test(api): pin the enum ordinals

[andiwand]

🤖 Generated with Claude Code

PR 10 of the v7 API plan.
Stacked on #839. Non-breaking — a test and a doc amendment.

The rule this enforces

27 public enums cross a binding by ordinal. JNI resolves constants with
values()[code] (jni_style.cpp:91); the apple and wasm mirrors are
positional too. Four headers say so, in comments:

// New entries go at the end: the bindings mirror this enum by ordinal.
// appended rather than sorted in, for the same reason as FileType
// Last because the bindings map this enum by ordinal.

A comment is not a guard. Reordering FileType compiles, links, passes the
suite, and then silently mis-maps in a consumer that was built against the old
jar. This pins all 214 enumerators so that fails here instead.

Appending stays silent by design — a new enumerator adds a line. Inserting or
reordering fails loudly, in the same commit.

Why not explicit values in the headers

The plan originally said to write = 0, 1, 2, … into the enums. Doing it, I
don't think that earns its place, and I have amended the plan rather than
leave the two disagreeing:

• 214 enumerators across 27 enums is a lot of noise in the public headers.
• It does not actually stop anyone. Someone reordering renumbers as they go and
  the header still compiles.

A test is what catches it. wasm/tests/enums.test.mjs already does exactly
this on the JS side; this is the C++ half, and the file says so.

Proof it works

Swapping soft and hard in HtmlTableGridlines — a two-line, entirely
plausible tidy-up:

test/src/enum_ordinals_test.cpp:232: Failure
test/src/enum_ordinals_test.cpp:233: Failure
[  FAILED  ] EnumOrdinals.html_table_gridlines

(reverted; the swap was only to check the guard bites)

Notes

The list was generated from the headers, not typed, so it cannot disagree with
them by transcription. 27 tests, one per enum, and they run in 0 ms.