Parent

Part of #3000 and follows #4179.

Goal

Add structured /doctor memory output and session-scale counters so memory reports can distinguish V8 pressure from large conversation/tool-result retention.

Motivation

A single process/V8 snapshot is useful, but Qwen Code OOMs are often caused or amplified by agent-specific state: long chat history, large tool results, UI history, or compression-time duplication. /doctor memory should expose safe aggregate counters before asking users for heavier heap snapshots.

Proposed solution

• Add a structured output path for /doctor memory, e.g. /doctor memory --json in non-interactive and interactive command paths.
• Extend diagnostics with safe aggregate session stats when available:
  • conversation/history item count
  • approximate total chars/tokens if an estimator is available
  • largest tool result size and total tool result size, without content
  • UI history item count, if accessible
  • current model/context window metadata when safe
• Gracefully report unavailable when the command context does not expose a counter.
• Keep all stats content-free: counts and sizes only.

Non-goals

• Do not serialize full conversation messages.
• Do not include tool result contents.
• Do not implement tool-result offload in this issue.
• Do not add heap snapshot generation here.

Acceptance criteria

• /doctor memory --json returns valid JSON in non-interactive mode.
• Human-readable output includes a Session scale or equivalent section when stats are available.
• Tests verify that session stats are aggregate-only and do not include message/tool content.
• The command still works in contexts without session/UI stats.