json.mqh: fix O(n^2) JSON serialization (46s -> 0.5s for a month of M1 candles)

[biohazardxxx]

Summary

JSONObject::toString() and JSONArray::toString() in json.mqh build their output by repeated string concatenation — every element copies the entire accumulated string, making serialization O(n²) in the output size. For large responses this dominates end-to-end latency: serializing ~30k MqlRates (~6 MB of JSON, i.e. one month of M1 candles via CopyRates) took ~46 seconds inside the expert.

This PR makes serialization linear:

• New JsonBuffer class: writes characters into a pre-grown ushort array (in-place StringToShortArray, amortized-doubling ArrayResize) and converts to a string once at the end.
• New virtual void serialize(JsonBuffer &buf) on JSONValue; JSONObject/JSONArray override it to serialize recursively into the single buffer. Scalar types keep their existing toString() (used by the base serialize).
• JSONObject::toString() / JSONArray::toString() keep their signatures and now delegate to serialize — no call-site changes anywhere.
• Applied identically to mq5/json.mqh and mq4/json.mqh (same defect in both).

Why an array buffer instead of string appends

Measured on MT5 build 5830: MQL string appends cannot be made linear. +=/StringAdd re-allocate the string to its exact new size and discard any capacity set with StringReserve, so an "efficient append" still copies the whole string per call — a StringReserve+StringAdd variant of this fix measured 20× slower than the original code (the re-reserve after every append doubles the copying). The ushort[]-backed buffer is the reliable in-place path.

Measured results (XAUUSD M1, live MT5 terminal, build 5830)

bars	before	after
1,000	116 ms	74 ms
8,000	3,012 ms	225 ms
16,000	15,668 ms	268 ms
32,000	50,968 ms	694 ms
Aug-2021 month range (30,281)	46,224 ms	548 ms (~84×)

Before: ms/bar grows 0.12 → 1.59 (quadratic). After: flat ~0.02 ms/bar (linear).

Testing

• Both MtApi5.mq5 and MtApi.mq4 compile with Result: 0 errors, 0 warnings; recompiled .ex5/.ex4 included (built against this branch's json.mqh).
• Runtime-verified against a live MT5 terminal: the scaling measurements above, plus functional checks on the same session (CalendarCountries, Print, PositionsTotal) confirming response JSON still parses correctly on the .NET side.
• MQL4 change is source-identical; compile-verified only.

🤖 Generated with Claude Code