The Cleanup TODO entry asked us to make `applyAlterTable` reject the
ALTER clauses it doesn't model instead of silently skipping them.
Prototyped that as a `--strict` opt-in flag (covering both
applyAlterTable AND the symmetric top-level skip path), then decided
the whole feature is out of scope for v1 — the silently-skipped
behaviour is intentionally permissive so raw mysqldump output keeps
parsing without manual editing, and the strict-mode plumbing wasn't
worth shipping for the use cases we have today.

Drop the TODO entry and document the actual behaviour as a deliberate
design choice instead of a future fix:

- TODO.md: removed the Cleanup section's only entry. Section
  itself goes too — `Cleanup` was a one-line bin and is now empty.
- AGENTS.md: the existing `applyAlterTable` paragraph now spells
  out that every other ALTER clause is skipped, and that the
  symmetric top-level skip exists for the same reason
  (mysqldump-output compatibility). Mentions the `--strict`
  prototype + v1 decision and points to CAVEATS.md.
- CAVEATS.md: new "Unmodelled SQL in desired-side files is silently
  skipped" section with the full inventory (which top-level
  statements skip, which ALTER clauses skip), the rationale
  (mysqldump compatibility), the impact (footgun for users who
  expected `ALTER TABLE t ADD COLUMN c INT` to land), and the
  workaround (express schema state via CREATE TABLE; out-of-band
  for DML).

No code changes.

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

Two valid nits:

#1 The CAVEATS.md "top-level statements skip" bullet listed the
three statement types that ARE handled (CREATE TABLE / CREATE
VIEW / ALTER TABLE) but didn't call out that user-written
`CREATE INDEX` rides on the ALTER TABLE AST shape and IS
supported. Reword so a reader doesn't conclude `CREATE INDEX`
is one of the skipped types.

#2 The "Impact" paragraph claimed "you'll only notice when the
next plan re-emits the underlying CREATE TABLE without the
column". That's wrong — `plan` emits diff DDL, not the
desired-side CREATE TABLE. The actual symptom is "no observable
plan output for that column": the diff is silent because the
column never made it into the desired model, so the next `plan`
either reports `-- No changes` or just shows the other diffs
in the file, and `dump` over the live DB shows the table without
the column. Rewrite the paragraph with the accurate observable
symptoms.

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

…nal intro exception

Reviewer noted the new section reads as contradicting the file intro's
"be explicit, fail loudly" stance. Add a one-sentence call-out at the
top of the "Why this isn't an error" paragraph so the contrast is
explicit instead of implicit.

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

…re validated

Reviewer noted the "comments — all parse successfully" wording reads
as if every `--`-prefixed line is silently skipped, but
`-- myschema:<name>` directive comments go through ValidateDirectives
and unknown/malformed shapes error out. Tighten "comments" to
"non-directive comments" in both CAVEATS.md and AGENTS.md and add a
parenthetical noting the directive exception.

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

… + IS typo

Three nits:

- Both CAVEATS.md and AGENTS.md described unhandled `ALTER TABLE`
  clauses as "silently skipped" without distinguishing the two
  paths: known target table → unhandled clause is silent; unknown
  target table → applyAlterTable fails fast with `ALTER TABLE on
  unknown table …`. Spell out both.
- "CREATE INDEX IS supported" → "is supported" (capitalised mid-
  sentence reads as a typo).

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