v1 partitioning support: dump → plan reports no diff for a
partitioned table, plan errors out (rather than emit ALTERs)
when the partition layout differs between desired and catalog.
ADD/DROP/TRUNCATE/REORGANIZE PARTITION generation is a future
PR; users manage partition changes by hand for now and the next
plan reconverges.

- model.Table.Partition *string carries the canonical, vitess-
  formatted PARTITION BY clause. nil = not partitioned. Both
  sides go through parser.NormalizePartitionOption so byte-equal
  comparison is meaningful.
- parser/partition.go normalises a *sqlparser.PartitionOption:
  trims the leading newline, lower-cases function-name and
  column-reference identifiers (vitess preserves the user's
  YEAR / dt casing but MySQL's SHOW CREATE TABLE always emits
  year / dt in lower case), and strips the per-partition
  `engine <name>` option that 8.0+ adds even when the
  table-level engine alone would suffice.
- parser/partition.go also exposes ExtractPartitionFromShowCreate,
  which lifts the `/*!50100 PARTITION BY … */` block out of a
  SHOW CREATE TABLE result, wraps it onto a trivial CREATE
  TABLE skeleton, re-parses with vitess, and runs the same
  normaliser the desired-side path uses. That symmetry is the
  whole reason the round-trip is bytewise.
- catalog/tables.go gets loadPartitions: list partitioned
  tables via information_schema.PARTITIONS (DISTINCT, one row
  per table), then SHOW CREATE TABLE each and pull the clause.
  Two-phase keeps cost proportional to partitioned-table count
  rather than total-table count.
- diff/tables.go's diffTable errors when current.Partition !=
  desired.Partition with a clear "manage by hand" message.
  Emitted before the column / index passes so the user gets
  the partition error immediately, not after a long plan.
- model.Table.SQL appends Partition on a fresh line so dump
  output reads cleanly.

SUBPARTITION BY is intentionally out of scope for v1: a
desired-side SUBPARTITION fails at parser parseCreateTable, and
a catalog-side SUBPARTITION is rejected by the
ExtractPartitionFromShowCreate guard so myschema won't try to
manage it. The user must drop SUBPARTITION manually before
bringing the table under management.

Tests:
- parser unit (partition_test.go): RANGE / LIST / HASH / KEY /
  RANGE COLUMNS round-trip through both pipelines and produce
  byte-equal output, absent partition leaves model nil,
  SHOW CREATE TABLE without partition block returns "", and
  both directions reject SUBPARTITION.
- end-to-end YAML: dump / plan / apply fixtures pin the no-diff
  round-trip; a separate plan fixture pins the "partition
  difference errors" path with a clear message.

Docs:
- CAVEATS.md "Partitioning is round-trip only in v1": full
  what-works / what-errors writeup + SUBPARTITION exclusion +
  rationale (RANGE/LIST vs HASH/KEY differ in the ALTER
  syntax, so diff-gen is its own PR).
- AGENTS.md "In scope": catalog reader now mentions partition
  clauses; "Not yet implemented" entry for partitions rewritten
  to point at CAVEATS.
- TODO.md: entry rewritten to scope the *generation* work,
  noting SUBPARTITION is not coming.

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

Four nits — one is a real bug, three are doc / shadowing fixes:

- ExtractPartitionFromShowCreate sliced from the *first* `/*!`
  to the *last* `*/`. SHOW CREATE TABLE on MySQL 8.0+ can emit
  earlier versioned-comment blocks (e.g. `/*!80016
  ENCRYPTION='N' */`) before the partition one, in which case
  the body would start with ENCRYPTION='N' and the function
  would return ("", nil) — silently dropping the partition
  metadata. Switch to LastIndex so we always grab the trailing
  block (the partition clause is always last when present), and
  pin the fix with a regression test that mixes ENCRYPTION + a
  RANGE partition.
- Inline comment in NormalizePartitionOption mentioned a
  fictional CreateSQL method. Tables render via
  `(*model.Table).SQL`; updated.
- model.Table.Partition's field comment claimed "both sides
  normalise via sqlparser.String(*sqlparser.PartitionOption)",
  but the implementation goes through
  parser.NormalizePartitionOption (which adds identifier
  lower-casing, leading-newline trim, and per-partition engine
  strip on top of vitess's formatter). Spell out what
  "canonical" actually means.
- Loop-local clause copy in catalog/tables.go::loadPartitions
  was named `c`, shadowing the `*Catalog` receiver. Rename to
  `clauseCopy`.

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

Doc-only nit. ExtractPartitionFromShowCreate's comment said it
extracts "including any SUBPARTITION BY …", but the function
deliberately errors when a SUBPARTITION clause is present
(SUBPARTITION is out of scope for v1 — CAVEATS.md "Partitioning
is round-trip only"). Fix the comment so future readers don't
expect SUBPARTITION pass-through.

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

Add a TestParseSQLPartitionRoundTrip case for `LIST COLUMNS(...)`
with `VALUES IN ('A', 'B')`. Pins two things at once:

- LIST COLUMNS(...) is on the supported list but wasn't actually
  exercised by the round-trip test.
- The normaliser only lower-cases function-name and column-
  reference identifiers — string literals (`'A'`, `'B'`) must
  round-trip case-preserved, because flipping `'A'` to `'a'`
  would change the value MySQL stores. Without an explicit test
  a future change to the rewriter could quietly break that
  guarantee.

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