"ignore struct length" flag

[Ponywka]

(Ported from vmihailenco#382 by vmihailenco#382 (comment) request)

In some scenarios, it is necessary to ignore the structure length.

One such example is Tarantool. Due to the specifics of the Lua language, a structure that has an empty value at the end (for example, {0, nil, 1, 2, nil}) is truncated to the last non-nil value and becomes {0, nil, 1, 2}.

More details are available in the Tarantool documentation.

Despite this behavior, the official Go module github.com/tarantool/go-tarantool/v2 still uses the github.com/vmihailenco/msgpack module in the decoder (see here), which reports an error in such cases:
msgpack: number of fields in array-encoded struct has changed.

This PR introduces an option to ignore this error and also fixes decoding when the number of fields differs from what is defined in the struct.

[xe-nvdk]

Thanks for porting this from upstream — the core change is small, correct, and the trailing skipNext(n) is the right call to keep stream position consistent for reused decoders. A few requests before merge:

1. Test-local constant shadows the package constant. types_test.go:36-38 defines ignoreStructLength uint32 = 1 << iota inside the test file, same name as the unexported package constant in decode.go:26. They evaluate to the same bit today by coincidence (both 1 << 0 in their respective iota groups), but if the package-level iota block is ever reordered the tests will still pass while real users break. Could you rename the test constant to something like testIgnoreStructLength, or drop it entirely and just call dec.IgnoreStructLength(true) directly from the test setup?

2. Doc comment needs more detail. decode.go:265 currently says only IgnoreStructLength disables field count verification during parsing. Users reaching for this won't know the actual semantics — that extra payload elements are skipped, and missing struct fields are left at zero value. Could you expand the comment to spell that out? Something like:

   // IgnoreStructLength disables the field-count check when decoding an
   // array-encoded struct. When enabled, extra elements in the payload are
   // skipped, and missing trailing fields are left at their zero value.
   // This is useful for sources like Tarantool that truncate trailing nils.

3. Naming consistency (optional). Every other flag in that iota block ends in Flag (looseInterfaceDecodingFlag, disallowUnknownFieldsFlag, disableAllocLimitFlag). Renaming ignoreStructLength → ignoreStructLengthFlag would match the dominant pattern. Not a blocker but worth doing while you're in there.

4. CHANGELOG entry. This is a public API addition and our project guidelines ask for divergence from upstream to be documented. Could you add an entry to CHANGELOG.md?

Once those are addressed I'll run the post-implementation review pass and we can get this merged. Thanks again!

[xe-nvdk]

Hi @Ponywka — following up on this one. The four items from my review above are still open, and while re-reviewing I found two more things worth addressing. Full list so you have everything in one place:

From the original review (still outstanding):

1. Test-local ignoreStructLength constant in types_test.go shadows the package constant — rename to testIgnoreStructLength or drop it and call dec.IgnoreStructLength(true) directly.
2. Expand the IgnoreStructLength doc comment to describe the actual semantics (extra elements skipped, missing trailing fields left at zero value).
3. Optional: rename ignoreStructLength → ignoreStructLengthFlag to match the other flags in that iota block.
4. CHANGELOG entry for the new public API.

New findings:

5. skipNext error is discarded. In decodeStructValue:

   if n > 0 {
       d.skipNext(n)
   }

   skipNext returns an error (it calls d.Skip() in a loop, decode_query.go:150). Dropping it means a truncated or malformed tail is silently swallowed and the decode returns success. Should be:

   if n > 0 {
       if err := d.skipNext(n); err != nil {
           return err
       }
   }

6. The loop rewrite runs unconditionally. The n-- / break / trailing-skip bookkeeping applies even when the flag is off. It's functionally equivalent in that case — the length check above guarantees n == len(fields.List) — but it adds per-field work to the struct-decode hot path, which this fork is specifically tuned for. Either gate the new logic behind the flag, or post a BenchmarkStructUnmarshal before/after showing it's free. Happy with either.

The feature itself is well-motivated and I want to land it. If you're short on time, say the word and I'll finish these off on top of your commit with authorship credit preserved — no problem either way.