Subagent context usage is not tracked or visible in parent agent statistics

Problem Description

Currently, subagent context/token usage is completely isolated from the parent agent'''s statistics. This leads to:

1. Invisible subagent context consumption: The parent agent'''s status bar shows only its own context usage, giving users no visibility into how much context their subagents are consuming.
2. No compaction trigger for subagents: Auto-compaction only considers the parent context, so subagents can grow unbounded without triggering compaction.
3. No inspection mechanism: There is no command, API, or UI feature to check a subagent'''s current token count or context usage percentage.

Current Behavior

• Each subagent gets its own Context instance with a separate context.jsonl file (src/kimi_cli/subagents/core.py:59)
• Each subagent gets its own KimiSoul instance (src/kimi_cli/subagents/core.py:85)
• StatusUpdate wire messages only contain the current soul'''s context metrics (src/kimi_cli/soul/kimisoul.py:951-961)
• The UI status bar (src/kimi_cli/ui/shell/prompt.py) only renders the parent soul'''s StatusSnapshot
• The Vis API /subagents endpoint only returns raw file sizes (bytes), not token counts

Expected Behavior

Users should be able to:

1. See subagent context usage aggregated in the parent agent'''s status, OR
2. Have a way to inspect individual subagent context usage (e.g., a slash command or API endpoint)

Possible Solutions

1. Aggregate in parent status: Sum active subagents''' token_count into the parent'''s StatusSnapshot.context_usage
2. Expose subagent stats: Add token_count/context_usage fields to AgentInstanceRecord and expose via Vis API
3. New slash command: Add /subagents command to list running subagents with their context usage
4. Event propagation: Forward subagent StatusUpdate events to the parent wire stream

Additional Context

The Context class already tracks _token_count internally, so the data exists — it'''s just never read or exposed for subagents.

Start with subagent lifecycle and context handling in src/kimi_cli/subagents/core.py, then trace StatusUpdate and StatusSnapshot in src/kimi_cli/soul/kimisoul.py and rendering in src/kimi_cli/ui/shell/prompt.py. Review the Vis API /subagents endpoint and choose a defined exposure or aggregation path; done means subagent context usage is observable and compaction behavior is addressed without regressing parent statistics.