feat(tools): preserve sampled output for selective rereading

[caihongwei2006]

Summary

Long command results currently keep only their head and tail, and later context pruning can erase the saved-output reference. Sample across the full result and retain a task-scoped reference so the agent can inspect omitted lines without rerunning the command.

The preview starts with an 8,000-character budget, distributed across ceil(length / 8000) samples. Each edge keeps at least 2,000 characters through additional allocation; the remaining gaps are evenly spaced. A 36,000-character result therefore returns 2000 + 1600 + 1600 + 1600 + 2000 source characters, with four 6,800-character gaps.

tool_output(output_id, block_id) expands a numbered omitted block into its original inclusive line range; previews include the exact expansion call. The reader also supports selected-line reads and 8k-bounded continuation that stops at the selected block’s final line. Sanitized originals survive idle session rebuilds and are reclaimed on explicit task close; context and persistence guards retain retrieval references. Command/script capability groups include the reader, while legacy explicit whitelists keep the existing file/read fallback. No new production dependencies.

Test Plan

• Sampling, edge minima, Unicode boundaries, and rendered range labels
• Block-to-line mapping, block-bounded continuation, selective reads, concurrent task isolation, and sanitized storage
• Reference preservation across context/persistence pruning and cleanup after task closure
• Full npm test on the isolated branch based on current main
• npx tsc --noEmit
• npm run build
• Manual live-agent verification