perf(batch): avoid retaining clean fix results

[luojiyin1987]

Problem

In fix mode, batchLint() accumulates full BatchLintItem results including fixedResult.result (the complete fixed Markdown text) for ALL files, even clean ones. The allResults array is only consumed by getFixDevMetrics() which only needs path, fixedResult.metrics, and fixedResult.convergence.

This causes unbounded memory growth proportional to total input size:

worker working set: bounded (concurrency threads)
result retention:   unbounded (grows with file count × file size)

Example: 10k clean files × 1 MiB each → ~10 GiB retained in allResults.

Solution

Project clean fix items in the worker to a compact form before returning across the thread boundary. The full Markdown text is never structured-cloned back to the main thread for files that don't need writing.

Changes

• src/types.ts — Add CompactFixedResult interface and isFullFixedResult type guard
• src/utils/lint-worker.ts — Project clean items to compact form (only metrics/convergence)
• src/cli/run-lint.ts — Use type guard in write path

Closes #187