fix(cli): respect atomic END boundaries (CLI-2405)

[7ttp]

TL;DR

Prevents valid BEGIN ATOMIC functions from being truncated when an identifier contains end

Prob/sol

The splitter treated end inside names such as pending as the closing END,
causing migrations to fail with SQLSTATE 42601
now fixed by:
Recognizing END only at PostgreSQL identifier boundaries in both SQL splitters

ref:

• closes Statement splitter closes a BEGIN ATOMIC body at any identifier ending in end — the closer lacks the boundary check the opener already has #6562

[github-actions]

Superseded by a newer AI review

🤖 AI Review

Five of Claude's seven findings are confirmed; two are refuted. The most significant issue is a pre-existing but real BEGIN ATOMIC splitting bug for top-level CASE expressions. The digit-before-dollar guard also diverges from PostgreSQL tokenization, while three confirmed nits concern maintainability or documentation. Codex reported no findings.

Findings

Severity	Location	Category	Sources	Claim
🟠 MAJOR	apps/cli/src/command-internal/sql-split.ts:142	correctness	claude	A top-level CASE expression inside BEGIN ATOMIC is mistaken for the function body's closing END, causing valid SQL to be split prematurely.
🟡 MINOR	apps/cli/src/command-internal/sql-split.ts:42	correctness	claude	The new dollar-quote guard mistakes a preceding numeric literal's digit for proof that an identifier is open, unlike PostgreSQL's lexer.
⚪ NIT	apps/cli/src/command-internal/sql-split.ts:142	maintainability	claude	The pendingEnd identifier-continuation branch bypasses the nested state, coupling correctness to ReadyState's current identifier-rune behavior.
⚪ NIT	apps/cli/src/command-internal/sql-split.ts:18	documentation	claude	The non-obvious PostgreSQL-specific reason for treating every code point at or above 0x80 as an identifier character is undocumented.
⚪ NIT	apps/cli/src/command-internal/sql-split.ts:23	maintainability	claude	hasIdentifierRuneBefore leaves its two-code-unit window unexplained and handles negative offsets inconsistently with its Go counterpart.

Refuted findings (kept for transparency, not posted as review comments)

• apps/cli/src/command-internal/sql-split.unit.test.ts:68 (test-coverage): The new identifier-before-$ behavior lacks meaningful observable regression coverage outside BEGIN ATOMIC tests.
  Refuted: The existing cases exercise the guard through observable splitter output and fail without it. A standalone example would cover the same path using malformed SQL, contrary to the trusted convention against redundant assertions.
• apps/cli-go/pkg/parser/state.go:218 (maintainability): AtomicState's delimiter comparisons should be replaced by a separate boolean or enum because its delimiter has only two possible values.
  Refuted: The delimiter is not merely a discriminator: it is the value endsWithDelimiter must match. Adding a separate boolean would duplicate state and create a possible inconsistency without correcting any demonstrated behavior; the small string conversions are not a substantive defect.

Stats

Claude findings: 7 · Codex findings: 0 · Confirmed: 5 · Refuted: 2 · Uncertain: 0

---

Models: claude-opus-5 + gpt-5.6-sol · Trigger: auto · Workflow run

This review runs once per PR. A maintainer can request another with a /ai-review comment.

[7ttp]

/ai-review

[github-actions]

🤖 AI Review

Both independent reviews completed: Claude reported six findings and Codex reported none. All six concerns are confirmed, though the comment-density issue predates this PR and the overlapping-comment bug requires malformed SQL. The most substantive issue is a regression that can leave a nested AtomicState stale and swallow subsequent statements.

Findings

Severity	Location	Category	Sources	Claim
🟡 MINOR	apps/cli-go/pkg/parser/state.go:271	correctness	claude	A nested END-delimited AtomicState that returns nil on its confirming semicolon remains stored in the enclosing AtomicState, preventing the enclosing parenthesized state from closing.
🟡 MINOR	apps/cli-go/pkg/parser/state.go:66	parity	claude	isBeginAtomic uses broader, differing Unicode whitespace classes in Go and TypeScript instead of the PostgreSQL lexer whitespace class introduced by this change.
🟡 MINOR	apps/cli/src/command-internal/sql-split.ts:68	correctness	claude	isCommentsAndWhitespace and BlockState handle overlapping block-comment delimiters differently, allowing END to close a body after the FSM has treated text as statement content but the helper has consumed it as an unterminated comment.
⚪ NIT	apps/cli-go/pkg/parser/state.go:51	dead-code	claude	Parenthesis-delimited AtomicState instances initialize and update END-specific bookkeeping that cannot affect their behavior.
⚪ NIT	apps/cli/src/command-internal/sql-split.ts:220	comments	claude	Several newly added inline comment blocks exceed the trusted repository convention of one or two lines, while the file remains above the documented comment-density limit.
⚪ NIT	apps/cli/src/command-internal/sql-split.unit.test.ts:109	test-coverage	claude	The TypeScript tests cover pending END only when the confirming character is a semicolon, leaving the non-semicolon and EOF paths untested.

Stats

Claude findings: 6 · Codex findings: 0 · Confirmed: 6 · Refuted: 0 · Uncertain: 0

---

Models: claude-opus-5 + gpt-5.6-sol · Trigger: manual · Workflow run

This review runs once per PR. A maintainer can request another with a /ai-review comment.