fix(mirror): stop a split multi-byte character corrupting mirrored output

[clawedassistant26]

The bug

followFile hands the mirror whatever bytes a poll tick happened to catch, then decodes each slice on its own with toString("utf8"). That boundary lands wherever it lands, and regularly that is in the middle of a multi-byte character. Each half decodes to U+FFFD, so the character is destroyed in both directions.

Two boundaries do it:

• the 65536-byte read slice inside a single tick (Buffer.allocUnsafe(Math.min(65536, size - offset)))
• the gap between two ticks, when script -f flushes part of a character and the rest lands after the next poll

This matters because of what engines actually print. A framed TUI is built out of box-drawing characters (│ ┌ ─ ┤ └), and every one of those is three UTF-8 bytes. Any mirrored session whose output ran past the read slice showed a wall of replacement characters where the frame should be.

Reproduced on unmodified main

x * 65535 followed by ─ (U+2500, bytes e2 94 80), so the character's first byte is the last byte of read one:

slices        = 2
round-trips   = false
U+FFFD count  = 3
tail          = "xxxxx���done"

And split across two ticks, héllo written as h + first byte of é, then the rest:

got: "h��llo"

Both round-trip exactly after the fix.

The fix

Decode through StringDecoder instead of per-slice toString. It holds an incomplete tail back until the rest of it arrives. Three small things go with it:

• the decoder is reset on resync, because a half-character held back belongs to the transcript that was just replaced
• the decoder is flushed in stop(), so a genuinely truncated tail (killed child) is still reported rather than swallowed
• onChunk is skipped when a slice decodes to nothing, which also means the first flag in engines.mjs still marks the real opening slice

21 insertions, 3 deletions, and 9 of the added lines are comment.

Tests

New test/pty-utf8-split.test.mjs, 9 tests.

4 are the bug (fail before, pass after): a character straddling the 64KB boundary; a character split across two ticks; a box-drawing UI larger than the read slice; and astral characters (4-byte emoji, 2 code units) with the pad length swept so the boundary walks through every byte of the character.

5 are controls that pass both ways, so the fix cannot buy clean decoding by dropping or reordering output: ASCII still streams in order and drains on stop; a truncated tail is still emitted and not swallowed; a replaced transcript still resyncs; a file that does not exist yet is still picked up; stop() is still idempotent.

Fail-before verified with git checkout -- src/pty.mjs: 4 fail / 5 pass unpatched, 9/9 patched. Full suite 385 pass, 0 fail.

Deliberately not in this PR

• stripScriptBanner has the same class of split. The header regex is anchored to the start of a slice, so if script(1)'s banner is ever cut in half by a tick, neither half matches and the bookkeeping line reaches the mirror. Real, but it needs a small buffering decision about how long to hold the opening slice back, which is a design call rather than a fix.
• Rotation reads through a stale fd. On resync the offset resets and statSync sizes the new file, but readSync still uses the fd opened on the old inode. Reopening on resync would fix it. Separate defect, and a transcript in a per-session temp dir is unlikely to rotate.

Happy to do either if you want them.