[Feature] Per-invocation --fast flag for headless claude -p

Summary

Fast mode is currently toggleable via /fast (interactive) or "fastMode": true in settings.json (persistent, and inherited by headless sessions). There is no per-invocation flag.

Request: add --fast / --no-fast as claude -p flags so fast mode can be enabled (or explicitly disabled) for a single non-interactive invocation, without changing a global/persistent setting.

Motivation

Automation fleets run mostly cost-sensitive batch work, which correctly belongs on standard mode (the docs themselves recommend standard mode for "batch processing or CI/CD pipelines" and "cost-sensitive workloads"). But a few latency-critical headless paths — e.g. a real-time, user-facing response endpoint — would benefit from fast mode on just those calls.

Today the only headless option is the global "fastMode": true in settings.json, which is all-or-nothing and risks billing an entire fleet via usage credits. A per-invocation flag would give the same granular control that --effort and --model already provide per call.

Proposed behavior

• claude -p --fast ... enables fast mode for that invocation only.
• claude -p --no-fast ... forces standard mode for that invocation even if "fastMode": true is set globally.
• Mirrors the existing per-call precedence of --model and --effort.

Environment

• Claude Code CLI 2.1.156
• Subscription auth (no API key)