feat(data/json): unify JSON escapers and add streaming NDJSON emitter

[octalide]

Closes #338.

What

Moves the streaming ASCII escaper out of the mach compiler's mach.cli.json and into std.data.json as the one JSON emission home, unifying it with the tree emitter's #337 structural escaper into a single escape core.

One escaper, explicit policy knob

Both emission surfaces now escape string bodies through one core (escape_unit) selected by EscapePolicy:

• ESCAPE_VERBATIM (structural-utf8-verbatim) — escapes ", \, and the C0 controls only; every other byte, including UTF-8 lead/continuation bytes, passes through verbatim. The tree emitter emit() uses this, so valid UTF-8 round-trips as UTF-8.
• ESCAPE_ENSURE_ASCII — as above for bytes < 0x80, but each byte >= 0x80 is UTF-8 decoded (full RFC 3629 validation) and emitted as a \uXXXX escape (astral code points as a surrogate pair); invalid UTF-8 emits the U+FFFD replacement escape, so every emitted byte is printable ASCII. The streaming NDJSON surface uses this.

The tree emitter now renders through a counting io.Writer (BufWriter) instead of a bespoke Emitter, so both surfaces share one sink and one escaper. emit() keeps its exact measure-when-buffer-short semantics (returns the required length even when the buffer is too small).

Streaming NDJSON surface (full flat + nested/array)

Ported verbatim from the current-dev mach.cli.json (which mach#1838 extended past the original flat surface): Object, Array, object_begin/object_end, the nesting primitives object_end_value/field_object_begin/field_array_begin/array_end/array_object_begin, and the members field_str/field_str_or_null/field_null/field_i64/field_bool, plus write_json_string. The Object.first/Array.first comma-state machine is ported unchanged. object_end closes via object_end_value and writes the sole NDJSON line terminator, so a whole nested object lands on one line. Public names match the compiler's exactly, so the mach-side rewire (refactor/338-json-to-std, rewiring both testing.mach and diagnostic.mach/flush_json) is a use swap with no call-site churn.

Placement rationale

Kept in std.data.json itself (not a sibling module). The one escape core stays private, shared by both public emission surfaces; a sibling module would force exporting the escaper or adding a third module just to share it. "One home" = one module.

Byte-identity

• Tree emitter: the pre-existing exact-byte emit tests (null/number/bool, array/object round-trips, quote/backslash/control escapes, object keys) all still pass unchanged.
• Streaming emitter: the exact-byte write_json_string tests were ported over from mach.cli.json (identical inputs and want byte arrays); mach#1838's compose tests (nested object + array-of-objects {"a":1,"loc":{"line":2,"col":3},"rel":[{"label":"x"},{"label":"y"}]}\n, and empty array {"rel":[]}\n) were ported verbatim; and a flat full-object sequence test asserts {"n":7,"s":"a\"b","b":true,"o":null,"z":null}\n byte-for-byte (exercising field_null).

Scope

• #337/#340 parse/emit representation-asymmetry warnings preserved intact; #340 stays open.
• CHANGELOG.md [Unreleased] entry added.

Verification

mach build . and mach test . green — 588 passed, 0 failed (36 in std.data.json: 23 tree + 13 streaming). Both CI checks (build + cross-riscv64) pass.

Note: #338 is a mach-std issue; this merge closes it on the mach-std side. The paired mach-side refactor (refactor/338-json-to-std) lands separately after the mach-std release and mach.lock bump.

🤖 Generated with Claude Code