Opt-in flag (default off; env MYSCHEMA_BULK_ALTER) that folds
*consecutive* same-table single-spec ALTER TABLE statements in the
table-change bucket into one multi-spec ALTER. The combiner is
order-preserving and runs only on tableDiff.Stmts — not on the
RenameStmts / FKDropStmts / FKAddStmts buckets — so the diff
pipeline's cross-table FK ordering invariants stay intact.

Excluded (each acts as a run separator):
  - FK adds / drops live in dedicated FK buckets so cross-table
    ordering survives. CAVEATS.md "Bulk-alter does not combine FK
    operations" explains the rationale.
  - Partition ops (REORGANIZE / ADD / DROP / COALESCE / TRUNCATE /
    EXCHANGE PARTITION). MySQL's grammar rejects the trailing-comma
    multi-spec form for these clauses; detected via the existing
    partitionOpInsertPos helper.
  - Standalone CREATE INDEX (different statement shape).
  - CREATE / DROP / RENAME TABLE.

Tests: 12 unit cases in bulk_alter_test.go (combineSameTableAlters
and splitCombinableAlter both 100%) covering the headline behaviour,
edge cases (nil/single/no-semicolon/back-ticked), and run-separator
shapes (different-table, non-contiguous, partition, CREATE INDEX,
malformed, DROP TABLE). Two YAML fixtures pin the end-to-end
behaviour: same-table column folds combine, FK ops + standalone
CREATE INDEX stay separate.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Apply the three review items that pay for themselves; skip the
help-text expansion (per prior user feedback that --bulk-alter help
is already long enough — the same applies to the rename exclusions).

- CAVEATS.md gains a "--bulk-alter interacts with --alter-algorithm /
  --alter-lock" subsection. Combining specs into one ALTER means
  MySQL applies a single ALGORITHM= / LOCK= choice to the whole
  statement, not per-spec. A run mixing INSTANT-eligible specs with
  one that needs INPLACE / COPY (DEFAULT CHARSET, generated columns,
  some MODIFY COLUMNs) becomes a COPY-or-fail under
  --alter-algorithm=INSTANT — the documentation now warns
  operators to plan first and verify each spec before opting in.
- testdata/apply/bulk_alter_convert_charset_with_column_add.yml:
  end-to-end fixture for `-- myschema:convert-charset` folded with
  a same-table ADD COLUMN under bulk-alter. CONVERT TO CHARACTER
  SET has different syntax / rebuild semantics than DEFAULT CHARSET
  (data rewrite vs. table-default flip), so the existing
  bulk_alter_charset_with_column_change.yml didn't cover it.
- options_test.go: TestAlterOptionKongBulkAlter pins the kong tag
  wiring for --bulk-alter and MYSCHEMA_BULK_ALTER. The YAML
  fixtures set BulkAlter directly via Go, so a typo in the kong
  field-name → flag-name mapping or env tag content would
  otherwise slip past CI. Mirrors the existing pre-sql / allow-drop
  kong-parse pins.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- options.go: --bulk-alter help string now lists `RENAME COLUMN /
  RENAME INDEX` alongside FK / partition / standalone CREATE INDEX.
  CLI users hitting --help get the same exclusion picture as
  CAVEATS.md / AGENTS.md.
- AGENTS.md: drop the false "and the partition shapes that aren't
  on the excluded list below" tail from the Combinable bullet.
  The excluded list covers every partition op the diff currently
  emits (ADD / DROP / COALESCE / REORGANIZE / TRUNCATE / EXCHANGE
  PARTITION), so no partition shape is combinable today. Mention
  CONVERT TO CHARACTER SET in the combinable list since
  bulk_alter_convert_charset_with_column_add.yml pins it.
- bulk_alter.go: rewrite the partitionOpInsertPos comment.
  Previous "a partition op anywhere after the table name
  disqualifies the whole stmt" overstated the helper —
  partitionOpInsertPos only matches a partition keyword
  *immediately* after the table name (the same anchor
  appendAlterHints uses for splicing). The new comment names
  that anchor and the why (one alter_specification per stmt
  in the diff layer's output puts the partition op exactly
  there).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address Copilot review on PR #77 raising two potential evaluation-
order traps. Probe fixtures confirm MySQL 8.0+ accepts both combined
forms; promote both to durable regression pins so a future MySQL
behaviour change (or a regression in the combiner) surfaces in CI
rather than in production:

- bulk_alter_first_after_chain.yml: `ADD COLUMN head_a INT FIRST,
  ADD COLUMN head_b INT AFTER head_a;` — the AFTER clause targets
  the column the preceding spec just inserted at FIRST. MySQL
  resolves it against the post-FIRST column list.
- bulk_alter_add_column_with_check.yml: `ADD COLUMN price INT,
  ADD CONSTRAINT chk CHECK (price >= 0);` — the constraint
  references the just-added column. Same evaluation-order class as
  the RENAME COLUMN + MODIFY trap (CAVEATS.md), but on this path
  MySQL 8.0+ resolves the CHECK against the post-spec state, so
  the combined form applies cleanly.

