fix: filter precedence, files-from errors, NUL/CRLF and filter-rule rejections

Address an independent c-review of the files-from/filter feature:

- .rsync-filter precedence now matches rsync: evaluate the innermost
  (current) directory's rules first, then ancestors, then the command-line
  base (--filter/-C), so a deeper file's '+' can re-include what a shallower
  '-' excluded (regression tests in both scan modes; single-thread and -m).
- --files-from: a listed entry missing on disk and an empty list are now hard
  errors surfaced pre-transfer in send_files, send_files_multithreaded,
  dry-run and --list-only; '.' (whole tree) and empty listed dirs stay valid.
- scanner_path_relative now handles a transfer root of / (previously the
  scanner aborted on children of /).
- Reject unsupported rsync filter syntax explicitly (no silent no-ops):
  +/- modifiers other than '/' (! C s r p x) and rules beginning with ':'/'.'
  /'!' (merge/dir-merge/list-clear shorthands). Docs updated.
- -0/--from0 NUL mode preserves entry bytes (no CR/LF trimming); only newline
  mode trims. Absolute-entry error message no longer includes the newline.
- --no-from0/--no-cvs-exclude registered as negatable booleans.
- RSYNC_COMPAT rows updated for the precedence, rejection list, NUL-mode
  detail and the documented O(entries x files) scalability bound of the
  allow-set (Summary unchanged: 62/3/5/1/76 = 147).

src/shared/file_list.c

@@ -39,16 +39,19 @@ static bool string_list_add(StringList* list, const char* text) {
 /* Validate and normalize one entry. Returns:
  *   1 -> added to `out`
  *   0 -> blank entry, skip
- *  -1 -> invalid (message set in `err`) */
-static int normalize_entry(const char* raw, size_t len, StringList* out, char* err,
-                           size_t err_size) {
-  /* Trim the trailing newline/carriage-return from line mode. */
-  while (len > 0 && (raw[len - 1] == '\n' || raw[len - 1] == '\r'))
-    len--;
+ *  -1 -> invalid (message set in `err`)
+ * `strip_line_endings` trims a trailing CR/LF (line mode only); NUL mode keeps
+ * the entry bytes verbatim so names ending in CR/LF survive. */
+static int normalize_entry(const char* raw, size_t len, bool strip_line_endings, StringList* out,
+                           char* err, size_t err_size) {
+  if (strip_line_endings) {
+    while (len > 0 && (raw[len - 1] == '\n' || raw[len - 1] == '\r'))
+      len--;
+  }
   if (len == 0)
     return 0;
   if (raw[0] == '/') {
-    snprintf(err, err_size, "absolute path entries are not allowed: '%s'", raw);
+    snprintf(err, err_size, "absolute path entries are not allowed: '%.*s'", (int)len, raw);
     return -1;
   }
   /* Reject NUL bytes inside a token defensively (NUL-delimited mode splits on
@@ -134,7 +137,7 @@ FileListSet* file_list_load(const char* path, bool null_separated, char* err, si
   bool ok = true;
   char delim = null_separated ? '\0' : '\n';
   while (ok && (n = getdelim(&line, &line_cap, delim, fp)) != -1) {
-    int r = normalize_entry(line, (size_t)n, &raw, err, err_size);
+    int r = normalize_entry(line, (size_t)n, !null_separated, &raw, err, err_size);
     if (r < 0) {
       ok = false;
       break;

src/shared/filter.c

@@ -19,6 +19,13 @@ static bool rule_text_is_unsupported_word(const char* p, size_t len) {
   return false;
 }
 
+/* rsync include/exclude rule modifiers we do NOT implement. A rule whose +/- is
+ * immediately followed by one of these is rejected instead of being silently
+ * parsed as a literal pattern. */
+static bool is_unsupported_rule_modifier(char c) {
+  return c == '!' || c == 'C' || c == 's' || c == 'r' || c == 'p' || c == 'x';
+}
+
 FilterRule* filter_rule_parse(const char* line, char* err, size_t err_size) {
   if (err && err_size > 0)
     err[0] = '\0';
@@ -34,7 +41,6 @@ FilterRule* filter_rule_parse(const char* line, char* err, size_t err_size) {
   while (len > 0 && (text[len - 1] == '\n' || text[len - 1] == '\r'))
     text[--len] = '\0';
 
-  FilterAction action = FILTER_ACTION_EXCLUDE;
   const char* p = text;
   while (*p == ' ' || *p == '\t')
     p++;
@@ -44,11 +50,35 @@ FilterRule* filter_rule_parse(const char* line, char* err, size_t err_size) {
     return NULL;
   }
 
+  FilterAction action = FILTER_ACTION_EXCLUDE;
   if (*p == '+' || *p == '-') {
     action = *p == '+' ? FILTER_ACTION_INCLUDE : FILTER_ACTION_EXCLUDE;
     p++;
-    /* Accept the rsync word forms include/exclude. */
+    /* rsync attaches rule modifiers directly to the +/- (e.g. "-s foo"). Only
+     * the '/' anchor modifier is supported; anything else is a clear error
+     * rather than a silently-ignored literal. */
+    if (*p != ' ' && *p != '\t' && *p != '\0' && is_unsupported_rule_modifier(*p)) {
+      snprintf(err, err_size,
+               "filter rule modifier '%c' is not supported (only the '/' anchor after +/- "
+               "is implemented; put a space between +/- and the pattern)",
+               *p);
+      free(text);
+      return NULL;
+    }
+    while (*p == ' ' || *p == '\t')
+      p++;
   } else {
+    /* ':' (dir-merge) and '.' (merge) are rsync filter-rule shorthands. At the
+     * start of a rule they mean "merge this file", so reject them instead of
+     * silently turning them into inert exclude patterns. */
+    if (*p == ':' || *p == '.' || *p == '!') {
+      snprintf(err, err_size,
+               "filter rule starting with '%c' is not supported (merge/dir-merge/list-clear "
+               "shorthands are not implemented; use +/- include/exclude rules)",
+               *p);
+      free(text);
+      return NULL;
+    }
     const char* sp = p;
     while (*sp != '\0' && *sp != ' ' && *sp != '\t')
       sp++;
@@ -68,16 +98,17 @@ FilterRule* filter_rule_parse(const char* line, char* err, size_t err_size) {
       action = FILTER_ACTION_EXCLUDE;
       p = sp;
     }
+    while (*p == ' ' || *p == '\t')
+      p++;
   }
 
-  while (*p == ' ' || *p == '\t')
-    p++;
   if (*p == '\0') {
     snprintf(err, err_size, "filter rule has no pattern");
     free(text);
     return NULL;
   }
 
+  /* A pattern beginning with '/' is anchored (either as "-/foo" or "- /foo"). */
   bool anchored = false;
   if (*p == '/') {
     anchored = true;

src/shared/filter.h

@@ -18,8 +18,12 @@
  *      contains a .rsync-filter file for per-directory rules)
  *   a trailing '/' makes the rule match directories only
  *
- * Unsupported rsync rule types (merge/dir-merge/hide/show/protect/risk/clear,
- * rule modifiers other than '/') are rejected with a clear error.
+ * Rejected explicitly (no silent no-ops): the rsync merge/dir-merge/list-clear
+ * shorthands written as a rule that starts with ':' or '.' or '!', the
+ * merge/dir-merge/hide/show/protect/risk/clear words, and every include/exclude
+ * rule modifier other than '/' (! C s r p x). The pattern must be separated
+ * from +/- by a space (or a single '/' anchor), exactly like rsync's
+ * "-s foo"/"-p ..." modifier syntax is refused.
  */
 
 typedef enum {