v1.1.6

FluidCore v1.1.6 — Constant-Width Ink, Unified Stroke Renderer, Memory Budget, & Stylus Barrel Buttons

FluidCore v1.1.6 is an ink-fidelity, memory-governance, and stylus-input release. It also hardens the Linux release pipeline, which had been structurally unable to produce a complete release.

1. Constant-Width Ink: Stroke geometry no longer tapers with pressure. A stroke is drawn at exactly the width the user selected, along its entire length.
2. Unified Stroke Renderer: All four ink renderers (document pane, infinite canvas, PDF export, excerpt crop tiles) now share a single Cairo rasterizer instead of four disagreeing copies.
3. Memory Budget: A single authority partitions the cache budget across page tiles, excerpt tiles, and undo history, sized from physical RAM and trimmed when idle.
4. Stylus Barrel Buttons: Barrel presses on a pen now erase instead of panning the canvas or opening a context menu.
5. Release Pipeline Repair: The Linux AppImage step could not succeed — its required tools were never installed — and two jobs raced to publish half-releases. Both are fixed.

---

🌟 What's New & Fixed in v1.1.6

1. Constant-Width Ink — StrokeWidthModel

New canonical header src/libfluidcore/geometry/StrokeWidthModel.h is now the single source of truth for ink width, derived by every renderer and by the eraser hit-test.

• Removed the pressure taper. Width was previously baseWidth * (0.25 + 0.75 * p), so the selected width meant something different at every point along the same stroke.
• Fixes a detached blob at pen lift. On lift, a stroke's tail collapsed toward the 0.25x floor while the round end cap kept its full radius, leaving what looked like a rendering fault at the end of otherwise clean lines.
• Fixes save/reload weight drift. Pressure is still captured and persisted with each stroke but no longer modulates geometry, so a reloaded stroke re-renders at the same apparent weight the user drew.
• Fixes eraser hit-radius mismatch. The eraser sized its hit radius from a pressure the ink no longer used.
• Shared highlighter opacity. kHighlighterAlpha = 0.5 is defined once, so a highlight reads identically in the pane, on the canvas, and in exported PDF.
• NaN-safe clamping. renderedWidth() and clampPressure() reject NaN, and a floor of kMinStrokeWidthPx = 0.5 keeps antialiasing from degenerating into a dotted line.

2. Unified Stroke Renderer — StrokeRenderer

New src/app/services/StrokeRenderer.{h,cpp} is the single Cairo entry point for ink geometry. The document pane overlay, the infinite canvas, PDF export, and excerpt crop tiles all call it.

• Fixes polygonal ink at high zoom. The old code flattened every span into a fixed 3–4 chords and stroked each chord separately. A fixed chord count is resolution dependent, so ink looked smooth at 50% zoom and visibly faceted at 200%.
• Fixes beaded outlines. Chained round caps at stepping widths produced a dashed-looking edge. The Centripetal Catmull-Rom spline is now emitted as cairo_curve_to spans and stroked once, so a single stroke is a single composite and Cairo's stroker subdivides adaptively at any zoom.
• Fixes export/display drift. The four renderers had independent copies of the rasterizer that disagreed on both flattening density and the pressure-to-width equation, so exported ink was narrower than what was on screen.
• Reflection end conditions. Phantom points reflect past each end of the polyline, so the spline is defined across the whole stroke rather than only between interior samples.

3. Memory Budget — MemoryBudget

New src/app/services/MemoryBudget.{h,cpp} arbitrates every byte-budgeted consumer in the app.

• Fixes a 2x budget overrun. Four independent caches each declared kDefaultMaxBytes = 64 MB and each enforced it only against its own surfaces, so a session could legitimately authorize 256 MB against a documented idle target of 120 MB (TRD.md §4). Nothing arbitrated between them and nothing ever called setMaxBytes().
• RAM-derived total. recommendedTotalBytes() derives the pool from physical RAM, clamped to [16 MB, 512 MB], falling back to 64 MB when RAM cannot be determined.
• Named slices. PageTiles, ExcerptTiles, and UndoStack each receive an explicit share. The partition is deliberately page-tile heavy, because page tiles back the reading experience the user is looking at right now, excerpt tiles are card thumbnails that re-render cheaply, and undo history is depth-bounded in practice.
• The two UndoStacks split their slice. There is an independent stack for the document pane and the workspace canvas, so undoBytesPerStack() divides evenly. UndoStack lives in libfluidcore and must not depend on the GUI (ADR-0001), so the frontend sets each stack's budget explicitly at construction rather than the engine reading the pool.
• Idle trim and restore. trimToIdle() lowers every slice to 25% of its configured size, releasing cache surfaces back to the allocator instead of holding them until exit; restoreFromIdle() puts them back. Both are safe to call repeatedly.

4. Stylus Barrel Buttons — StylusButtonRouter

New src/app/services/StylusButtonRouter.{h,cpp} maps GDK stylus buttons and device events onto ink intents.

