fix: make the Read hook work on Windows

[learte4]

On Windows the plugin silently records nothing. The hook runs, exits 0, and never writes a line — so .skill-observer/logs/skills.jsonl stays empty forever and there is no error to point at.

Two causes, both in main() of scripts/skill-tracker.sh:

1. jq.exe emits CRLF. Every value read out of the jq -r block carries a trailing \r, so tool_name is "Read\r". The [[ "${tool_name}" == "Read" ]] guard never matches and the script exits before reaching any logging code.
2. file_path arrives with backslashes (D:\proj\.claude\skills\foo\SKILL.md), which never match the *".claude/skills/"* pattern in the guard below it.

Either one alone is enough to kill logging; together they make it look like the hook simply isn't registered.

Fix

Strip the CR from the five values, and normalize the separators before the path guard:

tool_name="${tool_name%$'\r'}" file_path="${file_path%$'\r'}"
session_id="${session_id%$'\r'}" offset="${offset%$'\r'}" limit="${limit%$'\r'}"
[[ "${tool_name}" == "Read" ]] || exit 0
file_path="${file_path//\\//}"

Both are no-ops on macOS and Linux: there is no CR to strip and no backslash to replace. ${var%$'\r'} removes at most one trailing CR and leaves LF-only values untouched.

Reproducing

On Windows with Git Bash, read any SKILL.md under .claude/skills/ with the hook installed. Before this change the log file is never created; after it, the skill_loaded event appears as expected.

I have been running this patch locally since 2026-07-17 across a few dozen sessions, which is how the second issue in the pair surfaced — fixing only the CRLF gets you past the tool_name guard and straight into the backslash one.

Related: I opened a separate PR for an unrelated performance issue in the same function. The two touch different lines a few apart; whichever lands first, the other needs at most a trivial rebase.