feat: results web viewer

[christso]

Problem

AgentV results are viewed in the terminal or by reading JSONL files. For larger eval suites, a visual interface makes it much easier to:

• Browse and filter results
• Compare runs side-by-side
• Drill into individual case evaluator scores
• Share results with team members

Proposal

Add a local web viewer that serves eval results:

# View results from a completed run
agentv view results.jsonl

# View and compare two runs
agentv view before.jsonl after.jsonl

# Auto-open after eval
agentv run --target my-agent evals/ --view

Features

• Results table — sortable by score, duration, cost, pass/fail
• Case detail view — expand to see all evaluator scores and reasons
• Run comparison — side-by-side delta view (reuses existing compare logic)
• Filters — by pass/fail, evaluator type, score range
• No external dependencies — self-contained, runs locally

Design Principles

• CLI-first remains primary — the viewer is supplementary, not required
• No server signup — runs locally on localhost
• No persistent backend — reads JSONL files directly
• Lightweight — minimal framework, fast startup

Why This Matters

• Visual exploration catches patterns that tables miss
• Comparison view is the most common evaluation workflow
• Accessibility for non-CLI users (PMs, stakeholders)
• Stays true to AgentV principles: local, no signup, no overhead

Acceptance Criteria

• agentv view command serves local web UI
• Results table with sorting and filtering
• Case detail expansion
• Run comparison view
• No external service dependencies
• Lightweight and fast startup

[christso]

Closing per design principle #1: "CLI wrappers that consume AgentV JSON/JSONL output for post-processing (aggregation, comparison, reporting)." A results viewer is exactly this pattern — a separate tool that reads JSONL, not a core feature. Can be built as an independent package.