json: document that an unquoted float can be rounded twice

[farmer-data]

Refs #647. This documents the behaviour rather than changing it, so the issue can stay open to track a real fix through #482.

An unquoted JSON float with a fraction or exponent reaches json_helpers::float as the f64 serde parsed and is then narrowed with as f32, so it is rounded twice. A quoted decimal, and an unquoted integer in the i64/u64 range, is rounded once. A decimal close enough to an f32 midpoint becomes that midpoint as an f64 and then rounds to even, so 1.0000000596046448 decodes to 1.0 unquoted and 1.0000001 quoted.

The visitor sees only the f64, and an f64 sitting exactly on a midpoint does not say which side the decimal was on, so the helper cannot correct it. Rounding once needs the token text: arbitrary_precision (#482), or a raw-value capture, which would tie the helpers to serde_json while they stay generic over Deserializer. As the issue suggests for that case, this PR documents the behaviour:

• json_helpers::float rustdoc: when the two paths differ (by one f32 ulp), the issue's example, the overflow-threshold case (rejected unquoted, accepted quoted), that protobuf C++ narrows the same way while protobuf-go rounds once, and that only a producer can avoid it, by quoting.
• A pointer to it on FloatValue's Deserialize impl (buffa-types/src/wrapper_ext.rs) and in DynamicMessage::from_json, since users of generated messages, wrappers and reflection never name json_helpers directly.
• A paragraph in the JSON section of docs/guide.md, next to the arbitrary_precision note.
• Tests pinning the double-rounded and single-rounded results, and the single rounding of integer tokens, for the generated helper and for DynamicMessage::from_json. They compare exact bits, so a future single-rounding fix will fail them and prompt a docs update. The example input gives the same f64 with or without serde_json's float_roundtrip.
• A Changed changelog fragment marked as documentation, with no behaviour change.

Verification

• task lint, task doc and task test-local (4186 passed, 0 failed), with the repository-pinned protoc 33.5.
• Reviewed with the repo's rust-code-reviewer and rust-api-ergonomics-reviewer per CLAUDE.md: no Critical or High findings. Addressed the Mediums: the midpoint bound in the rustdoc no longer assumes float_roundtrip, the note now also appears on FloatValue, DynamicMessage::from_json and in the guide, and the guidance says only the producer can avoid the difference. Also addressed the Lows: integer tokens are excluded (with a test), the overflow case is noted, and the changelog is Changed with "No behaviour change".

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[iainmcgin]

[claude code] Pushed one commit (befada9) that removes .changes/unreleased/changed-20261013-062603.yaml. The PR is documentation plus two tests that pin current behaviour, and a docs-only change carries no fragment here: a Changed entry would read as a behaviour change in the release notes. Everything else is as you wrote it.