[ata-sesli]

Summary

This moves the file-discovery module (src/discover/) from C to Zig, following the same approach as the earlier git, simhash and watcher port: the C headers (discover.h, userconfig.h) stay the interface, and the Zig files export the same cbm_* functions. No caller changes. After this PR, src/discover/ has no .c files.

Discovery should behave exactly as before: the same files are found, the same directories are skipped, and .gitignore, .git/info/exclude, .cbmignore and the global excludes file are applied the same way.

Commits

Each commit builds and passes the full test suite on its own.

1. refactor(foundation): move discover's stat and realpath helpers to compat_fs (C only). Moves the platform-specific file helpers out of discover.c unchanged, as cbm_stat_utf8, cbm_stat_no_follow and cbm_canonical_path in foundation/compat_fs.{h,c}. On Windows they use _wstat64, skip reparse points and call _fullpath; on POSIX they use stat/lstat/realpath. This keeps the Windows #ifdef code in C, so the Zig walk calls one helper on every platform.

2. refactor(discover): port gitignore, userconfig and language to Zig. Replaces gitignore.c, userconfig.c and language.c with .zig counterparts.

   • The language tables keep their entry order, because lookups stop at the first match.
   • userconfig.zig parses JSON through the same yyjson calls as before, so duplicate keys and parse errors are handled the same way.
   • scripts/generate-lang-code.py now prints Zig table rows.

3. refactor(discover): port the directory walk to Zig. Replaces discover.c with discover.zig. The port covers:

   • the skip lists and the ignore-file chain;
   • cache-directory pruning and the 512-frame walk stack;
   • resolving the git common dir, the enclosing repository and ancestor ignore files;
   • tracked-file lookup;
   • the CBM_DISCOVER_TEST_FAIL_PHASE/_NTH fault-injection hooks.

   It keeps the C version's fixed buffers and truncation behaviour. Arrays and strings returned to C use the C allocator, so cbm_discover_free and cbm_discover_free_excluded work unchanged.

Build wiring:

• build.zig and Makefile.cbm drop the C sources and list the Zig files in CBM_ZIG_SRCS.
• src/cbm_zig.zig runs checkExports on the new modules, so a signature that drifts from its C header fails the build.
• scripts/license-policy.json and CONTRIBUTING.md point at the new file names.

Differences from the C code

• cbm_gitignore_match_result is no longer exported. No header declared it; discover.c declared it locally. discover.zig now calls the matcher through a Zig import.
• cbm_gitignore_merge_dup_hook_for_test is kept as an exported variable. It is a test hook.
• excluded_strdup fault path: when that injected fault fires, the Zig code frees the copied string. The C code leaked it. This only runs under fault injection.
• Cache-path respelling: the C code guarded it with a non-Windows #ifdef. The Zig code keeps the same guard as if (!is_windows).

Verification

• Full test suite with CBM_TEST_TIERS=all (all language tiers on), macOS arm64: 6,447 passed on main (7484e7e) and 6,447 passed after each of the three commits.
• Old vs new CLI. main and this branch each indexed the same repositories, each into its own cache. The index_repository response (status, excluded dirs and counts), the sorted File paths and the node counts per label were identical in every case. The repositories were:

  • a scratch fixture that covers:

    • nested .gitignore files, including re-included files;
    • .git/info/exclude and .cbmignore;
    • core.excludesFile from ~/.gitconfig and from the XDG config dir;
    • a tracked file inside an ignored dir;
    • a linked worktree;
    • a subfolder without its own .git (exercising enclosing-repository lookup);
    • a symlinked repository root;
    • a plain folder with no git;
    • a cache dir inside the repository, reached both directly and through the symlink.

    It was indexed in full, fast and moderate modes.

  • this repository's source tree and a larger external repository, each in full and fast modes.

• Table parity. A script extracted all 14 tables from main's C sources and from the Zig sources (extensions, filenames, compound extensions, language names, tiers, user-config language names, skip dirs, suffixes, filenames, patterns, ignored JSON files) and compared them in order. All 14 were identical.
• Discover-phase time (warm runs): this repository took 7 ms before and 5 ms after. The larger repository took 16–18 ms both before and after.
• Cross-compile of libcbm_zig.a succeeded for x86_64 and aarch64 Linux (gnu), x86_64 Windows (gnu) and aarch64 macOS, and the discover exports were present in each archive.
• Header drift check: I temporarily changed a parameter type in discover.h. The build failed with cbm_discover_free: parameter 1 is c_int in Zig but c_long in its C header. The change was then reverted.
• Lint: clang-format-diff (clang-format 22.1.8) is clean on the changed C lines. zig fmt --check, git diff --check and tests/test_github_workflows.sh all pass.

Limits

• Windows was only cross-compiled. It was not run, because PR CI runs Linux amd64 only.
• Stale comments: a few comments still mention src/discover/discover.c. They are in tests/repro/*.c, src/pipeline/pass_pkgmap.c and one line of discover.h. They are left as they are to keep this diff focused on the port.