Defend against ACP server reporting used > size

[timvisher-dd]

Closes xenodium#364

The ACP server (claude-agent-acp) has a bug where model switches cause used to exceed size in session/update notifications. For example, switching from Opus 1M to Sonnet 200k drops size to 200000 while used keeps growing past it (observed: 419574/200000 = 209.8%). This results in nonsensical context indicators and percentages e.g. (from a real session)

 Context: 420k/200k (209.8%)
  Tokens: 32 in · 11k out · 11m cached (11m total)
    Cost: USD75.96

While I intend to get a fix for this improper reporting into claude-agent-acp, agent-shell should be robust against it. To that end, when the garbage usage data is observed, the UI now signals unreliable data instead of showing nonsense:

• Context indicator shows ? with warning face when used > size
• Formatted usage shows raw numbers with (?) instead of a bogus percentage
• A regression test replays the real observed ACP traffic from the model-switch scenario so this class of bug is caught going forward

This also adds comprehensive ERT test coverage for agent-shell-usage.el: notification updates, indicator scaling/colors, compaction replay, token saving, and number formatting.

For the claude-agent-acp side of this see agentclientprotocol/claude-agent-acp#412

Test plan

• All 21 ERT tests pass
• checkdoc clean on both files
• byte-compile clean on both files
• Manual baking verification