docs: all-or-nothing TOCTOU caveat and two-section manifest budget

The walker's all-or-nothing guarantee holds only while the destination is not
concurrently modified (rehearsal and delete are separate walks).  The
protected-prefix list shares the 16 MB MAX_MANIFEST_BYTES budget with the
keep-set and each section is capped at MAX_MANIFEST_ENTRIES; an over-budget
frame is rejected on the receiver with STATUS_ERROR rather than truncated.

RSYNC_COMPAT.md

@@ -154,17 +154,24 @@ then reported an error (a truncated deletion); it now removes nothing and fails
 with an error naming the bound. Directories count toward the bound. A directory
 that still holds entries the walker leaves in place (a protected excluded file,
 a kept manifest entry, a symlink) is left behind rather than failing the run —
-matching rsync's "cannot delete non-empty directory" behaviour.
+matching rsync's "cannot delete non-empty directory" behaviour. The
+all-or-nothing guarantee holds only while the destination is not concurrently
+modified: rehearsal and delete are two separate walks, so a concurrent change
+between them (another process adding or removing destination entries) can make
+the actual deletion diverge from the counted set.
 
 Manifest size: the sender's keep-set and protected-prefix collections (streaming
 or early pre-scan) are unbounded, but the receiver rejects a manifest beyond
-`MAX_MANIFEST_ENTRIES` (1 048 576 entries) / `MAX_MANIFEST_BYTES` (16 MB of
+`MAX_MANIFEST_ENTRIES` (1 048 576 entries, applied to EACH section — a frame can
-paths, counted across both sections) as a hard protocol error. In the commit
+therefore total up to 2 097 152 entries) / `MAX_MANIFEST_BYTES` (16 MB of paths,
+counted across BOTH sections) as a hard protocol error. A heavily filtered
+source whose exclusion list grows large thus fails the run cleanly on the
+receiver (STATUS_ERROR) instead of being silently truncated. In the commit
 modes this only means the deletion is refused after the data already arrived; in
 the early modes (`--delete-before`/`--delete-during`) the manifest is the first
-frame, so an oversized keep-set aborts the whole transfer BEFORE any data is
+frame, so an oversized keep-set or protected list aborts the whole transfer
-sent. Keep the source tree small enough for the receiver's manifest caps when
+BEFORE any data is sent. Keep the source tree small enough for the receiver's
-using the early timing.
+manifest caps when using the early timing.
 
 Early-delete ACK wait: after committing a large deletion (up to
 `MAX_SERVER_DELETE_COUNT` removals) the receiver's `STATUS_OK`/`STATUS_ERROR`

src/shared/utils.h

@@ -37,7 +37,11 @@ typedef struct {
    max_delete is not SIZE_MAX the run is all-or-nothing: extras are counted
    first and DELETE_WALK_LIMIT_EXCEEDED is returned (with nothing removed) when
    the count would exceed the cap.  `deleted_out` optionally receives the number
-   of entries actually removed. */
+   of entries actually removed.  The all-or-nothing guarantee holds only while
+   the destination tree is not being concurrently modified: the rehearsal pass
+   and the delete pass are two separate walks, so a concurrent change between
+   them (another process adding/removing entries) can make the second pass
+   delete a different set than the first one counted. */
 DeleteWalkResult delete_extras_limited(const char* dest_root, ArrayList* manifest,
                                        size_t max_delete, const DeleteSkipEntry* skips,
                                        int skip_count, size_t* deleted_out);