perf: skip redundant deep copy in CostPolicy.apply_token_limit

[shrdgn]

What & why

Every panel-member call (openfusion/panel.py:_call_member) and every judge synthesis call (openfusion/synthesize.py:synthesize) deep-copied the request body twice: once by the caller to isolate it from the shared original, then a second time inside CostPolicy.apply_token_limit (openfusion/cost.py).

This is pure wasted CPU/memory that scales with panel size and debate rounds, and is invisible in normal tests since correctness isn't affected — only latency/memory for large multi-turn or multimodal (e.g. base64 image) request bodies.

Fix: add an already_isolated: bool = False keyword-only flag to apply_token_limit. When True, the caller is asserting it already owns an exclusive copy of body, so the method mutates and returns it in place instead of deep-copying again. Default behavior — and the existing "does not mutate the original" contract relied on by server.py's two call sites — is unchanged.

How it was tested

• ruff check . passes
• pytest -q passes (476 passed, no live network)
• New behavior has a test (tests/test_cost.py: identity check that no extra copy is made, in-place mutation, and that reject_over_limit still raises correctly under already_isolated=True)
• Docs updated if config / request surface / defaults changed — n/a, no public request/config surface changed
• No secrets, prompts, or response bodies added to logs or metrics
• Quality/cost claims backed by a reproducible bench/run.py number (if applicable) — n/a, this is a copy-elimination perf fix, not a quality/cost tradeoff; behavior is bit-for-bit identical, verified by the existing test suite

Notes for reviewers

Only the two hot-path callers that provably already own an isolated copy at the call site were switched to already_isolated=True:

• panel.py:_call_member — deep-copies request_body into body and already mutates it (.pop("model"), .pop("stream")) before calling apply_token_limit.
• synthesize.py:synthesize — deep-copies request_body into judge_body and already mutates it (sets messages, pops model, strips tool fields) before calling apply_token_limit.

server.py's two callers (policy.apply_token_limit(body, ...)) were deliberately left on the default (deep-copying) path since they call it directly on the request body without a prior isolating copy of their own.

---

Generated by Claude Code