Franklin Agent 3.36.0 — Claude Opus 5 is the default Opus

`opus` and `claude` now resolve to `anthropic/claude-opus-5`. It is live on the BlockRun gateway at $5/$25 per 1M tokens — the same price as Opus 4.8, so this is a straight capability upgrade with no cost delta and nothing to opt into. The Smart Router's COMPLEX and REASONING tiers point at it, `franklin init` writes it as `ANTHROPIC_DEFAULT_OPUS_MODEL`, and the savings-vs-Opus figure in `/stats` and the panel now baselines against it.

Opus 4.8 leaves the visible picker rather than sitting next to a strictly-better entry at an identical price. Its `opus-4.8` shortcut stays live, as do `opus-4.7`, `opus-4.6`, and `opus-4.5`. Opus 4.8 and 4.7 remain in the router's fallback chain.

Opus 5 never receives the extended-thinking flag. It thinks by default — omitting the `thinking` field runs adaptive, unlike 4.8 and 4.7 where omitting it meant no thinking at all — and it rejects `budget_tokens` outright.

The context window stays pinned at 200k until a real >200k call is verified end-to-end; max output is set to the full 128k.

Also fixed: compacting a long Opus session was failing outright. Anthropic requires `1024 <= budget_tokens < max_tokens`; Franklin sent `min(max_tokens, 16384)`, which violates both ends. Compaction tiers an Opus session down to Sonnet 4.6 and asks for exactly 16,000 tokens, so every attempt to summarize a long Opus session hit a 400. Subagents (16,384) and the turn evaluator (512) were on the same path. Franklin now reserves a quarter of the ceiling for the visible answer and drops the thinking block when no legal budget exists.