docs: document LLM-as-judge vs Granite Switch latency overhead with real numbers

[planetf1]

The README implies Mellea is a drop-in improvement without mentioning the very different latency profiles of the two requirement-checking modes. This causes surprises in production.

Evidence from eval-mFG (70 rows, Granite 8B on H100):

Task category	LLMaJ Mellea overhead	Switch Mellea overhead
constrained-format	3.2× native	0.8× native
structured-output	3.7× native	0.8× native
factual-hard	3.4× native	0.9× native
dual-constraint	0.9× native	0.9× native
reasoning	0.5× native	1.5× native

Switch is consistently faster because requirement checks are adapter forward passes (milliseconds). LLMaJ overhead ranges from 0.5× to 3.7× depending on whether the first draft passes (fast) or needs a revision — a full LLM judge call per requirement, per round.

What good docs look like:

## Requirement checking modes

### Granite Switch (aLoRA adapters) — recommended for production
Requirement checks run as adapter forward passes inside the Switch model.
Latency overhead: ~1.1× native (adapter check adds milliseconds, not seconds).
Requirements: GRANITE_SWITCH_API_BASE must point at a granite-switch endpoint.

### LLM-as-Judge (default, no extra infra)
Each requirement check is a full LLM call (~4-8s on a loaded H100).
With 3 requirements and 1 revision budget: up to 7 LLM calls per request vs
1-2 for native inference. Observed overhead: 0.5×-4× native depending on
first-draft pass rate.
Use this mode for: local development, CI, environments without a Switch server.

This information would help users choose the right mode at design time rather than discovering the latency cliff in production.

[planetf1]

Closing this — the ask was too thin to stand as an issue on its own. The aLoRA adapters docs already note the minimal overhead of adapter checks vs full context inference, which covers the core of the latency difference. If there's appetite for concrete numbers and a "choose your mode" guide, that's better addressed as a docs PR than tracked as an open issue.