Commit the limiter's audible state on the click and defer only the host's latency restart

setInstrumentParams now arms a sticky pending restart that flushLatencyRestart drains from the sync tick; setState and the bake's adopt flush at their own tails.

src/shell/instrument/CLAUDE.md

@@ -107,7 +107,7 @@ declared ahead of the instrument slots at that member in `reasampler_processor.h
 ## Modules
 
 - `reaper_bridge` — READ-ONLY bank consumer: receives bank snapshots from the extension and exposes them as a read-only view. **Never writes to the extension's bank** — this is a load-bearing invariant; no mutation path exists in this module. It owns TWO prefix-guarded ext-state write entry points, `writeUsageExtState` (`rsusage_`) and `writeBakeExtState` (`rsbake_`), each refusing every other key; neither weakens the read-only-*bank* invariant, because neither payload is bank state and `banks`/`view`/`tail`/`assign` stay structurally unwritable. Both PROVE the write by reading the key back (`wire::extStateWriteLanded`) — `SetProjExtState`'s own return cannot speak for one key, so testing it was a guard that could never fire, and the bake's "could not publish" refusal was consequently unreachable. It also owns the bake crossing — `extensionActionAvailable` / `invokeExtensionAction` (`NamedCommandLookup` + `Main_OnCommandEx` with `getReaperParent(3)`, the instance's OWN project tab, as `proj` — a request, not a DAW-verified guarantee; see the header) and `projectTempoBpm`.
-- `reasampler_processor` (`shell/instrument/`: `reasampler_processor.cpp` lifecycle + `process()`, `processor_state.cpp` component-state I/O + UI-thread parameter accessors, `processor_reload.cpp` the off-audio-thread `reloadInstrument`/publish family — Q-W2v, T4-12 split; `process()` and its per-block work stay ONE TU on purpose, no cross-TU call on the per-sample path) — VST3 `SingleComponentEffect` shell: declares event-input bus + **permanently stereo** output (GA fix: dynamic mono↔stereo bus renegotiation deleted; `ChannelMode` is now decode-only), marshals MIDI note-on/off into the VoiceEngine, renders audio; owns off-audio-thread `reloadInstrument` + atomic pointer swap so `process()` does no allocation, no file I/O, no bridge calls. The instance state is `{loaded capture id, one InstrumentParams}`, and `reloadInstrument` resolves + decodes exactly that one capture into the `SampleData` the engine plays. **Self-contained playback (pS):** `ComponentState` v10 adds a `SampleRefs` table — per referenced sample, a project-relative path + decode intrinsics (root, loop, channels, displayName); `reloadInstrument` decodes directly from `SampleRefs`, bank-free (plays with the extension absent). The bank/bridge is a browser source: loading a capture copies its reference in; the reopen-heal timer + poll-to-play apparatus are removed. `retireIdleDrain()` retires fully-idle drain snapshots on the UI-timer cadence. Voice-param edits (`setVoiceCount`/`setVoiceMode`/`setMonoTrigger`) rebuild the engine from the already-decoded `SampleData` via the drain-slot swap — no bank re-read, no WAV re-decode, no audible cut to ringing tails. **FB1:** applies the post-mixer `masterGainLinear` (from `ComponentState` v8) as a per-sample ramp over the summed output — no zipper noise. **GA v9:** `channelModeExplicit_` flag persisted; `channelModeFor()` auto-defaults the mode from the loaded capture's channel count when the flag is not set. **pS:** `ComponentState` bumped v9→v10 (`SampleRefs` table); pre-v10 blobs lift to empty refs and re-save self-contained. **pS-usage:** publishes instance usage (held `SampleRefs` paths) to `rsusage_<instanceGuid>` at the tail of `reloadInstrument` (off audio thread) via `reaper_bridge::writeUsageExtState`; `ComponentState` bumped v10→**v11** (`instanceGuid` field); pre-v11 blobs mint guid on first publish. **The master bus:** the summed output runs `voice mixer → master gain → limiter (core/instrument/engine/limiter) → output bus`, with the meter tapped at the bus output POST-limiter and published as relaxed atomics — per-channel peak and smallest limiter gain ACCUMULATED across every block since the UI last read, plus the latched clip (the meter Gotcha below owns why). The limiter's enable is persisted in the parameter set (params payload v15) and mirrored onto the audio thread by `setInstrumentParams`, the single funnel every writer already goes through. That mirror is also what `getLatencySamples()` answers from — the plugin's FIRST latency reporting: 0 bypassed, the lookahead engaged. `setLimiterEnabled` requests the host's `restartComponent(kLatencyChanged)`, UI thread only and never from `process()`; it is a LATENCY restart with the bus untouched, NOT the retired per-mode `kIoChanged` bus renegotiation the invariant above forbids.
