MySQL automatically removes any index whose every column is dropped
in the same ALTER TABLE flow. Issuing an explicit `DROP INDEX i`
afterwards then errors with `Error 1091 (42000): Can't DROP 'i'; check
that column/key exists`.

The bug was first surfaced while writing the wildcard YAML fixture in
PR #21 — that fixture sidestepped the case by exercising
table+constraint+view drops together. With this fix, --allow-drop=all
on a column-with-its-only-index now applies cleanly.

diff/tables.go:
- diffTable computes a `droppedColumns` set (current columns absent
  from desired) once and passes it to diffIndexes.
- diffIndexes adds an early-skip: pure DROP INDEX whose every part
  references a dropped column is silently dropped from the plan.
  Indexes with expression parts (no Column field) don't qualify —
  we don't track expression-column dependencies here.

Tests:
- TestDiffDropColumnSuppressesDependentIndex (single-col index)
- TestDiffDropAllColumnsSuppressesCombinedIndex (multi-col, all dropped)
- TestDiffDropPartialColumnKeepsIndex (multi-col, only some dropped —
  index DROP must still fire so the column drop succeeds)
- TestDropPolicyWildcard rewritten to a clean "wildcard allows
  everything" assertion using a separate table; the column-dependent
  case is its own test now.

YAML fixtures (3 added):
- apply/drop_column_with_dependent_index — single-column case
- apply/drop_two_columns_with_combined_index — multi-column suppression
- apply/drop_one_of_two_index_columns — partial coverage; verifies the
  diff still emits DROP INDEX + new CREATE INDEX

Verified manually:
  apply --allow-drop=all on `users(id, legacy, KEY i (legacy))` →
  desired `users(id)` now succeeds (previously errored 1091).

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

Three fixes — one bug, two doc / fixture corrections.

🔴 Bug: the new dependent-index suppression ran regardless of drop
policy. If a user passed --allow-drop=index but NOT
--allow-drop=column, the column drop would be suppressed (sent to
DisallowedDropStmts) yet my code still treated the column as
"going away" and silently dropped the explicit DROP INDEX. Net
effect: an orphaned index that nobody was managing.

Fix: gate the dropped-columns set on dc.IsDropAllowed("column"). If
the user hasn't allowed column drops, dropped is nil and
allPartsDropped never matches, so DROP INDEX is emitted as before.

New regression test:
TestDiffDropIndexNotSuppressedWhenColumnDropDisallowed.

🟢 Docs:
- diff/tables_test.go comment said `diff.allPartsDropped` —
  the helper is in this package so the qualifier was misleading.
- testdata/apply/drop_one_of_two_index_columns.yml had a stale
  comment claiming "this case would still fail at apply time"
  with `verify_no_drift: false`, but the apply actually succeeds
  (MySQL folds the multi-col index down to its surviving column
  on DROP COLUMN; we then DROP INDEX the leftover and CREATE
  the desired narrower form). Reworded the comment to match
  reality and dropped `verify_no_drift: false`; the harness now
  re-plans after apply and confirms no drift.

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

Copilot review caught a remaining 1091 case my earlier fix didn't
cover: the DROP+CREATE replacement path. When an index keeps the
same name but moves to different columns AND the original columns
are being dropped (e.g. KEY i (a) → KEY i (b) with column a gone),
the diff emits DROP INDEX i + CREATE INDEX i (b). MySQL auto-removes
the old index during DROP COLUMN a, so the explicit DROP INDEX still
errors 1091.

Drop the `!ok &&` qualifier from the suppression — the check now
applies whether the index is being purely removed or replaced. The
desired-side loop still emits the new CREATE INDEX, so the
replacement converges.

Tests:
- TestDiffReplaceIndexSuppressesDependentDrop (Go) covers
  KEY i (a) → KEY i (b) with column a dropped.
- testdata/apply/replace_index_with_column_drop.yml runs the same
  scenario end-to-end, and the harness's drift check verifies
  apply converges to the desired schema.

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

Make droppedColumns the single source of truth for the dropped-column
set. It now returns []string (current's iteration order); diffColumns
iterates the slice instead of duplicating the loop, and diffTable
materialises the same slice into a map[string]bool for allPartsDropped
lookups in diffIndexes. Behaviour unchanged.

Addresses Copilot PR #23 review #3 inline comment.

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