Add exact tool approval review and explicit workflow continuation

[DrDrewCain]

Guarded agent calls now appear in run inspection with the exact tool arguments, selected model, workflow step and tool revision. Users can approve or deny a call, then explicitly continue only the selected saved decisions. Saving a decision does not execute a tool or invoke a model.

The console verifies approval and continuation receipts against the immutable run request, preserves large integer argument literals, and makes directional Unicode controls visible without changing the bound arguments. Ambiguous continuations retain their original ID and batch across status refresh and same-run reopening; server-saved continuations can be recovered after browser reload. Cancelled and unknown-outcome runs cannot continue.

Validation:

• 500 tests pass; TypeScript, production build and packaged asset checks pass. The existing large-bundle warning remains.
• The committed packaged-browser regression covers ambiguous-response retry identity across refresh and pending/history directional-control rendering.
• Additional isolated Chrome/native HTTP checks passed at desktop and 390px mobile widths: double-click produces one decision POST, approval alone produces no effects, denial skips the handler, only selected parallel calls execute, cancelled/unknown runs do not replay, failed-admission recovery reuses the exact continuation, and a real server restart preserves the saved decision and executes once after explicit continuation.
• Independent scoped review cleared both reproduced findings: retry identity lost on refresh and misleading directional-control display.

Stacked on Web19 (feature/agent-tool-catalog). Requires native approval runtime PR87. Native stages and this UI milestone do not complete the broader reference gap work. Merge coordination with Fable remains pending.