Two inline comments, both valid.

parser/directive.go
  - classifyInlineLine: extend the unnamed-index guard to the two-word
    keyword forms (UNIQUE KEY, UNIQUE INDEX, FULLTEXT KEY/INDEX,
    SPATIAL KEY/INDEX). Without this, `UNIQUE KEY (col)` tokenizes
    with tokens[2]="(col)" and the directive mis-attached to a name
    of "(col)", later surfacing as "target index '(col)' not found".
    Mirrors the guard already in the plain KEY/INDEX branch.
    (Copilot #1.)
  - tokenize doc fixed: previous comment claimed it peeled off
    `","`, `"("`, `"("`, but the implementation only trims `,(`.
    Comment now matches reality. (Copilot #2.)

Tests
  - parser/directive_test.go: TestExtractInlineRenamesUnnamedTwoWordIndexIsUnsupported
    covers the new guard for `UNIQUE KEY (col)` and
    `FULLTEXT INDEX (col)`.

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

Round 1 review on PR #75 caught eight issues; this commit addresses
all of them.

1. (review #1) loadPreSQL treated any non-empty PreSQL string as
   "set", including whitespace-only values. Whitespace-only env
   var (`MYSCHEMA_PRE_SQL=`) would trigger the mutually-exclusive
   error against a legitimate --pre-sql-file, and skip the no-op
   short-circuit. Fix: TrimSpace both fields before deciding "set".

2. (review #2) loadPreSQL read --pre-sql-file with os.ReadFile,
   which doesn't support the repo's existing `-` for stdin
   convention used by parser.ReadSQLFile (the desired-SQL file
   args already accept it). Fix: route through parser.ReadSQLFile
   so `--pre-sql-file=-` works.

3. (review #3 + #4) runPreSQL was invoked AFTER connect(), so
   flag-validation errors (both flags set, missing file) could be
   masked by a downstream connection failure, and the client
   opened a DB connection it then threw away. Split into:
   - loadPreSQL: validate / read, no DB contact (called BEFORE connect)
   - execPreSQL: run on conn (called AFTER connect)
   apply.go and plan.go updated accordingly.

4. CLI-level mutual exclusion: tag both PreSQL and PreSQLFile with
   kong's `xor:"pre-sql"` so kong rejects the both-set case at
   parse time before our code sees it. The runtime check in
   loadPreSQL stays in place for programmatic API callers
   (Apply / Plan invoked from Go without going through kong).

5. (review #5 / #6 / #7 / #8) Test reshuffle. The original
   suite over-relied on success-only smoke tests that would have
   silently passed even if pre-SQL were skipped:
   - TestApply_PreSQLString and TestApply_PreSQLMultiStatement
     only checked NoError.
   - TestApply_PreSQLAppliesToSession claimed to be the
     "strongest behavioural pin" but probed nothing — the
     connection it would have probed is closed by Apply's defer.
   - TestPlan_PreSQLString same issue.
   Fix: load-bearing pins now feed INVALID pre-SQL and assert
   the apply / plan aborts with a wrapped "pre-sql" error
   containing the exact failing piece. The multi-statement
   split test now uses a payload with an invalid SECOND piece
   and asserts the error references that piece exactly (proves
   the split happened — without splitting, the driver would
   reject the whole concatenated string with a different error).

6. YAML migration: the new harness field `pre_sql` lets
   inline-payload tests live as testdata/apply/pre_sql_*.yml and
   testdata/plan/pre_sql_*.yml fixtures (3 + 1 = 4 new fixtures).
   File-related and programmatic-API tests stay in Go because
   the YAML harness has no `pre_sql_file` field (would need
   fixture-relative path handling, out of scope).

Coverage is unchanged at the package level; loadPreSQL stays at
100% and execPreSQL is exercised by all the real-execution paths.

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