fix(encoding): decode git's file names as Python does, and write LF (Windows); km 1.0.2

[pasrom]

Problem (reported from a Windows machine)

km validate over the whole repo crashed on Windows as soon as a tracked .md had a non-ASCII character in its name. km decoded git's output with the system code page (cp1252) while git writes file names as UTF-8, so künstler.md came back as a name that does not exist: FileNotFoundError before anything was checked. gen-index, demote and serve did not crash but silently skipped the file: gen-index reported "up to date" while the doc was missing from its list. Linux and macOS were fine, so CI never saw it.

Fix

• One helper, paths.git(), decodes git's output with the file-system encoding, the way Python turns file names into text: UTF-8 on Windows and macOS, the locale's encoding on Linux. Each listed name opens its file on all three. Undecodable bytes are kept (surrogateescape) instead of crashing; git writes its messages in the locale.
• validate lists files through common.tracked_md(), with git called once per run; a tracked doc deleted but not yet staged no longer crashes it.
• promote reads its child validator as UTF-8; console output escapes characters the console lacks.
• Every file km writes goes through paths.write_text(): UTF-8 with LF. On Windows km wrote CRLF, so generated _index.md files showed as changed.

Guards

• Tests run km with EncodingWarning as an error.
• tests/encoding_smoke.sh: a brain with an ü and an ő in its file names; gen-index lists them and writes LF, validate reports them, the work tree stays as committed. A new Windows CI job runs it (the release waits for it); team_smoke runs it under the macOS locales that behave like Windows.
• On Linux a non-UTF-8 locale changes the file-name encoding too. The smoke job generates a Latin-1 locale and a test checks a non-ASCII name is still validated there.
• A git error in a Latin-1 locale is reported, not a crash.

Test plan

• Reproduced first: crash with a Latin-1 locale, clean with UTF-8
• Suites 30/30, 30/30, 92/92 + encoding 4/4 locally; CI green on Linux and Windows
• The Windows job bites: on the old code it fails through the EncodingWarning guard, and without the guard because gen-index leaves the non-ASCII file out (throwaway PRs throwaway: windows job without the fix #7, throwaway: windows job on the old code, no guard #8, closed)
• Mutations: decoding as cp1252 fails encoding_smoke; removing console escaping fails it; removing the existence filter fails the deleted-doc tests
• Linux Latin-1 (Docker): decoding as UTF-8 regardless of platform, the state before review, fails the new test; the fs-encoding decoding and the old code pass
• /simplify and /code-review (high) on the final diff, findings folded in; write_jsonl building the bundle in memory kept (a brain's served bundle is small)

Version bump to 1.0.2: merging is the release.