context window
lookup: dynamic catalog (refreshed every 24h) → static tokenlens → 128k default.

- specialist-agent: token budget from model context window instead of MAX_CONTEXT_CHARS=100k
- critic-agent/reflexion-agent: tokenCounter.clip() instead of diff.slice(0, 3000/2000)
- diff-extractor: raw diff clipping at 30k tokens
- engine.ts: refreshCatalog() at startup, stopRefresh() at shutdown

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)

package.json

@@ -15,6 +15,7 @@
     "hono": "^4.11.9",
     "lodash-es": "^4.17.21",
     "openai": "^4.87.3",
+    "tokenlens": "^1.3.1",
     "zod": "^3.25.1",
     "zod-to-json-schema": "^3.25.1"
   },