+- `reasampler_processor` (`shell/instrument/`: `reasampler_processor.cpp` lifecycle + `process()`, `processor_state.cpp` component-state I/O + UI-thread parameter accessors, `processor_reload.cpp` the off-audio-thread `reloadInstrument`/publish family — Q-W2v, T4-12 split; `process()` and its per-block work stay ONE TU on purpose, no cross-TU call on the per-sample path) — VST3 `SingleComponentEffect` shell: declares event-input bus + **permanently stereo** output (GA fix: dynamic mono↔stereo bus renegotiation deleted; `ChannelMode` is now decode-only), marshals MIDI note-on/off into the VoiceEngine, renders audio; owns off-audio-thread `reloadInstrument` + atomic pointer swap so `process()` does no allocation, no file I/O, no bridge calls. The instance state is `{loaded capture id, one InstrumentParams}`, and `reloadInstrument` resolves + decodes exactly that one capture into the `SampleData` the engine plays. **Self-contained playback (pS):** `ComponentState` v10 adds a `SampleRefs` table — per referenced sample, a project-relative path + decode intrinsics (root, loop, channels, displayName); `reloadInstrument` decodes directly from `SampleRefs`, bank-free (plays with the extension absent). The bank/bridge is a browser source: loading a capture copies its reference in; the reopen-heal timer + poll-to-play apparatus are removed. `retireIdleDrain()` retires fully-idle drain snapshots on the UI-timer cadence. Voice-param edits (`setVoiceCount`/`setVoiceMode`/`setMonoTrigger`) rebuild the engine from the already-decoded `SampleData` via the drain-slot swap — no bank re-read, no WAV re-decode, no audible cut to ringing tails. **FB1:** applies the post-mixer `masterGainLinear` (from `ComponentState` v8) as a per-sample ramp over the summed output — no zipper noise. **GA v9:** `channelModeExplicit_` flag persisted; `channelModeFor()` auto-defaults the mode from the loaded capture's channel count when the flag is not set. **pS:** `ComponentState` bumped v9→v10 (`SampleRefs` table); pre-v10 blobs lift to empty refs and re-save self-contained. **pS-usage:** publishes instance usage (held `SampleRefs` paths) to `rsusage_<instanceGuid>` at the tail of `reloadInstrument` (off audio thread) via `reaper_bridge::writeUsageExtState`; `ComponentState` bumped v10→**v11** (`instanceGuid` field); pre-v11 blobs mint guid on first publish. **The master bus:** the summed output runs `voice mixer → master gain → limiter (core/instrument/engine/limiter) → output bus`, with the meter tapped at the bus output POST-limiter and published as relaxed atomics — per-channel peak and smallest limiter gain ACCUMULATED across every block since the UI last read, plus the latched clip (the meter Gotcha below owns why). The limiter's enable is persisted in the parameter set (params payload v15) and mirrored onto the audio thread by `setInstrumentParams`, the single funnel every writer already goes through. That mirror is also what `getLatencySamples()` answers from — the plugin's FIRST latency reporting: 0 bypassed, the lookahead engaged. The host's `restartComponent(kLatencyChanged)` is issued by `flushLatencyRestart` alone, UI thread only and never from `process()`; it is a LATENCY restart with the bus untouched, NOT the retired per-mode `kIoChanged` bus renegotiation the invariant above forbids. Why it is split from the commit is the Gotcha below.
