Anchor an entering unit's sprite on its own tile, not its map slot

[Giszmo]

Fixes #230 — globs jump back one square when they enter a building.

Cause

PR #218 (3ad40cf3b, merged today) changed Game::drawUnit to derive the sprite origin from the tile the draw loop found the unit on rather than from the unit's own position:

-	map.mapCaseToDisplayable(unit->posX, unit->posY, &px, &py, viewportX, viewportY);
+	int px = x * 32, py = y * 32;

That was needed: re-converting the canonical position collapses a wrapped tile visible at both screen edges into a single occurrence, so a unit standing on a map seam lost its opposite-edge copy.

The two tiles agree for every action that clears the old map slot and claims the destination one. Entering a building is the deliberate exception — Unit::handleActionEnteringBuilding leaves the map slot on the tile the unit is leaving ("we don't delete the unit on the map because we have to draw it while it is entering") while posX/posY already name the building tile. Game::integrity() enforces that invariant, and test/EnteringUnitSaveHarness.cpp asserts both halves of it.

So the arrival interpolation px -= (dx*(255-delta))>>3 started from the tile behind the unit: the glob snapped back a square and walked into the square it had just left.

Fix

Keep the per-occurrence anchor, and recover the unit's own tile from it by folding in the signed wrapped offset to posX/posY. This is the same anchor-recovery idiom PR #218 already uses for building footprints in drawMapGroundBuildings, made signed because a unit's slot can sit one square behind its position, not only ahead of it. When the two tiles coincide — every case but entering a building — the result is x unchanged, so #218's seam behaviour is preserved exactly.

The arithmetic is in src/render/UnitDrawGeometry.h so it is testable without a GraphicContext.

Verification

Rendered proof. test/EnteringUnitDrawHarness.cpp (new) draws the actual state through Game::drawMapGroundUnits and compares framebuffers. Nothing in Game::drawUnit reads displacement, so at equal delta an entering unit must render pixel-for-pixel like the same step expressed as an ordinary walk onto the destination tile. Against the pre-fix anchor it reports exactly -32 px at delta 0 — one tile backwards, the reported symptom. With the fix all five sampled deltas match:

delta   0: entering sprite x 268..282, walking x 268..282
delta  63: entering sprite x 273..294, walking x 273..294
delta 128: entering sprite x 276..299, walking x 276..299
delta 191: entering sprite x 282..303, walking x 282..303
delta 255: entering sprite x 294..310, walking x 294..310
Entering-unit draw regression passed: 5 deltas render identically to the equivalent walk

A second pass in the harness captures terrain, glob and inn together so the animation can be looked at by eye. One step into an inn, before and after — top row is master at 87bdee0fd, where the glob starts a square left of where it is and arrives on the grass instead of the inn:

Also:

• test/UnitDrawGeometryTest.cpp (new, 4 cases): the coinciding case over every viewport/tile combination; the entering-building lag in all eight directions over every viewport/tile combination; the two seam occurrences staying distinct; and both properties together for an entry across the seam.
• Confirmed the test binds to the fix: with the helper stubbed back to the pre-fix return slotTile, testSlotLagsWhileEnteringBuilding and testSeamOccurrencesWhileEnteringBuilding fail; with the fix, ./TestsRunner reports OK (175 tests).
• scons -j4 release=1 server=0 builds clean, and EnteringUnitSaveHarness passes both CI invocations (40 directions/positions and corruption controls, plus the saved-fixture load).

One thing to apply by hand: the draw harness should run in CI, but the push token has no workflow scope so this branch cannot touch .github/workflows/build.yml. The job already installs xvfb and mesa for the fullscreen aspect harness, so the step is:

      - name: Build and run the entering unit draw regression
        run: |
          scons -j$(nproc) release=1 server=0 entering-unit-draw-test
          timeout 300s xvfb-run -a -s '-screen 0 1024x768x24' ./build/src/EnteringUnitDrawHarness

test/README.md carries the same snippet.

Base: 87bdee0fd.

🤖 Generated with Claude Code