PreCompact/PostCompact hooks have no visibility into what's being lost

Problem

I built a "lossless compaction" system using PreCompact and PostCompact hooks. PreCompact dumps critical state (current task, file paths, decisions, git branch) to disk. PostCompact reads it back and re-injects it.

It works, but it's held together with duct tape because the hooks get zero information about the compaction itself:

1. No context size info — I don't know how much is being dropped. Am I losing 10K tokens or 200K? The preservation strategy should differ.
2. No ability to influence preservation — I can save state externally, but I can't tell the compaction engine "these 3 messages are critical, keep them." My hook is essentially racing against compaction with no coordination.
3. No diff — PostCompact doesn't know what survived vs what was lost, so it blindly re-injects everything, often duplicating what the compaction already kept.

What I'd want

Pass richer metadata to compaction hooks:

{
  "pre_compact_token_count": 180000,
  "post_compact_target": 40000,
  "tokens_being_dropped": 140000,
  "message_count_before": 87,
  "message_count_after": 12
}

And ideally, let PreCompact return a preserve_hints array — short strings the compaction engine should try to keep:

{
  "preserve_hints": [
    "Currently on branch feat/agent-metrics",
    "Editing /Users/josh/hemp-route/lib/crm/accounts.ex",
    "Josh decided to use GenServer over Task.Supervisor"
  ]
}

Why this matters

Compaction is the #1 context loss event for long sessions. Power users who build memory systems around it are doing so completely blind. Even just exposing token counts before/after would be a massive improvement.