fix(windows): report why an atomic publish failed instead of blaming the repo

#1620: every index_repository run fails, every pass succeeds, the worker exits
0, no error-level line appears anywhere, and the user is told:

  Pipeline failed. Check repo_path exists and contains source files.

The repository is fine. The publish is failing on a filesystem permission
problem, and two separate defects conspired to hide it.

cbm_rename_replace discarded GetLastError() on Windows and returned a bare
CBM_NOT_FOUND. Callers then log errno - see finalize.rename_failed - so the one
field that should explain the failure held whatever an unrelated CRT call left
there. The Win32 error is now translated to errno, with ERROR_ACCESS_DENIED the
interesting case: MoveFileEx needs DELETE on the destination, which a cache file
created under an empty or foreign DACL does not grant.

cbm_pipeline_finalize_staged_generation returned PERSIST_FAILED on a failed
sidecar removal with no log at all. That is the silent path: a publish that
fails must say so.

Neither change fixes the underlying ACL problem. Together they turn a report
that misdirects the user at their own repository into one naming a permission
failure on a specific file - which is the difference between an unactionable
issue and a five-minute one.

platform + pipeline suites: 268 passed, 0 failed.

Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>

src/foundation/compat_fs.c

@@ -1039,9 +1039,54 @@ int cbm_rename_replace(const char *src, const char *dst) {
     wchar_t *wdst = cbm_path_to_wide(dst);
     int ret = CBM_NOT_FOUND;
     if (wsrc && wdst) {
-        ret = MoveFileExW(wsrc, wdst, MOVEFILE_REPLACE_EXISTING | MOVEFILE_WRITE_THROUGH)
-                  ? 0
-                  : CBM_NOT_FOUND;
+        if (MoveFileExW(wsrc, wdst, MOVEFILE_REPLACE_EXISTING | MOVEFILE_WRITE_THROUGH)) {
+            ret = 0;
+        } else {
+            /* Translate the Win32 error into errno so callers can report WHY.
+             *
+             * Callers log `errno` after a failed rename (see
+             * finalize.rename_failed in the pipeline). Without this the value
+             * is whatever happened to be left there by an unrelated CRT call,
+             * so on Windows the one field that should explain an atomic-publish
+             * failure was noise. #1620 is exactly that: an ACL problem surfaced
+             * to the user as "Pipeline failed. Check repo_path exists and
+             * contains source files" — blaming their repository — because
+             * ERROR_ACCESS_DENIED never reached the log.
+             *
+             * ERROR_ACCESS_DENIED is the interesting one here: MoveFileEx needs
+             * DELETE on the destination, which a cache file created under an
+             * empty or foreign DACL does not grant. */
+            DWORD error = GetLastError();
+            switch (error) {
+            case ERROR_ACCESS_DENIED:
+            case ERROR_WRITE_PROTECT:
+                errno = EACCES;
+                break;
+            case ERROR_FILE_NOT_FOUND:
+            case ERROR_PATH_NOT_FOUND:
+                errno = ENOENT;
+                break;
+            case ERROR_SHARING_VIOLATION:
+            case ERROR_LOCK_VIOLATION:
+            case ERROR_USER_MAPPED_FILE:
+                errno = EBUSY;
+                break;
+            case ERROR_NOT_SAME_DEVICE:
+                errno = EXDEV;
+                break;
+            case ERROR_DISK_FULL:
+                errno = ENOSPC;
+                break;
+            case ERROR_INVALID_NAME:
+            case ERROR_FILENAME_EXCED_RANGE:
+                errno = ENAMETOOLONG;
+                break;
+            default:
+                errno = EIO;
+                break;
+            }
+            ret = CBM_NOT_FOUND;
+        }
     }
     free(wsrc);
     free(wdst);

src/pipeline/pipeline.c

@@ -1806,6 +1806,15 @@ int cbm_pipeline_finalize_staged_generation(char *stage_path, const char *final_
     struct timespec t_fin;
     cbm_clock_gettime(CLOCK_MONOTONIC, &t_fin);
     if (cbm_remove_db_sidecars(stage_path) != 0) {
+        /* This returned PERSIST_FAILED with no log at all, which is how #1620
+         * presented: every pass succeeded, the worker exited 0, no error-level
+         * line was emitted anywhere, and the user was told "Pipeline failed.
+         * Check repo_path exists and contains source files" — pointed at their
+         * repository for a filesystem permission problem. A publish that fails
+         * must say so. */
+        char errno_text[16];
+        (void)snprintf(errno_text, sizeof(errno_text), "%d", errno);
+        cbm_log_error("finalize.sidecar_removal_failed", "errno", errno_text, "stage", stage_path);
         discard_generation_stage(stage_path);
         return CBM_PIPELINE_PERSIST_FAILED;
     }