Write the floats of json_view from their digits

[nlohmann]

Part of the stack for the zero-copy view (#5295). Depends on the previous PR (shortest doubles with Żmij).

Summary

json_view::dump() writes a float token of at most 15 significant digits from its digits, without converting it to a double and back. The output is unchanged: it is what json::dump() writes for the same value.

• Why this is exact. Two decimals of at most 15 significant digits are farther apart than the rounding interval of a normal double; this is the argument behind DBL_DIG. So the token's digits, without trailing zeros, are the shortest digits of its double. The library now writes exactly those (Żmij); with Grisu2 this would not hold.
• When it applies. The exponent must keep the value away from subnormals and overflow (leading digit between 10^−290 and 10^304).
• Other tokens (16 or more digits, or edited values) are converted from the digits already read, with the library's decimal_to_float() (Speed up the lexer: own float parser, string scan, and \u table #5738), so the token is not read twice.
• Doubles are written into the output directly, not through a local buffer.
• With NEON, the fixed layouts (12.5, 0.001, 100.0) are put together in vector registers by a table lookup of the digit bytes, as in the prototype. The library's portable layout copies the digits through a buffer at another offset, and a load that spans several recent stores waits until they reach the cache. x86 and JSON_VIEW_NO_SIMD use the library's layout.

Benchmarks

json_view::dump(), before vs. after this PR, on Apple M1 (best of 7 alternating processes):

file	before	after	ratio
numbers	450 µs	140 µs	0.31
marine_ik	5.19 ms	1.96 ms	0.38
mesh.pretty	1.66 ms	1.09 ms	0.66
canada (mostly 16–17 digits)	5.32 ms	4.57 ms	0.86
citm_catalog	144 µs	145 µs	1.00
twitter	73.2 µs	73.2 µs	1.00

canada's tokens mostly have 16 or 17 digits, so they are converted; its gain comes from writing in place and from the NEON layout.

Tests

• 20,000 float tokens with 1 to 17 significant digits, in every spelling: with and without a point, exponents with e/E and signs, leading zeros (0.000123), trailing zeros, and negative values, from about 1e−320 to 1e300 (subnormals included). json_view::dump() must equal json::dump() of the same text.
• On AArch64, the same tests check the NEON layout; on x86 and with JSON_VIEW_NO_SIMD, the library's layout.

Public API

No change.

---

Written by Claude Code.

🤖 Generated with Claude Code

[nlohmann]

CI fix for this PR (commit 977bb1b, plus all lower fixes merged in). CI had not run yet; this was found locally with gcc:16 (linux/amd64) and the CI flags:

• ci_test_gcc: -Werror=useless-cast on static_cast<std::size_t>(1 + (tokens() % 17)) and static_cast<std::size_t>(tokens() % (digits.size() + 1)) in the float-token test of unit-json_view.cpp. std::mt19937_64 yields std::uint_fast64_t, which is std::size_t on Linux. The numbers are now drawn through a lambda that casts a named std::uint64_t. That lambda also makes the std::string(count, '0') conversions explicit where std::size_t is 32 bits (MSVC Win32). The sequence of draws is unchanged.

Verified: unit-json_view builds and passes (C++11/C++20 with the CI flags, and a JSON_NOEXCEPTION build run as --no-throw). All view tests also pass a silkeh/clang:4 syntax check on this branch.

— posted by Claude Code on behalf of @nlohmann

[github-actions]

🔴 Amalgamation check failed! 🔴

The source code has not been amalgamated and/or formatted correctly, or BUILD.bazel is out of date.

[nlohmann]

The zero-copy view stack (#5295) was rebased onto develop and regrouped from 20 PRs into 10, one squashed commit each. The changes of this PR are now in #5633; its review comments stay here for reference.

---

Written by Claude Code.