discover: add output-size logging to fallback path for accurate missed-savings estimates

Problem

rtk discover estimates missed savings by matching command names against known RTK handlers. This produces false positives for commands that RTK handles but can't profitably compress — because the output is already small.

Example: 91 git push calls were flagged as ~5.8K tokens saveable. In practice, git push output is typically 1-3 lines (Everything up-to-date, or a brief ref update). RTK correctly falls back with 0% savings. No tokens were actually wasted.

The core issue: discover has no way to distinguish:

• grep that returned 3 lines (no real savings possible)
• grep that returned 3,000 lines (real savings missed)

Proposed fix

Add output-size logging to the fallback path. When RTK decides not to compress a command result, record the output token/byte count alongside the fallback event.

rtk discover (or a new rtk discover --accurate mode) could then surface only fallbacks where output exceeded a meaningful threshold — these are the genuine missed opportunities where context is being wasted.

What this enables

A reliable answer to: "are there commands producing large output that RTK isn't compressing?" — the actual question users care about. The current heuristic answers a different question ("does RTK have a handler for this command name?") that doesn't map to real context waste.

Current workaround

None. Users have to manually test commands or check rtk gain --history to identify false positives after the fact.