feat(format): expand the runtime safety gate
CI / Test (push) Successful in 37s
CI / Build (push) Successful in 52s

Replace the bare SemanticallyEqual call in every frontend (CLI fmt, LSP
formatting + rangeFormatting) with format.VerifySafe, which runs four
checks before any formatted output is emitted:

  - semantic equivalence  - the code token stream is unchanged
  - comment preservation  - no -- or /* */ comment is dropped, merged,
    split, reordered, or reworded (line endings / indentation normalised
    away; recurses into dollar-quoted bodies)
  - structural balance    - the () [] and BEGIN/CASE/IF/LOOP...END nesting
    profile matches, ignoring anything inside a comment or a string
  - idempotence           - a second format pass would not change it

On failure the CLI now prints the specific reason and keeps the original.

The comment check surfaced two real formatter bugs, both fixed in
formatBodyStatements:

  - multi-line /* */ comments inside a PL/pgSQL body had their interior
    lines re-split and reindented as if they were statements; they are
    now tracked and carried verbatim with the opening line
  - a column-0 -- line was glued onto the preceding line by the
    split-line-join, which merged consecutive comment lines into one

Regenerate testdata/corpus/test_mm_proc.pgsql (was carrying the mangled
output). TestCorpusIdempotentAndSafe now runs the full VerifySafe bundle;
add safety_test.go with targeted cases.
This commit is contained in:
Hein
2026-09-10 15:17:07 +02:00
parent 83b215fd25
commit 5cdec88299
9 changed files with 363 additions and 31 deletions
+16 -7
View File
@@ -60,13 +60,22 @@ testdata/corpus/ — real-world .pgsql procedures used as the safety/idempoten
4. **Graceful degradation**: any span the parser cannot handle is passed through verbatim
rather than corrupted.
5. **Runtime safety gate**: every frontend (CLI `fmt`, LSP `textDocument/formatting` and
`rangeFormatting`) calls `format.SemanticallyEqual(src, out)` before writing or returning
formatted output. It re-lexes both sides and compares non-trivia token streams
(case-insensitive for identifiers/keywords, exact otherwise, recursing into dollar-quoted
bodies). If it ever returns false, the formatter has a bug — the caller must refuse to
write/emit the result and keep the original source, never guess or best-effort it. This is
not just a test assertion (`pkg/format/format_test.go`); it is enforced at runtime so a
formatter bug can never silently drop or alter code.
`rangeFormatting`) calls `format.VerifySafe(src, out, style)` before writing or returning
formatted output. If it returns a non-nil error the formatter has a bug — the caller must
refuse to write/emit the result and keep the original source, never guess or best-effort
it. This is enforced at runtime, not just in `pkg/format/*_test.go`, so a formatter bug can
never silently drop or alter code. `VerifySafe` runs four checks:
- **Semantic equivalence** (`SemanticallyEqual`): re-lex both sides, compare the non-trivia
token streams — case-insensitive for identifiers/keywords, exact otherwise, recursing
into dollar-quoted bodies. Comments and whitespace are trivia and are ignored here.
- **Comment preservation** (`CommentsPreserved`): every `--` and `/* */` comment in `src`
reappears in `out`, in order, with the same content (line endings and indentation are
normalised away; dropping, merging, splitting, reordering, or rewording a comment is
not). Descends into dollar-quoted bodies.
- **Structural balance** (`StructurallyBalanced`): the `( ) [ ]` and BEGIN/CASE/IF/LOOP…END
nesting profile of `out` matches `src`, counting only real code tokens (comment and
string/dollar-quote contents are skipped).
- **Idempotence**: re-formatting `out` yields `out` unchanged.
**Line endings**: the formatter re-emits all layout with `st.Newline` (default `\n`), so a
CRLF input file is normalised to LF on write. This includes `\r\n` that sits *inside* a