feat(pipeline): add --llm-scores flag to pipeline bench

[christso]

Objective

Add a --llm-scores <file> option to agentv pipeline bench so LLM scores can be provided via a file path instead of stdin only.

Problem

pipeline bench currently only accepts LLM scores via stdin. This causes issues on Windows PowerShell where the < operator is reserved:

ParserError: The '<' operator is reserved for future use.

The workaround (Get-Content file | agentv pipeline bench <dir>) is non-obvious and undocumented in the skill instructions.

Proposed Change

Add an optional --llm-scores flag to pipeline bench:

# Existing stdin approach (still works)
cat llm_scores.json | agentv pipeline bench <dir>

# New file flag approach
agentv pipeline bench <dir> --llm-scores llm_scores.json

When --llm-scores is provided, read from that file instead of stdin. When omitted, fall back to stdin (backward compatible).

Implementation

In apps/cli/src/commands/pipeline/bench.ts:

• Add optional --llm-scores option (type: string)
• If provided, read from the file path; otherwise call readStdin() as today

Acceptance Signals

• agentv pipeline bench <dir> --llm-scores scores.json works
• Stdin still works when --llm-scores is omitted
• Works on both Windows and Unix

Non-Goals

• Changing the scores JSON format
• Auto-discovery of scores files in the export directory