fix: warn before context-heavy model switches

[danielkov]

Summary

Warn before a model switch when the estimated current transcript occupies at least 80% of the target model's context window. The Kit TUI offers Continue anyway, Compact, and Cancel, with compaction using the original model before applying the requested selection.

Motivation

Addresses #18: switching to a smaller context window can make both the next request and subsequent compaction fail if the new model is selected first.

Impact

Cancel preserves the selection, transcript, and input draft. Compaction failure or cancellation keeps the original model selected; a replacement that already completed is retained. Unknown transcript usage or target context limits allow an explicitly documented unchecked switch rather than forcing compaction or guessing a limit.

External ACP clients receive an actionable confirmation-required error instead of a Kit dialog. Both ACP versions accept the same one-shot confirmation extension for continuing or compacting before switching.

Technical details

Provider catalogs retain reported context-window metadata in memory. The guard uses ceil(latest_transcript_tokens × 1.20) >= target_context_window × 0.80, not accumulated historical usage or double-counted cached tokens.

Actor-owned confirmations bind the original selection, target, transcript, and cancellation generation. The TUI handles requests asynchronously and correlates completions with session generation and operation ID. A unique manual-compaction marker, consumed only after successful transcript replacement, gates selection changes; uncorrelated runtime notifications cannot complete a pending switch.

[kit-code-agent]

No issues found. The changes look good to merge.

[kit-code-agent]

No issues found. The changes look good to merge.