feat(Inference): disable extended-thinking budget for fast/snap-judgment calls

[caraka]

Problem

Every claude --print subprocess spawned by Inference.ts silently spends an extended-thinking budget. On snap-judgment payloads (classification, tab titles, session naming) this adds ~10s+ of pure latency with negligible quality benefit — and the thinking budget, not the model tier, dominates the cost.

Measured on a classification-shaped payload:

call	latency
claude --print --model haiku (default)	~14.9s
same call, MAX_THINKING_TOKENS=0	~3.2s
--model sonnet, MAX_THINKING_TOKENS=0	~3.6s

Downgrading the model barely helps; disabling the thinking budget is what collapses the latency.

Change

• Add a thinking?: boolean option to inference() and a --no-thinking CLI flag. When thinking: false (or --no-thinking), the call sets MAX_THINKING_TOKENS=0 in the subprocess env, disabling the extended-thinking budget.
• level: 'fast' implies it — fast is for snap tasks (quick generation, basic classification) by definition, so it no longer pays the thinking tax. standard and smart are unchanged unless they opt out explicitly.

Single file, +22/−2, no new dependencies. Parses/transpiles clean; CLI usage string updated.

Reproduction

time claude --print --model haiku --tools '' --output-format text \
  --system-prompt 'Classify the message as A or B. Output one letter.' <<< 'hello'

time MAX_THINKING_TOKENS=0 claude --print --model haiku --tools '' --output-format text \
  --system-prompt 'Classify the message as A or B. Output one letter.' <<< 'hello'

Note

The one behavior change is the fast-level default (fast now disables thinking). Happy to make that opt-in instead if you'd prefer zero default change — it's a one-line tweak.