feat: emit time_to_appetizer Braintrust metric

[yodem]

Implements sc-44755.

Adds a numeric, aggregatable `metrics.time_to_appetizer` on the trace —
user-perceived latency from request receipt to the appetizer being served
(seconds, one decimal). Only set when an appetizer is actually served; the
existing `metadata.appetizer` debug blob is unchanged.

• Emitted under Braintrust `metrics` (avg/percentiles), not metadata.
• Deterministic test runs the appetizer thread inline (no race).

Open question for Mickey (non-blocking): the metric stops at server enqueue,
which is lower than the client-paint timing referenced in the story. Confirm
this boundary is acceptable.

🤖 Generated with Claude Code

[gitvelocity-reviewer]

📊 Code Quality Score: 11/100

28 × 0.4 = 11.2 → 11

Category	Score	Factors
🔭 Scope	5/20	2 files, single subsystem (observability), no new public APIs or endpoints
🏗️ Architecture	4/20	Promotes time_to_appetizer from metadata blob to first-class numeric Braintrust metric; minor but intentional design decision with backward compatibility
⚙️ Implementation	5/20	Simple conditional dict construction and unpacking; round() for 1-decimal precision; no algorithmic complexity
⚠️ Risk	3/20	Purely additive change; backward compatible (metadata preserved); existing exception handling retained; no migrations or schema changes
✅ Quality	10/15	Good test coverage with sync executor trick to eliminate threading races; float type check, non-negative assertion, 1-decimal rounding check; missing negative test for no-appetizer-served case
🔒 Perf / Security	1/5	Metric is lightweight (single float); no security implications; no benchmarks needed at this scale

Was this score accurate? 👍 Yes · 👎 No

_{Scored by GitVelocity · How are scores calculated?}