[imakris]

Purpose

Plot font atlases retain a fixed 2048-square bitmap even when their glyph inventory fits in a smaller canvas. Enable automatic sizing while preserving glyph scale, CPU bitmap lifetime, and the cache memory budget.

Changes

Enable options_t::auto_atlas_size and include that policy in the font digest. Accept the resulting power-of-two disk-cache extents from 1 through 2048 only when the RGBA byte count matches. Preserve the disk format, 128 MiB CPU retention budget, actual-size accounting, and bitmaps needed for renderer recreation. Extend disk-cache coverage for supported and malformed extents and payload lengths.

Add persistent compiler caches to the four native CI workflows. Use Visual Studio 2026 and MSVC v145 on Windows, with explicit environment/compiler checks and Ninja so the compiler launcher is effective. Reuse prebuilt Qt and preserve dependency declarations tracking master.

The prerequisite vnm_msdf_text PR #2 has merged with all checks passing.

Validation

• All eight native CI jobs and CodeQL pass. Text suites pass 43 tests each on Linux, FreeBSD and macOS, and 44 on Windows, including FontDiskCache, FontAtlasCache, FontIdentity, FontRendererBounds, FontRendererDrawRanges and PlotRendererLifecycle.
• Native D3D11 Release comparison used three interleaved baseline/candidate pairs at DPR 1 and 2, each with 300 measured frames through the production renderer and GPU completion. Atlas RGBA storage falls from 16 MiB to 4 MiB at both DPIs. Frame timings show no repeatable material regression; no universal speedup claim is made.
• All 248 glyph metrics, kerning, coverage and glyph bitmap crops match. DPR 2 rendered pixels match exactly. DPR 1 differs at 16 of 864,000 pixels by at most one channel unit. Each variant recreates renderer resources with exact pixels, retained CPU atlas storage and one atlas upload per renderer.
• Compiler-cache statistics prove persistent reuse and FreeBSD cache save/copyback. Individual Linux cache stores reported errors, affecting only future reuse; subsequent validation recorded 165 hits out of 173 requests with no compilation or test failure.

Independent source and validation review found no remaining product blocker. One-off benchmark instrumentation is retained outside the repository and is excluded from the shipped changes.