Closes the Medium-severity TODO item from the PR #81 silent-drop
audit. Pre-fix, vitess parsed `col INT INVISIBLE` into
`cd.Type.Options.Invisible` (`*bool`) but `parseColumnDef` ignored
the field, so a `hidden_col INT INVISIBLE` desired-side column
silently lost the attribute on parse. The catalog reader (which
surfaces it via `information_schema.COLUMNS.EXTRA`) saw a permanent
diff against the parser side's empty state — drift never closed.

Wire the attribute through the four layers that need to know:

- model/column.go: new `Invisible bool` field on Column.
- parser/parser.go: parseColumnDef reads `opts.Invisible` (only
  the explicit-true case promotes; nil / false stays the visible
  default and emits nothing).
- model/table.go: columnDefSQL emits ` INVISIBLE` between
  AUTO_INCREMENT and COMMENT to match SHOW CREATE TABLE order.
- catalog/tables.go: loadColumns scans EXTRA for the literal
  "INVISIBLE" token (Contains, so it composes with other extras
  like AUTO_INCREMENT).
- diff/tables.go: columnEqual gains the Invisible mismatch check
  so a VISIBLE↔INVISIBLE toggle fires MODIFY COLUMN.

Tests:

- TestParseColumnInvisible (parser) pins the parsed-model shape.
- TestColumnDefSQLInvisible (model) pins the emitted token and the
  visible-default no-emit path.
- TestColumnEqual/invisible_differs (diff) pins the mismatch check.
- TestColumnInvisibleRoundTrip (catalog) pins the EXTRA → Invisible
  read for per-package coverage; the YAML round-trip exercises
  the same path transitively but doesn't register against catalog
  package coverage.
- testdata/apply/invisible_column.yml integration round-trip:
  visible → INVISIBLE column add, verify_no_drift: true. Toggling
  back (INVISIBLE → visible) is exercised manually and confirmed
  to emit `MODIFY COLUMN x int` (no INVISIBLE keyword).

Coverage on touched functions: columnDefSQL / columnEqual 100%,
loadColumns 91.9% → 94.6%, parseColumnDef 96.8% (all new
INVISIBLE branches covered).

TODO.md: drop the entry now that the fix is in.

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

Three real issues from the review:

- catalog/tables.go: ON UPDATE EXTRA parsing leaked the trailing
  INVISIBLE token into Column.OnUpdate (`DEFAULT_GENERATED on
  update CURRENT_TIMESTAMP INVISIBLE` → OnUpdate
  "CURRENT_TIMESTAMP INVISIBLE"), so dump emitted the keyword
  twice (`ON UPDATE CURRENT_TIMESTAMP INVISIBLE INVISIBLE`) and
  parser-side desired never compared equal. Fix: detect INVISIBLE
  first, strip the token from extraUp before parsing ON UPDATE /
  AUTO_INCREMENT / generated.
- catalog/tables_test.go: TestColumnInvisibleComposesWithOnUpdate
  pins the ON UPDATE + INVISIBLE composition (would have failed
  the regression above pre-fix).
- testdata/apply/invisible_to_visible.yml: end-to-end pin for the
  reverse toggle (INVISIBLE → VISIBLE), which only the YAML
  apply harness can exercise. Asserts the diff emits a bare
  `MODIFY COLUMN x int` (no INVISIBLE keyword, no explicit
  VISIBLE) and verify_no_drift confirms convergence.
- model/table_test.go: TestColumnDefSQLInvisible/`INVISIBLE sits
  between AUTO_INCREMENT and COMMENT` adds an index-comparison
  subtest so a future reorder of columnDefSQL fails CI rather
  than silently shipping wrong-order DDL — the previous Contains-
  only check was order-blind.

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

- model/table.go: ColumnDefSQL doc comment listed the emitted
  attribute order ending at AUTO_INCREMENT / COMMENT and didn't
  mention INVISIBLE — stale after this PR's emitter change.
  Update the attribute list and call out that the order is pinned
  by index-comparison assertions in model/table_test.go.
- parser/parser_test.go: TestParseColumnInvisible now exercises
  all three vitess Invisible *bool states explicitly:
  - INVISIBLE keyword → *bool=true → Column.Invisible=true
  - explicit VISIBLE keyword → *bool=false → Column.Invisible=false
  - no keyword → *bool=nil → Column.Invisible=false
  Previously only the nil and true branches were covered; the
  false branch (the `*opts.Invisible` deref's other arm) ran
  unverified. Splitting into named sub-tests pins each state
  individually so a future regression in the nil-vs-false
  handling fails a specific case.

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

- README.md / AGENTS.md: feature lists called out INVISIBLE only
  for secondary indexes. With column-level INVISIBLE wired in this
  PR, both surfaces should mention it. README's "Tables, columns
  …" bullet now includes "(incl. INVISIBLE on MySQL 8.0+)";
  AGENTS.md's "In scope (v1)" CREATE TABLE entry adds "column-level
  INVISIBLE on MySQL 8.0+" to its attribute list.
- parser/parser_test.go: TestParseColumnInvisible's doc comment
  said it pins "parser → model → emitter → catalog → diff"
  consistency, but the test only inspects the parsed model. Reword
  to describe the actual coverage boundary (parser-side vitess
  *bool decoding) and link to the dedicated tests for the other
  layers (TestColumnDefSQLInvisible, TestColumnInvisibleRoundTrip,
  TestColumnEqual/invisible_differs) so a reader navigating the
  feature can find each layer's pin without confusion.

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