fix: use transcript model to correct cross-session model contamination

[AniAggarwal]

Summary

Fixes the model display showing the wrong model when multiple Claude Code sessions are open and one changes models. The root cause is upstream (anthropics/claude-code#19570) — the model field in statusline JSON reads from global shared state instead of per-session state.

• Extracts the model from the per-session transcript (which has the correct message.model on each assistant entry) inside getTokenMetrics — no extra file read
• Overrides the stdin model in renderMultipleLines when the transcript model differs
• When the last transcript entry is a user message (API call in progress), returns null to avoid showing a stale model — self-corrects once the response arrives

Trade-off

Brief window during API calls where cross-session contamination can still show. Acceptable vs permanently displaying the wrong model. This workaround can be removed when Claude Code fixes the upstream bug.

This is a pretty big patch for a bug in upstream, so feel free to reject. I was getting quite annoyed with the status line changing between chats so figured I'd make a PR.

Closes #190

Test plan

• 3 new tests: model extraction, pending-user null, missing model null
• All existing tests updated and passing (67 tests across 6 files)
• tsc --noEmit clean
• Manual test: two parallel sessions, change model in one, both show correct models