• Fixes barrel press panning the canvas. GDK's Win32 backend already surfaces Wintab hardware as ordinary GDK input — a pen arrives as GDK_SOURCE_PEN with a GDK_AXIS_PRESSURE axis, the eraser end as GDK_SOURCE_ERASER, and barrel buttons as plain GdkEventButton presses on that pen device. Left alone, a barrel press was indistinguishable from a middle click, so it panned the canvas or opened a context menu instead of erasing.
• Stylus overrides mouse-oriented tools. select, text, crop, and rect_select are mouse modes that start a selection and consume the press. A stylus now overrides them; otherwise touching the page with a pen while the select tool was active selected text instead of drawing, forcing a mouse tool switch before every annotation.
• Eraser is honoured as selected. An earlier revision overrode unconditionally, which meant a stylus press with the eraser selected drew a stroke at the eraser's forced width instead of erasing.
• Barrel-button state machine. isBarrelButton() lets the eraser be held for the duration of a press.
• No Windows Ink required. This is plain GDK input routing, the layer the rest of the app is already built on.

5. Release Pipeline Repair

The Linux release could not have succeeded, and this is the direct cause of repeated re-release attempts.

• AppImage step was guaranteed to fail. appimagetool hard-die()s with "mksquashfs command is missing but required" and "desktop-file-validate command is missing but required" before doing any work. Neither squashfs-tools nor desktop-file-utils was in the runner's package list, so no AppImage was ever produced. Both are now installed in ci.yml and release.yml.
• Two jobs raced to publish. release-windows and release-linux each called softprops/action-gh-release for the same tag with no ordering. The first to finish published a live release containing only 3 of 6 assets, and a simultaneous create could return 422 already_exists. Release is now a single publish job gated on needs:, with fail_on_unmatched_files: true so a missing asset fails loudly instead of shipping a partial release.
• Node 20 actions can no longer run. GitHub removed Node 20 from hosted runners on 2026-09-23 with no opt-out, which stranded actions/checkout@v4, actions/upload-artifact@v4, and softprops/action-gh-release@v2. All are on Node 24 majors now.
• ubuntu-22.04 is deprecated. The image entered deprecation on 2026-09-17 with hard brownouts scheduled before its 2027-04-17 retirement. Both workflows are pinned to ubuntu-24.04; ubuntu-latest is also pinned because it migrates to ubuntu-26.04 between 2026-10-19 and 2026-11-19.
• Portable zip used Windows path separators. Compress-Archive under Windows PowerShell 5.1 writes share\glib-2.0\schemas\gschemas.compiled into the archive, which violates the ZIP specification. Non-Windows extractors materialised those as single flat files instead of a directory tree. The packager now writes entries by hand and then verifies that no entry contains a backslash and that the entry count matches.
• Windows checksums were a truncated table. Get-FileHash | Out-File emits a formatted table whose Path column is ellipsis-truncated at console width, and it is not parseable by sha256sum -c. Both platforms now emit a standard <hash> <bare-filename> file.
• Inno Setup could not accept a path with spaces. /DSourceDistDir= and /DOutputDir= were passed unquoted through a space-joining argument list, and fluidcore.iss used an unquoted OutputDir=. Both are now quoted.
• Cairo pin now fails loudly. curl -sSL had no --fail, so a 404 body was written to disk with exit 0 and pacman then failed on a non-zstd file. The pin is -fsSL and short-circuits when the installed Cairo is already the pinned version.
• Added the missing MSYS2 dependency. mingw-w64-ucrt-x86_64-gtk-update-icon-cache was required by CLAUDE.md and ops/CONTEXT.md but absent from both workflows, leaving the bundled share\icons\hicolor tree without a cache.
• Added .gitattributes. Without it, Windows checkouts silently rewrite LF to CRLF, which breaks ops/scripts/*.sh (bad interpreter: bash^M) and makes ops/patches/*.patch unapplyable because its context lines no longer match. Line endings are now pinned per file type.
• Added concurrency, timeout-minutes, and least-privilege permissions to both workflows, and workflow_dispatch to release.yml so a release can be re-run without deleting and re-pushing a tag.

---

📦 Platform Binaries & Packaging

• Windows 11 (Native):
  • Inno Setup installer: FluidCore-Setup-x64.exe (v1.1.6)
  • Portable zip: fluidcore-windows-x64.zip (v1.1.6) — now POSIX-separated
  • Checksums: checksums-windows-sha256.txt (now sha256sum -c compatible)
  • Resources: src/app/fluidcore.rc version set to 1.1.6.0
• Linux (Ubuntu / Debian / Fedora / Arch):
  • AppImage: FluidCore-1.1.6-x86_64.AppImage
  • Debian package: fluidcore_1.1.6_amd64.deb
  • Checksums: checksums-linux-sha256.txt
  • AppStream metainfo: resources/linux/org.fluidcore.platform.metainfo.xml updated with v1.1.6 release metadata

---

🧪 Verification & Test Results

• Windows Ninja Build: Clean build with ops\scripts\build-win.ps1 via MSYS2 UCRT64, Cairo pinned to 1.18.4-4.
• CTest Suite: 42/42 tests passed (100% pass rate), including 3 new suites — StrokeRendererTest, MemoryBudgetTest, StylusButtonRouterTest — plus extended ExcerptTileCacheTest and PageTileCacheTest.
• New coverage:
  • StrokeRendererTest — width constancy, NaN rejection, zoom-independent smoothness, pane/export parity.
  • MemoryBudgetTest — slice partitioning, RAM-derived clamping, idle trim/restore round-trip.
  • StylusButtonRouterTest — barrel-button classification, tool override precedence, genuine mouse passthrough.
• Memory & Scalability: ScalabilityBenchmarkTest passes against the $\le 1.2\text{ GB}$ ceiling.
• Offline guarantee: Zero runtime network calls; the Cairo pin and appimagetool download are build-time only.