[s2x]

What changed

empty("0") is true in PHP, so both emptiness checks rejected the one-character payload "0"
even though FastEncoder, FfiEncoder, NativeEncoder and the C++ core all encode it. Encoder
and QRCode now compare against '' explicitly.

QRCode still requires a valid URL, so "0" is reported there as
Invalid URL provided: "0" rather than as empty data — the old message was factually wrong.
The empty check stays first, so '' still yields Data cannot be empty and nothing else moves.
Non-URL payloads go through an encoder directly.

The other backends were swept and needed no change: FastEncoder already uses
strlen($url) === 0, FfiEncoder uses $data === '', NativeEncoder passes ZSTR_LEN to
C, and clib/src/encoder.cpp checks len == 0.

How it was tested

The "0" payloads are pinned by three new rows in tests/fixtures/qr_reference.csv. That file is
read by two independent mechanisms: tests/QrReferenceTest.php compares every row against
Encoder, FastEncoder and FfiEncoder, and clib/tests/test_csv_fixtures.cpp compares
the same rows against the C++ core — so one row pins four backends.

• Fixture generation. The 441 module bits per row were produced by a standalone C++ program
  linked against clib/src/encoder.cpp, with no PHP in the process, then compared against all
  three PHP backends: 3 payloads × 3 backends = 9 comparisons, all byte-identical. The fixture is
  not "whatever PHP now produces".
• The fixture has teeth. test_csv_fixtures accepts the new rows (1772 → 1775 data rows,
  0 failures) and rejects a copy of the "0" row with a single module bit flipped
  (1 tests, 1 failures, exit 1).
• tests/EncoderTest.php: asserts v1 / size 21 for "0", "00" and "0" repeated ten times
  (what the payload actually determines), and keeps the empty-string case throwing
  Data cannot be empty.
• tests/QRCodeTest.php: the empty-URL case, plus a test pinning "0" to
  Invalid URL provided: "0".

Red/green

Red was reproduced by reverting only src/ to the base commit 9c93e75, with the new tests and
fixtures in place:

$ git checkout 9c93e75 -- src/Encoder.php src/QRCode.php
$ vendor/bin/phpunit --filter 'data set "0' tests/QrReferenceTest.php

 ✘ Encoder matches reference with data set "0 ECL=L"
   CrazyGoat\ScanMePHP\Exception\InvalidDataException: Data cannot be empty
   src/Encoder.php:87
 ✔ Encoder matches reference with data set "00 ECL=L"
 ✔ Encoder matches reference with data set "0000000000 ECL=L"
 ✔ Fast encoder matches reference with data set "0 ECL=L"
 ✔ Ffi encoder matches reference with data set "0 ECL=L"
 ERRORS! Tests: 9, Assertions: 16, Errors: 1.

Red lands on exactly Encoder matches reference for "0" while FastEncoder and FFI pass — the
backend disagreement, expressed in the repo's own terms. Note that only the "0" row goes red;
"00" and "0"×10 were never broken, so those rows are guard rails rather than reproducers.

The other new tests on base code:

$ vendor/bin/phpunit tests/EncoderTest.php tests/QRCodeTest.php
 ✘ Encode accepts zero only payload with data set "single "0""
 ✘ Zero payload is reported as invalid url not empty data
 Tests: 32, Assertions: 75, Errors: 1, Failures: 1.

testEmptyUrlThrowsEmptyData passes both before and after the fix — '' behaviour is unchanged
by design — so it is deliberately not claimed as a red case.

Green with the fix restored:

• bin/lint.sh — All checks passed. composer validate --strict — valid.
• composer test — 5487 tests, 20153 assertions, OK. (8 deprecations + 4 PHPUnit deprecations
  are pre-existing imagedestroy() calls; the new tests add none.)
• ctest in clib/build — 3/3 passed, with the extended CSV.

Review round 2

Both blocking review findings were in the tests and are addressed:

• F1 — the parity test was a tautology. Accepted and deleted rather than repaired.
  Encoder::encode() delegates to FastEncoder::encodeVersion() for every version ≤ 27 and all
  three payloads resolve to v1, so both sides of the comparison were the same call; the test
  stayed green even with FastEncoder made to return a wrong matrix. Comparing two PHP encoders
  cannot work while that fast path exists, so the coverage moved to the fixture rows above. The
  misleading docblock sentence was deleted with the test.
• F2 — testEncodeAcceptsZeroOnlyPayload pinned nothing. Accepted. assertInstanceOf was
  guaranteed by the return type, the version check held for any constructible Matrix, and
  assertCount($matrix->getSize(), $matrix->getData()) counted rows (nested bool[][]), so it
  would fail confusingly with a 21-vs-441 diff on any getData() refactor. All three are gone.

Non-blocking points: the CHANGELOG "Both sites" wording is fixed, and the PR body's red/green
claim now names the specific cases that go red instead of implying all of them. src/Builder.php:39
keeps its empty() — it tests shell_exec() output of which/where, never a payload. The
README does not document that QRCode is URL-only; that gap is being filed separately.

Checklist

• Tests added or updated, and all checks pass
• CHANGELOG.md updated under [Unreleased]
• Documentation updated (no doc changes needed; the QRCode-is-URL-only gap is filed separately)
• Everything is written in English
• The PR title is a Conventional Commit (feat: ..., fix: ...)

Fixes #75