feat(engine): expose context_input_budget_for_route for external budget reuse

[h3c-hexin]

Problem

context_input_budget_for_route — the internal input-budget computation (window − window-dependent output reservation − headroom) that is the single source of truth for the preflight / emergency-recovery / capacity-trim decisions — is pub(super) to core::engine.

External hosts / wrappers that embed the engine and derive their own compaction or pressure trigger line have no way to reuse it, so they end up mirroring the constants: the ≥500K large-window threshold, TURN_MAX_OUTPUT_TOKENS, the output-reservation split in route_output_reservation_for_window, and the headroom. When the engine later changes any of these, the mirrored copy still compiles but silently diverges — the host computes a different budget than the engine actually enforces. (In our case the host's "compact now" line drifted above the engine's emergency line, so the graceful summarize path stopped firing and every long session fell through to emergency recovery.)

Change

Make context_input_budget_for_route pub and re-export it from core::engine, so a host can compute the exact same budget the engine uses. Pass input_tokens = 0 to get the full route budget.

Purely a visibility change:

• No behavior change; the function body is untouched.
• All parameters and the return type are already public (ApiProvider, RouteLimits, usize).
• cargo check -p codewhale-tui passes.

This lets downstream embedders reuse the engine's budget math instead of re-deriving (and drifting from) its internal constants.

[Hmbown]

Merged into main for the 0.8.67 finalization train as #3968 (25cce4074096b22dc9b531a7810f67dc0391e3e6). Thanks @h3c-hexin. I treated this as behavior-neutral API reuse surface, and it compiled cleanly in the local scratch train with the other accepted PRs.