Both round-trip empty (verify_no_drift: true).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address Copilot review on PR #77. The previous "Drop one of the
flags if the combined ALGORITHM/LOCK choice doesn't fit" was
misleading because `appendAlterHints` always appends the
user-supplied ALGORITHM= clause to every generated ALTER TABLE,
regardless of `--bulk-alter`. So dropping `--bulk-alter` alone
splits the run into per-spec ALTERs but each one still carries
`ALGORITHM=INSTANT` — a non-INSTANT-eligible spec still fails on
its own statement.

Rewrite the section to call this out and split the recovery
guidance:

- Loosening the algorithm pin (or dropping `--alter-algorithm`
  entirely) is the only fix that changes what algorithm MySQL
  applies.
- Dropping `--bulk-alter` recovers the *failure shape* (each spec
  succeeds or fails on its own ALTER instead of one combined
  abort), but doesn't help any spec succeed on its own.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address Copilot review on PR #77:

- CAVEATS.md: the previous bullet asserted "two ADD COLUMN adds
  ... is fine — both qualify" for `--alter-algorithm=INSTANT`. INSTANT
  eligibility depends on exact 8.0.x version, column position, and
  presence of generated columns ahead of the new one — myschema's
  baseline is "8.0+" without pinning a minor, so the absolute claim
  was misleading. Reword to "when every spec individually qualifies
  for the pinned algorithm, the combined ALTER inherits that
  qualification" with an explicit per-spec verification disclaimer.
- testdata/apply/bulk_alter_table_comment_with_column_add.yml:
  end-to-end fixture for `COMMENT=` (table_option) folded with a
  same-table ADD COLUMN. Different grammar shape from per-column
  alter_specifications; combinable list documents it but no apply
  pin existed. The diff orders the COMMENT= spec before column adds
  (matching diff_all.go's bucket assembly), so the asserted shape
  is `ALTER TABLE t COMMENT='new', ADD COLUMN ...`.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address Copilot review:

- options.go: AlterOption struct doc previously read "carries the
  MySQL online-DDL hints that get appended to every generated
  ALTER TABLE / CREATE INDEX statement" — accurate when the type
  only held AlterAlgorithm / AlterLock, but now the same struct
  also carries BulkAlter. Restructure the doc as a per-field
  bullet list so Go docs and editor hovers describe both
  responsibilities accurately.
- testdata/plan/bulk_alter_combine_with_index_add_separate.yml:
  drop the hard-coded `diff/tables.go:728` line reference and
  point at the CAVEATS.md "Secondary-index ADDs" subsection
  instead, so the rationale survives unrelated edits to
  diff/tables.go.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address Copilot review on PR #77 with two more end-to-end pins:

- bulk_alter_charset_with_modify_column.yml: counterpart to
  bulk_alter_charset_with_column_change.yml. The CHARSET +
  ADD COLUMN shape was covered, but DEFAULT CHARSET +
  MODIFY COLUMN is the more fragile path because string columns
  inherit the table charset only when their per-column metadata
  isn't already pinned. CAVEATS.md flags the two-apply convergence
  rule there; bulk-alter doesn't change that, so verify_no_drift:
  false matches change_table_charset.yml's baseline. The fixture
  pins that the combined first-apply ALTER applies syntactically
  and that the bucket order (tableCharsetCollationSQL before
  diffColumns) survives a refactor.
- bulk_alter_drop_column_with_drop_index.yml: bulk-alter
  counterpart to drop_one_of_two_index_columns.yml. When only
  some of an index's columns are dropped, MySQL keeps the index
  (folded down to surviving columns), so the diff still emits
  DROP INDEX + CREATE INDEX rebuild. Under bulk-alter the
  DROP COLUMN + DROP INDEX fold into one ALTER while the CREATE
  INDEX stays standalone (separate-bucket rule). Pins that
  `ALTER TABLE t DROP COLUMN a, DROP INDEX i_combined;` is
  accepted when the index's other column survives, and that the
  rebuild then lands on the post-ALTER state. (The
  replace_index_with_column_drop.yml shape — drop *all* index
  columns — uses dependent-index suppression and doesn't emit
  DROP INDEX at all, so bulk-alter has nothing to fold there.)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>