Context % uses limit.context instead of limit.input for split-window models #51251

Description

For models with a separate input limit (e.g. gpt-5.3-codex: context 400k, input 272k), the context % in the TUI divides by limit.context. Compaction uses limit.input (see usable() in packages/opencode/src/session/overflow.ts), so the % shown is lower than how close the session actually is to compacting.

Example with gpt-5.3-codex at 272k tokens: TUI shows 68%, but the input window is full.

Affected places:

• packages/tui/src/component/prompt/index.tsx (prompt footer)
• packages/tui/src/feature-plugins/sidebar/context.tsx (sidebar)
• packages/tui/src/routes/session/subagent-footer.tsx (subagent footer)
• packages/opencode/src/cli/cmd/run/runtime.boot.ts (run limits map)

Steps to reproduce

1. Use a model whose models.dev entry has limit.input lower than limit.context (e.g. openai/gpt-5.3-codex)
2. Run a long session
3. Compare the context % with when compaction triggers