Bulk provider switch with remembered last batch (usage-limit escape) #360

Problem

When you hit a usage limit on one provider while running many agents on it, every one of those agents is stuck. Today the only escape is the single-agent Switch Provider command, one focus-and-switch at a time. With ~20 agents on the throttled provider, that's 20 manual operations under time pressure.

Feature

A "Switch Agents to Another Provider…" modal (command palette) that:

1. Moves a batch of Claude/Codex agents to the other provider in one action — pick direction (Codex → Claude / Claude → Codex) and scope (All / Selected projects), preview the affected agents, confirm.
2. Remembers the most recent batch so the same set can be sent back to its origin provider — from a banner inside the same modal (no separate command, no keybind). The return is optional; it stands until used or superseded.

This is a bulk wrapper + a memory record over the existing single-agent switchProvider machinery (translate transcript → new file/id → replaceSession).

Decisions

• Round-trip, not auto-undo: "return" re-translates the live session back to its origin provider so interim work done on the borrowed provider is preserved (rather than snapshot-restoring the pre-switch transcript and dropping it).
• Mid-turn agents are switched along with the rest, but the preview surfaces how many in-flight turns will be interrupted (no skip toggle) — the trigger is "get everyone off this provider".
• One level of memory, in-memory for the working session (not persisted: batch holds current-launch SessionIds that get remapped on reload).

Design doc

docs/superpowers/specs/2026-06-24-bulk-provider-switch-design.md