fix: own legend label bytes in LegendEntryLabel forwarder (macOS issue #9)

[borisbat]

Summary

Fixes #9 — on macOS the Y2-routed series' legend label came back empty in the v2 plot snapshot. Root-caused as a dangling internal-pointer alias, not a label-resolution lag.

LegendEntryLabel (src/dasIMPLOT.main.cpp) returned the raw const char* from ImPlot::...GetLegendLabel(i) = Legend.Labels.Buf.Data + NameOffset, a pointer straight into ImPlot's per-frame Legend.Labels buffer. daslang aliases a const char* return (no copy), and capture_legend stored the alias. The buffer is valid at capture time (post-EndPlot), but the snapshot is serialized a frame later over the live API — by then the next frame's BeginPlot/SetupFinish → Legend.Reset() has cleared and (on macOS) realloc'd the buffer, leaving the das string dangling. Windows/Linux kept the same allocation and rewrote identical bytes in place, so the alias still read correctly there; macOS surfaced empty/garbage labels.

Diagnosing on a macOS box showed it's worse than "empty": both labels came back as garbage bytes, not just the Y2 one — the test only asserted pressure, so temp's corruption went unnoticed. Same root cause.

Fix (C++ wrapper, not the das caller)

We own the LegendEntryLabel forwarder (it's hand-written, unlike the generated bindings), so the right place to own the bytes is the wrapper itself — matching the rest of the string-returning API. Threaded das::Context* / das::LineInfoArg* and return context->allocateString(...), so the returned das string owns a heap copy that stays valid regardless of ImPlot's buffer churn. Same idiom as dasImgui's text_range_string / ImGTB_Slice.

(First pass cloned on the das side with clone_string in capture_legend; moved to the wrapper per review — keeps callers from having to remember to clone, consistent with the rest of the API.)

Also removed the macOS skip in test_multi_axes — the pressure/Y2 legend assertion now runs on all platforms.

Verification (macOS, headless, CI-equivalent)

Reproduced and fixed in a fresh workspace mirroring tests.yml exactly (own daslang-src + dasImgui + dasImguiImplot, Ninja Release build, daspkg install --global both packages):

• Before: forcing the skipped assertion → wait_for_series_shown(... "pressure" ...) times out; raw legend dump shows garbage labels for both series.
• After: labels resolve to temp / pressure; full integration suite passes 13/13 (--isolated-mode --isolated-mode-threads 4 --headless), including the de-skipped test_multi_axes.
• Both changed .das files lint clean.

Audit (same internal-pointer pattern elsewhere)

Swept every native string return that das might store:

• LegendEntryLabel — the one live bug. Fixed here.
• GetColormapName (generated binding) — same internal-ImGuiTextBuffer pattern (Text.Buf.Data + TextOffsets[cmap]), but no das caller stores it today → latent only. If it's ever wired to das, give it the same allocateString treatment.
• GetStyleColorName / GetMarkerName — static string literals → safe.
• dasImgui text_range_string / ImGTB_Slice — already use context->allocateString(...) → safe; combo-getter callbacks are transient → safe.

🤖 Generated with Claude Code