[NaNIsaac]

Adds two subcommands for Korean text that arrives as escaped UTF-8 bytes in file names and URLs:

hanfix inspect _ED_8A_B8_EB_A6_AC                     report: type, status, reason, candidate
hanfix decode --type underscore-hex _ED_8A_B8_EB_A6_AC   트리
hanfix decode --type percent %ED%8A%B8%EB%A6%AC          트리
printf '%s' '_ED_8A_B8' | hanfix decode --type underscore-hex

CLI compatibility

inspect and decode are subcommands only as the first argument. Anything else is unchanged:
hanfix -- inspect still converts the word, and the shell functions call hanfix --to en -- $last, so hf
and its Korean twin never reach the new code. tests/cli.rs covers both, plus hanfix ㅑㅜ넫ㅊㅅ → inspect.
The only way an existing call changes is a bare hanfix inspect / hanfix decode meant as English keys to
convert into Hangul; hanfix -- inspect does that now.

Contract (tested)

• underscore-hex: _ + two upper-case hex digits, bytes 0x80-0xFF only. backup_2026_10_29, hello_bed,
  hello_world and the first _ of report__ED_8A_B8.pdf are kept; __ is never a space.
• underscore-hex runs also end at a byte that cannot start a UTF-8 character (80-BF, C0, C1, F5-FF); that _XX
  stays text and scanning goes on: IMG_8123.JPG, DSC_9001.JPG, test_99.txt unchanged with exit 0,
  _ED_8A_B8_99.txt -> 트_99.txt, IMG_8123__ED_8A_B8.jpg -> IMG_8123_트.jpg. Inside a character every
  escape is taken, so _ED_8A (incomplete), _ED_C3_A9 and _DB_backup (_DB is a lead byte) still fail.
  Percent is unchanged: every %XX is an escape.
• Error messages show at most the first 40 characters of the failing run, plus its length.
• --type given twice is a usage error.
• percent: %XX, any byte, either case. % without two hex digits is text (100% done, %GG). + is kept.
• Consecutive escapes are one byte sequence and must be complete, valid UTF-8. Otherwise decode exits 1,
  writes nothing to stdout and one hanfix: error: ... line to stderr. No U+FFFD substitution. The whole
  text is checked before anything is printed.
• Text with no escapes: decode prints it unchanged, exit 0.
• inspect exits 0 on any text (plain, invalid, mixed); valid / ambiguous (both kinds present) /
  invalid, with a rule-based reason, no confidence numbers.
• Input on stdin is read strictly as UTF-8 for both subcommands (exit 1 otherwise); the existing conversion
  keeps its lossy read.
• Usage errors (missing or unknown --type, unknown option) exit 2, as before.

Files

• src/encoding.rs: detection, validation and decoding as pure functions with unit tests; no new crates.

• src/main.rs: subcommand dispatch at the top of run(), help and usage text, read_stdin() shared with
  the existing path. usage_error uses eprintln! (same output) so cargo clippy -- -D warnings passes.

• tests/cli.rs: runs the built binary: stdout/stderr/exit contract, stdin, usage errors, subcommand words
  as text, existing conversions.

• README.md, README.ko.md: new "Decode escaped Korean text" section, usage, --type, exit status, and a
  PowerShell Rename-Decoded function that previews old -> new by default and renames only with
  -Apply, skipping names that fail to decode, do not change, are not valid file names, or would overwrite.

• tests/shells.ps1: extracts that README block, checks the Korean README has the same code, and runs it
  on a temp folder (preview changes nothing; -Apply renames only the decodable, free name).

• Rename-Decoded first probes hanfix decode --type percent -- %41 and throws hanfix too old: update with hanfix --update unless it prints A: hanfix 0.2.7 and older read decode ... as text and exit 0, so without the probe the example would rename every file. tests/shells.ps1 checks this with a stand-in for the old binary (a function calling hanfix -- @args); nothing is renamed.

• inspect adds Note: lower-case _xx is not treated as an escape for runs such as _ed_8a_b8.

• Rename-Decoded also skips results that are more than one line or contain control characters (no CR to LF
  rewrite), end in a dot or space (Windows would drop them), or are a Windows device name; it collects every
  name in the folder and every name planned in the run, ignoring case, so the preview lists exactly what
  -Apply does. Tested in tests/shells.ps1 (percent folder: preview equals apply), also run in Windows
  PowerShell 5.1.

• Version 0.2.8.

Not in this PR: mojibake (planned next) and the release.

local-ci: 7504d53 | cargo test --release --locked (10 unit + 7 CLI pass), cargo clippy --all-targets --locked -- -D warnings clean, pwsh tests/shells.ps1 with this build first on PATH (3 ok), README example in Windows PowerShell 5.1 (preview correct) on Windows | not run: tests/shells.sh (bash/zsh/fish, Linux only; init scripts untouched), macOS, Termux