Match external config semantics exactly

When reading, writing, or transferring configuration governed by an external format or tool (e.g. git’s config files), replicate that tool’s full semantics rather than the common-case behavior, and cover each semantic edge case with a dedicated test. Config-file parsing has subtle rules that are easy to miss on a first pass: repeated keys must resolve using the tool’s own precedence (e.g. “last assignment wins” for git’s scalar keys, not “first match”), escaping/quoting rules must be applied uniformly and exactly as the source tool defines them (git decodes \\, \", \n, \t, \b regardless of quoting, with quoting only affecting whether whitespace/comment characters are literal), and encoding artifacts such as a leading UTF-8 BOM must be stripped exactly as the authoritative parser does. Similarly, when code copies, detaches, or otherwise transfers a configuration-bearing structure (e.g. splitting a submodule’s git directory from its superproject), ensure every relevant config file is carried over — not just the primary one (e.g. config but also config.worktree, or remote/branch settings) — so no setting is silently dropped. Each of these edge cases should land with its own named regression test (e.g. core_worktree_repeated_key_honors_last_occurrence, core_worktree_decodes_unquoted_escapes, core_worktree_skips_a_leading_utf8_bom) so future refactors can’t reintroduce the gap.