feat(p6-batch): residual-batch write/read driver + codec

Implements the client-only residual-batch feature end-to-end:
- src/shared/batch.{c,h}: self-contained single-file batch codec using the
  existing chunk_serialize/chunk_deserialize codec (byte-identical by
  construction).  Magic+format-version header (metadata mode is persisted into
  the header so a batch is self-describing across machines), length-prefixed
  chunk records, bounded reads that reject malformed/truncated/oversized
  records cleanly.
- src/client/client_send.c: write_batch_from_source (deterministic separate
  scan pass, loads every chunk's file images, emits header+records) and
  apply_batch_to_dest (local apply to a destination root via
  file_save_to_disk_full).  No wire change, no server involved.
- src/client/client_validation.c: --write-batch XOR --only-write-batch;
  --read-batch exclusive with both; --read-batch needs only a DEST,
  --only-write-batch only a SOURCE.
- src/client/client_cli.c: main() drives the three batch modes without
  connecting/transferring for read/only-write; --write-batch runs the live
  transfer (single-threaded so the config survives) then emits the batch.
- tests/test_batch.{c,h} (unit: byte-identical roundtrip with and without
  metadata; bad-magic/truncated/oversized rejection) + tests/integration/
  test_batch.py (only-write no-server, read-batch no-source roundtrip,
  --write-batch with a live transfer, conflict rejections).
- clang-format: realign PART-1 config.h comment block.

No PROTOCOL_VERSION bump, no config-frame field, no server flag.

tests/runner.c

@@ -1,4 +1,5 @@
 #include "test_array_list.h"
+#include "test_batch.h"
 #include "test_chunk.h"
 #include "test_change_list.h"
 #include "test_checksum.h"
@@ -50,6 +51,7 @@ int main() {
   RUN_TEST(test_array_list);
   RUN_TEST(test_shared_utils);
   RUN_TEST(test_chunk);
+  RUN_TEST(test_batch);
   RUN_TEST(test_change_list);
   RUN_TEST(test_config);
   RUN_TEST(test_credentials);