Windows console write uses wrong low surrogate range (0xDCEE instead of 0xDC00)

[SebTardif]

Bug

In library/std/src/sys/stdio/windows.rs, the low surrogate detection range is 0xDCEE..=0xDFFF instead of the correct 0xDC00..=0xDFFF.

This misses surrogates 0xDC00 through 0xDCED, causing:

• The trailing low surrogate of a split surrogate pair to not be flushed
• The UTF-8 byte count calculation to be wrong (counts 3 bytes instead of 1 for the low surrogate portion)
• The Write trait to report fewer bytes written than consumed

Affects supplementary Unicode characters (emoji, CJK Extension B, mathematical symbols) written to the Windows console when the write buffer splits mid-surrogate pair.

Fix

Change both instances of 0xDCEE to 0xDC00 to cover the full low surrogate range.

[SebTardif]

Upstream Context

• Windows console Stdin::read_vectored panics with a small first buffer and a pending surrogate rust-lang/rust#158096 (OPEN): "Windows console Stdin::read_vectored panics with a small first buffer and a pending surrogate." Related surrogate handling bug in the same Windows console code path.
• Reading from and writing to the Windows console should be lossy rust-lang/rust#116871 (OPEN): "Reading from and writing to the Windows console should be lossy." Proposes making console I/O lossy instead of erroring on non-UTF-8.
• Insufficient test coverage for Stdin::Read on Windows "console mode" rust-lang/rust#93055 (OPEN): "Insufficient test coverage for Stdin::Read on Windows console mode."
• stdio: Handle unicode boundaries better on Windows rust-lang/rust#23344 (CLOSED): "stdio: Handle unicode boundaries better on Windows." Original tracking issue for this class of bugs.

Our finding (#377) is specifically about the write path using 0xDCEE instead of 0xDC00 as the low surrogate range start when splitting surrogate pairs across console writes. This corrupts all characters above U+FFFF (emoji, CJK supplementary, mathematical symbols) on Windows console output.

[SebTardif]

Upstream context and cross-reference

This is a typo in library/std/src/sys/stdio/windows.rs where 0xDCEE is used instead of 0xDC00 as the low surrogate range boundary. The correct range is U+DC00..=U+DFFF, as confirmed by library/core/src/wtf8.rs (line 95: trail @ 0xDC00..=0xDFFF) and library/core/src/char/methods.rs (line 2571: 0xDC00).

The bug was likely introduced in rust-lang/rust PR #91754 by Patrick-Poitras (merged Jan 2022), which added the surrogate handling logic for Windows console read(). The 0xDCEE value appears to come from the UTF-16 encoding of U+120EE (CUNEIFORM SIGN GABA), whose surrogate pair is D808 DCEE. A developer likely used the low surrogate from a specific test character as the range start. The code was later refactored by @thomcc in PR #101193 and by @ChrisDenton in PR #134534, but neither caught the wrong constant.

Impact: When WriteConsoleW performs a partial write splitting a surrogate pair, low surrogates in 0xDC00..0xDCED (238 out of 1024 possible values, ~23% of supplementary-plane characters) are mishandled: the orphaned surrogate is not flushed, and the byte count overestimates by 2, desynchronizing the caller's buffer position.

No upstream issue or fix PR exists for this bug.

Downstream: The yazi terminal file manager copied this code verbatim into yazi-tty/src/windows.rs and has the same 0xDCEE bug.

Fix with regression test: PR #383

[joboet]

Upstream fix: rust-lang#158128