fix: support Dolt 2.3.1 GTID replication

[weicao]

Review-only Dolt 2.3.1 replication update

This PR is the post-#484 rebuild of the Dolt 2.3.1 replication change. It is
not a release or merge authorization.

Locked identity

• Signed head: e20e51bc77345d898ae3dd4c6f7cfcd03986348b
• Unique parent: ba805622a2b0c855e904e44e158112b216c8724b
• Tree: 26c91f5ffefec93fccf1149fd175e91b80c94c03
• GitHub signature: verified=true, reason=valid
• Scope: exactly 6 files, +977/-259
• Report SHA-256: e7ae857904d85463951dce5a85450455b9a4152b850d265725be3b46a9039d1e
• Evidence archive SHA-256: f6ab77f8c613421628ccf7633cf97b7b4c936d954ab8be95a6babc7f060e8153

The parent is the merged PR #484 Dockerfile fix (ba805622); the Dockerfile
is not part of this PR. Rejected heads 3a0d73b5, d54fcc18, and local
unsigned bce4878 are not in the signed commit ancestry. Old approvals do not
transfer.

Implementation boundary

The applier preserves main's QueryFlush barrier, reconnect transaction discard,
encoded/current file-position loading, two-argument non-Dolt dump path, and
transaction ordering. It adds only dual Dolt identity detection and
BinlogThroughGTID (0x04) for confirmed Dolt 2.3.1 MySQL56 GTID positions.
The replication script now suffixes both container names and uses those names
for all Docker operations and DSNs, preventing concurrent collisions and
cleanup of pre-existing fixed-name containers.

Fresh evidence

• Candidate image myduckserver:task61-post484:
  sha256:cd15802b761da1f4ee6763c0bbcc7c094f7f07375cc6a2803ad91dca677405c9
• Exact-parent image myduckserver:task61-parent-ba805:
  sha256:518653912d43d5bd9aab96abce9c8606f54f319093da6d80403b6569a107c5ee
• Dolt candidate: exit 0 through initial/incremental/STOP-START/process-restart,
  exact rows 2 -> 3 -> 4 -> 5, GTID through :1-8, IO/SQL Yes, errno 0.
• Dolt exact parent: STOP/START exit 1 at the locked main baseline (expected 4,
  got 7); initial/incremental errno 0.
• MySQL candidate and parent: exit 0.
• MariaDB candidate and parent: exit 0 with the same known Last_SQL_Errno=1105
  unknown-event baseline.
• PostgreSQL candidate and parent: both stop at the same 2-row incremental
  boundary; this is shared replication behavior, not candidate-only.
• Focused capability/dump/QueryFlush/position tests, precompiled restart probes,
  compile-only tests, build, tidy, format, shell syntax, ShellCheck and
  diff-check pass. Two static rounds are recorded as
  0b5159a7b0495d1e2df1c470d55c370877247438a9a7357365b114cccbafb802 and
  8529f7135ceb56aaa7adc7dfb9e9ce12894ff344eb5b0720d2967cf2d8597ef9.
• The macOS host full-binlog command timed out after 601 seconds because its
  go run . service never listened; all fixtures failed with connection
  refused before product assertions. This is environment evidence, not a
  candidate-only regression. The precompiled candidate/parent package result
  is the shared AutoRestartReplica 1105 / row-event flag 0x10 baseline.

Do not merge, enable auto-merge, publish an image, move a release tag, or move
latest from this PR. A new independent non-author review must bind this exact
head before any integration decision.

[leon-ape]

Independent review approved for exact signed head e20e51b (tree 26c91f5ffefec93fccf1149fd175e91b80c94c03). Reviewed the six-file diff, frozen candidate/parent evidence, hosted four-source replication matrix, exact-parent CI baselines, and fresh local focused/repeated/race/build/static gates. No merge, auto-merge, release, tag, or image action taken.