M+Minify++

Production assessment

Is Minify++ production-grade?

The previous scoped pass was withdrawn after a real CSS failure. Following repairs and materially stronger evidence, the decision is now PASS WITH KNOWN LIMITATIONS — REVALIDATED AFTER REOPENING.

Production-grade for the documented scope.The restored pass follows a clean committed-package test and fresh-browser website validation. It is not a claim of universal syntax, validating-parser behavior or untested platform support.

Why the answer is yes

  • All current direct correctness, semantic, cross-format, CLI and Nift integration suites pass.
  • 15,459 generated JavaScript programs execute before and after minification with equivalent observed behavior.
  • 180 JSX/TSX programs pass reference parsing and idempotence checks.
  • 115 non-JavaScript documents exercise HTML, CSS, JSON, XML and SVG contracts.
  • Representative modern stylesheets pass an independent normalized PostCSS semantic-tree comparison.
  • 7,000,000 deterministic mutations pass across all seven format modes under both ordinary and ASan/UBSan builds.
  • The standalone and Nift-embedded implementations pass an exact 20-file synchronization gate.
  • Performance and peak-memory measurements provide a repeatable regression reference.
  • make distcheck builds a clean committed git archive and runs the complete standalone suite.
  • Public limitations and the reopened decision are documented rather than obscured.

The fuzz campaign materially changed the assessment

Fuzzing did not merely run and report green. It first found three genuine token-manufacturing families:

CSS       / * or * /       → accidental comment delimiter
JSX       < name            → JSX-significant opener
HTML/XML  separated prefix → <!-- or <![CDATA[

The original campaign found three token-manufacturing families, but a later real website build still exposed unsafe ordinary CSS boundaries: dimensions, descendant selectors, quoted component values, transform lists and modern color-function arguments could be merged. The reopened audit then found JavaScript block-comment delimiters manufactured across whitespace, JSX roots detected inside JavaScript comments, and unterminated quoted constructs accepted by several format scanners.

Each finding paused the broader campaign. The reproducer was minimized, neighboring cases were tested, the boundary logic was fixed, and focused regressions were retained. The ordinary-build campaign then reached 7,000,000 successful mutations, and CSS gained an independent semantic-tree oracle so idempotently wrong output cannot satisfy the evidence wall by itself.

A permanent self-consistency invariant

minify(input) succeeds
        ↓
output = transformed artifact
        ↓
minify(output) must not reject output

This does not claim that every format always reaches a byte-identical fixed point after one pass. Strict idempotence is a stronger, separate contract. Second-pass acceptance protects the more fundamental promise that Minify++ does not emit text outside its own supported language.

Known limitations

  • Platforms: this assessment directly validates Linux x86-64. macOS and Windows remain explicit evidence gaps.
  • XML/SVG: these are conservative lexical modes, not complete validating XML parsers.
  • Embedded languages: HTML raw script/style bodies and JavaScript template contents are preserved rather than recursively minified.
  • TypeScript: there is no general TypeScript input mode. JSX has tested TSX-aware boundaries but Minify++ does not compile TS.
  • Finite evidence: large differential and fuzz corpora increase confidence; they do not prove all future syntax.
  • Performance: current figures are host/fixture regression references, not universal promises.
  • Filesystem durability: staged replacement is recoverable, but there is no claim of power-loss durability or ownership/timestamp preservation.

What would reopen the decision?

A supported input changing meaning, a crash/hang or unbounded-growth case, a scanner change without equivalent adversarial evidence, drift between standalone and embedded copies, a regression in file replacement, a failed clean-package build, or public claims expanding beyond tested platforms would all require reassessment.

The candid conclusion

This production-grade decision rests on a bounded job, controlled failure behavior, meaningful adversarial testing, family-level fixes and claims that match the evidence. Its withdrawal and later restoration demonstrate that status is maintained and revocable—not permanent by ceremony.

Build Minify++ Inspect the test evidence