feat(acp): report Pi context window usage

[KorenKrita]

Summary

Forward Pi's authoritative context-window occupancy through the standard ACP usage_update notification.

Pi already exposes the active context via get_session_stats.contextUsage, but pi-acp did not forward it. ACP clients therefore had to show no context data or guess from model names and cumulative token totals. This is especially visible for custom/extended-context models (for example, a configured 1M window can be shown as 258K or 200K).

What changed

• map contextUsage.tokens to ACP used
• map contextUsage.contextWindow to ACP size
• publish usage:
  • after session/new
  • after session/load
  • after model changes
  • after agent_settled, before resolving session/prompt
• skip updates when Pi reports tokens: null or invalid/non-integral values
• never substitute cumulative tokens.total
• omit cost because Pi's stats do not carry an explicit currency
• make get_session_stats best-effort with a 1s RPC timeout
• remove timed-out requests from the RPC pending map and swallow late responses

The agent_settled boundary is important on current main: agent_end may be followed by retry, compaction, or queued continuation work.

Compatibility

Older Pi versions that do not expose contextUsage continue to work; no usage update is emitted. Immediately after compaction, Pi may return tokens: null, so clients retain the previous reading until a trustworthy estimate is available.

This is a current-main implementation of the behavior discussed in #48, #75, and #87. Those PRs are now conflicting and/or target the older agent_end lifecycle.

Validation

• npm run format -- --check
• npm run typecheck
• npm run lint
• npm test — 110 passed, 0 failed
• npm run build
• git diff --check

A live Pi 0.83 RPC probe returned:

{
  "contextUsage": {
    "tokens": 100,
    "contextWindow": 1000000,
    "percent": 0.01
  }
}

The adapter maps that to:

{
  "sessionUpdate": "usage_update",
  "used": 100,
  "size": 1000000
}

Tests cover initial/new and loaded sessions, both model-switch entry points, agent_settled delivery ordering, cancellation, null/invalid usage, RPC failures/timeouts, pending-request cleanup, and late-response handling.