Opt-in, machine-wide crash-safe unfocused-responsive mode

To keep MCP commands responsive while the editor is unfocused — the normal state during a session, since the user is looking at the chat — the toolkit lowers the global editor setting interface/editor/unfocused_low_processor_mode_sleep_usec (default ~100000 µs ≈ 10 fps) to a boosted value while an authenticated client is connected, then restores it on the last disconnect. The engine re-reads that setting on every focus-out (editor_node.cpp), so writing the EditorSetting is the correct, clean integration point.

The original implementation held the pre-boost value only in memory and mutated the setting silently and unconditionally. That had five problems: it is an obtuse global side-effect (the key is machine-wide — every project on that editor version), it was not optional, its state was invisible, and — the real footgun — set_setting does not persist immediately (editor_settings.cpp); it reaches disk only when some later save() flushes it. Once flushed, a crash (or a concurrent second editor) left the global stranded at the boosted value, and on the next launch the leftover boosted value was read back as the “original” — so the true default was lost permanently. The crash and concurrency failures share one root cause: reading a boosted value as the original.

Decision: Option A — keep the dynamic connect-boost/disconnect-restore behaviour, but make it opt-in (default on), visible, machine-wide crash-safe, and conflict-aware.

• Opt-in, tunable, in Editor Settings (not Project Settings). Two new keys: mcp_toolkit/performance/keep_editor_responsive_unfocused (bool, default true) gates the behaviour, and mcp_toolkit/performance/unfocused_responsive_sleep_usec (int, default 16666 = 60 fps; not clamped) is the boosted value. These are the only EditorSettings the toolkit owns — every other toolkit setting is a mcp_toolkit/* ProjectSetting. The exception is deliberate: the effect is machine-global, the preference is personal (battery/CPU), it must not be committed to project.godot/VCS, and “off” is only truly off at global scope (a per-project opt-out cannot stop a concurrent project from boosting the shared global).
• Machine-wide, version-keyed, first-writer-wins, conflict-aware backup. The true original is persisted once to unfocused_sleep_backup_<major.minor>.json in the registry dir (registry_client.gd::registry_dir()), under the existing registry file lock. A second instance that connects while a backup already exists does not overwrite it (so it never captures a boosted value as the original). Restore — on disconnect, and as a startup self-heal — reads the live value C: if C equals the value we wrote (B), restore the stored original O; if C != B, a human/other tool changed it, so keep C. Either way the backup is deleted. The backup stores both O and B so the compare works even if the configured boosted value changed between sessions.
• Default 60 fps (not 30). Interactive users can’t perceive 30 vs 60 (both dwarfed by agent thinking time), but automated rapid-fire runs (smoke/sweep/ dispatch) pay the poll latency directly — ~20–30 s added per full smoke run at 30 fps. 60 favours the dev/test loop and max snappiness, is zero behavioural change from before, and is tunable down to 30 (documented power-saver) or off.
• Visible state. A 3-state dock indicator (Off / On (idle) / On · active · {fps} fps) plus an inline toggle (mirrors the audit-enabled toggle, but read/writes the EditorSetting) and activation/deactivation log lines. The wizard discloses the behaviour and flags that it lives in Editor Settings. No toast — EditorToaster is not reliable on 4.2, and a missable notice is worse than a steady indicator.

Considered alternatives

• Remove the boost and educate users to set the EditorSetting themselves (Option B). Rejected: the unfocused state is the common case for MCP, so B is laggy by default.
• One-time wizard sets the setting once, always-on (Option C). Rejected: always-on wastes background CPU even with no client connected; A boosts only while connected.
• Wizard gates A behind explicit consent (Option D). Folded into A — consent is delivered by the dock indicator + wizard note, with no blocking step (default is on).
• Per-instance backup (in user://…/project_instance_<hash>/). Rejected: the protected key is machine-global but user:// is per-project — narrower than the thing being backed up — so it leaves a persistent cross-project crash leak (true original siloed in a crashed project that may never reopen). The backup must be machine-wide.
• Reference-count active connections across instances to fix the transient concurrent de-boost. Rejected for 1.0: a refcount leaks on crash and leans on Windows dead-instance GC that OS.is_process_running reports unreliably. The residual transient dip (a disconnecting instance de-boosts a still-connected peer until its next fresh connect) is never persistent corruption and already existed before this change.
• WebSocket reader thread instead of the frame-rate lever. Rejected for this iter and filed post-1.0 (Plan/Ideas/PostRelease/2026-06-07-websocket-reader-thread-latency.md): it removes the 4-frame pickup batching but not the execution floor (command execution must marshal to the main thread at the main-loop cadence, and multi-yield mutations pay one frame per yield regardless). Complementary, not a replacement, and a sizeable threading project.

Known limitations (documented, not coded)

• Exact-B manual write is undetectable. If the user sets the key to exactly B while it is already B, the engine de-dupes the same-value write (editor_settings.cpp — changed only when p_value != old), emitting no settings_changed and recording nothing in get_changed_settings(), so no listener could catch it. On restore the toolkit treats it as its own and reverts to O.
• Transient concurrent de-boost. See the refcount note above — a brief responsiveness dip, never persistent corruption.

Full source-verified analysis (engine line references, blast radius, cost, rejected alternatives): Insights/unfocused-throttle-analysis.md. Grilling session: Plan/Reference/GrillingSessions/2026-06-07-unfocused-throttle.md. Supersedes the silent in-memory boost added in iter 28b (unfocused-throttle-fix).