fix: preserve split UTF-8 across PersistentTerminal feed() calls

[aarcamp]

When reading from a PTY, byte chunks arrive at arbitrary boundaries. Multi-byte UTF-8 characters (emoji = 4 bytes, CJK = 3 bytes, accented chars = 2+ bytes) can be split across consecutive read() calls. The old code in PersistentTerminal.feed():

// OLD - each call decoded independently
str = data.toString("utf-8")      // Buffer path
str = new TextDecoder("utf-8").decode(data)  // Uint8Array path

Both of these treat each chunk as a complete, self-contained UTF-8 sequence. When a multi-byte character is split (e.g., the first 2 bytes of a 4-byte emoji in one call, the remaining 2 in the next), each independent decode produces U+FFFD replacement characters (�) instead of the intended character.

[remorses]

is there a way to remove hasBinaryDecoderState state and use the decoder directly? cc @codex review

[aarcamp]

is there a way to remove hasBinaryDecoderState state and use the decoder directly? cc @codex review

Yes, it's an optimization. We could swap it out for simpler logic as follows -- it just means some extra TextDecoder() allocations in certain conditions:

diff --git a/src/ffi.ts b/src/ffi.ts
index 79bc6a3..6a4244c 100644
--- a/src/ffi.ts
+++ b/src/ffi.ts
@@ -255,7 +255,6 @@ export class PersistentTerminal {
   private _rows: number
   private _destroyed = false
   private _streamDecoder = new TextDecoder("utf-8")
-  private _hasBinaryDecoderState = false

   constructor(options: PersistentTerminalOptions = {}) {
     if (!native) {
@@ -297,14 +296,9 @@ export class PersistentTerminal {
     this.assertNotDestroyed()

     if (typeof data === "string") {
-      if (this._hasBinaryDecoderState) {
-        const flushed = this._streamDecoder.decode()
-        this._streamDecoder = new TextDecoder("utf-8")
-        this._hasBinaryDecoderState = false
-        if (flushed.length > 0) {
-          native!.feedTerminal(this._id, flushed)
-        }
-      }
+      // Discard any partial UTF-8 bytes from prior binary feeds before
+      // crossing into a plain string boundary.
+      this._streamDecoder = new TextDecoder("utf-8")

       if (data.length > 0) {
         native!.feedTerminal(this._id, data)
@@ -312,7 +306,6 @@ export class PersistentTerminal {
       return
     }

-    this._hasBinaryDecoderState = true
     const decoded = this._streamDecoder.decode(data, { stream: true })
     if (decoded.length > 0) {
       native!.feedTerminal(this._id, decoded)
@@ -337,7 +330,6 @@ export class PersistentTerminal {
     this.assertNotDestroyed()
     native!.resetTerminal(this._id)
     this._streamDecoder = new TextDecoder("utf-8")
-    this._hasBinaryDecoderState = false
   }

   /**
@@ -414,7 +406,6 @@ export class PersistentTerminal {
     if (this._destroyed) return
     this._destroyed = true
     this._streamDecoder = new TextDecoder("utf-8")
-    this._hasBinaryDecoderState = false
     native!.destroyTerminal(this._id)
   }

I can either push this update, or add comments to the existing PR so it's clearer what's happening, just lmk. :)

[remorses]

Yeah I like this one better. I will publish after you push this commit