 - `reasampler_editor` — VST3 `IPlugView` LICE editor shell: hosts a LICE-drawn child window; the Sample face is home and Browse is a modal picker over it. Split on the Sample face's BAND axis, mirroring the pure `sample_bands` allocator: `editor_session` (session/bridge state, caches, commit-and-reload), `editor_controls` (the ONE `faceLayout` band resolve every paint and hit-test path shares, the node-drag bounds, the value labels, and the per-instance controls the parameter set does not carry — the parameter-set binding itself is the pure `core/instrument/ui/deck_values` module this only adapts int ids onto), `editor_models` (the orthogonal half: which stored struct each transient editor selection names — the staged-envelope pack/unpack, the drawn contour, and the three velocity curves), then matching paint and input sets — `editor_paint`/`editor_input` (dispatch + drag router + hover dispatch), `_chrome`, `_waveform`, `_deck` — plus the two band-independent surfaces (`_browse` for the modal picker, `_curve` for the velocity-curve popup) and `editor_platform` (IPlugView/Win32 window plumbing). Shared internals in `editor_internal.h`, no TU of its own. Drop-onto-editor ingest is NOT shipped (deferred).
 - `reasampler_embed` — implements `IReaperUIEmbedInterface` so the instrument draws inline in the TCP/MCP without a plugin-owned HWND; delegates layout to `embed_strip`. A read-only readout: the loaded capture across the keyboard span with its root marked, plus the activity level. It takes no mouse input (there is nothing on the strip to select).
 - `editor_stroke` — the editor's LICE side of the analytic stroker: builds a coverage mask with the pure `core/ui/stroke_aa` and blends it into the bitmap ONCE, writing straight to the bitmap's bits (the arithmetic matches LICE's own mode-0 combine, so a stroke composites identically to every other kit draw). Every radial and spline stroke on the editor routes through `strokeArcAA` / `strokePolylineAA` / `strokeLineAA`. Holds the draw-thread-only scratch mask and arc point list — reuse, not a hidden dependency: threading a canvas through the eight paint sites would grow those signatures to carry an allocation detail. Deliberately does NOT touch `shell/panel/draw_kit`: the waveform stroke, the docked bank panel and the browse cards are out of this seam's blast radius.
@@ -125,16 +125,22 @@ declared ahead of the instrument slots at that member in `reasampler_processor.h
   the very instance whose frame is on the stack. Deferring by one tick is same-thread and
   in-instance — it is NOT a cross-process poller/nonce handshake, and it must not grow into
   one.
-- **The limiter toggle arms on the same principle, and for the same reason.** Its commit
-  requests the host's `restartComponent(kLatencyChanged)`; a host that services that
+- **The limiter toggle splits its commit: the sound is inline, the HOST NOTIFICATION arms.**
+  Its commit needs the host's `restartComponent(kLatencyChanged)`; a host that services that
   synchronously runs `setActive(false)`/`setActive(true)`, and OUR `setActive(true)` calls
-  `reloadInstrument()` — a WAV re-decode plus disk I/O. Inline from `WM_LBUTTONDOWN` that
-  whole cycle runs with `SetCapture` held. The click writes the editor's own snapshot and
-  paints at once; the sync tick calls `setLimiterEnabled`, so **the audio and the reported
-  latency follow the click by up to one tick.** It sits AFTER the drag guard with the bake:
-  the restart rebuilds the instance, which mid-drag would yank the edit surface exactly as a
-  reload would. Every other writer of the parameter set (`setState`, the bake's adopt) still
-  commits and restarts immediately — none of them is inside a mouse handler.
+  `reloadInstrument()` — a WAV re-decode plus disk I/O, which inline from `WM_LBUTTONDOWN`
+  would run nested in a mouse handler. So the click commits the parameter set, the audio-thread
+  mirror and the latency reader at once, and `setInstrumentParams` only ARMS a pending restart
+  that `flushLatencyRestart` delivers. The editor's sync tick is the general drain and sits
+  AFTER the drag guard with the bake (the restart rebuilds the instance, which mid-drag would
+  yank the edit surface exactly as a reload would); `setState` and the bake's adopt flush at
+  their own tails, because they can commit with no editor open. The arm is a sticky bool, so
+  toggling twice inside one tick still costs exactly one restart.
+  **The residual:** between the commit and the flush the host's delay compensation is out of
+  step with the plugin by `limiterLookaheadSamples` (2 ms — `round(0.002 · rate)`, the
+  detector's 4-sample group delay INSIDE that budget, not on top), bounded by one 500 ms tick.
+  Narrowing it further means a second deferral mechanism (a posted window message) rather than
+  the tick — deliberately not built.
 - **The MASTER meter's ballistics ride the sync tick, and that tick is 500 ms.** They run
   BEFORE the tick's in-flight-drag guard on purpose — a drag suppresses the reload poll, but
   the bus keeps sounding. Elapsed time is measured (`GetTickCount64`), never assumed from the