Add local model support: run order, warmup, and per-scenario metrics

[jeanfbrito]

Summary

Three features that make ToolCall-15 practical for benchmarking local GPU models served by on-demand backends like llama-swap.

All new env vars are opt-in — existing setups without them continue to work identically.

1. Configurable run order (RUN_ORDER)

• scenario (default): runs each scenario across all models before moving on
• model: runs all 15 scenarios for one model before swapping to the next — avoids constant GPU reloads

2. Model warmup (WARMUP_ENABLED)

• When true, sends a lightweight max_tokens=1 request before each model's benchmark run to trigger GPU loading (120s timeout)
• Dashboard shows a "Warming up" indicator with the model name during loading

3. Per-scenario performance metrics

• Captures duration, token usage (prompt + completion), turns, and tool call count per scenario
• Duration shown directly in each result cell as it completes
• Aggregate metrics (total time, avg time/scenario, tok/s, total tokens) displayed in score cards
• Full per-scenario breakdown on hover tooltip

Bonus: Trace dialog fix

• Renamed FailureDialog → TraceDialog — now shows correct status (Passed / Partial / Failed / Timed out) instead of always showing "Failed"

Changed files

File	Change
lib/llm-client.ts	warmupModel() function + extract TokenUsage from API responses
lib/benchmark.ts	ScenarioMetrics and AggregateMetrics types, aggregate computation in scoreModelResults()
lib/orchestrator.ts	Model-first execution path, warmup integration, timing and token capture
components/dashboard.tsx	Warmup UI state, metrics in cells and score cards, TraceDialog with correct status
app/globals.css	.metrics-strip, .cell-duration, button layout fix for .result-icon-shell
.env.example	New RUN_ORDER and WARMUP_ENABLED vars with documentation
README.md	"Local Model Configuration" section, updated execution model and dashboard behavior docs

Backward compatibility

Both new env vars default safely when absent:

• RUN_ORDER undefined → scenario-first (existing behavior)
• WARMUP_ENABLED undefined → no warmup (existing behavior)

No breaking changes to types — all new fields are either optional (metrics?: ScenarioMetrics) or always populated by their sole producer (scoreModelResults()).

Test plan

• npm run typecheck passes
• Default config (no new env vars) behaves identically to main
• RUN_ORDER=model completes all scenarios per model before swapping
• WARMUP_ENABLED=true shows "Warming up" and preloads model before scenarios
• Duration appears in each cell immediately after scenario completes
• Score cards show aggregate metrics after run finishes
• Clicking a pass cell shows "Passed" in trace dialog
• Hover tooltip shows duration, turns, tool calls, tokens