Phase 1 / Forge / Tier 0 pre-freeze and C1.1 cost

.github/workflows/ci.yml

@@ -138,6 +138,29 @@ jobs:
         # figure rather than a steady state: the same cell ran in 1 minute when the
         # cache served it.
         precision: [false, true]
+        # M1.1.15.1 / H1 — ONE EXTRA CELL, AND DELIBERATELY NOT AN AXIS.
+        #
+        # `std.debug.assert` is compiled to NOTHING in ReleaseFast. The matrix above
+        # carries {Debug, ReleaseSafe} only, so every assert in the tree is verified in
+        # exactly the two modes where its breach costs nothing and in NEITHER of the
+        # modes a game ships. H1 was born in that gap: the entity-major premise of
+        # `Forge3DModule.dedupEntities` was held by an assert, so the guard was absent
+        # precisely where its breach silently returns an entity twice.
+        #
+        # ONE CELL AND NOT SIX, and the reason is the shape of the class it detects:
+        # a release-stripped assert is MODE-dependent, neither platform- nor
+        # precision-dependent. `-Doptimize` decides whether the assert exists at all;
+        # the OS and `-Dphysics_f64` cannot make it reappear or vanish. So one cell
+        # detects the whole class, and completing the axis would cost 50 % of the
+        # matrix for the same detection.
+        #
+        # **DO NOT "COMPLETE" THIS BY SYMMETRY.** If a future finding is
+        # platform-dependent or precision-dependent, that is a different class and it
+        # needs its own reason written here — not this one extended.
+        include:
+          - os: ubuntu-24.04
+            mode: ReleaseFast
+            precision: false
     runs-on: ${{ matrix.os }}
     # M0.1 hotfix — bumped from 10 to 20 min: Windows ReleaseSafe on the
     # 2-vCPU runner spends ~3 min on `zig build` then ~7 min on
@@ -186,7 +209,11 @@ jobs:
     # work, ~45 % inside the 20-min budget on the slowest runner. ReleaseSafe
     # keeps the cache: its near-cold recompile is what the 55-min budget exists
     # for, and it is the leg the cache was added for in the first place.
-    timeout-minutes: ${{ matrix.mode == 'ReleaseSafe' && 55 || 20 }}
+    # ReleaseFast joins ReleaseSafe on the 55-minute budget: both are optimising
+    # builds, and the 20 minutes were sized for Debug. Sizing the new cell to Debug's
+    # budget would make it fail on compile time rather than on the class it exists to
+    # detect, which is the worst kind of red.
+    timeout-minutes: ${{ (matrix.mode == 'ReleaseSafe' || matrix.mode == 'ReleaseFast') && 55 || 20 }}
     steps:
       - uses: actions/checkout@v6
 
@@ -881,7 +908,7 @@ jobs:
     # permanent, where a dispatch input leaves no trace in the repository at all.
     # Absent the trailer the job skips and costs nothing.
     #
-    # PERMANENT, not scaffolding. M1.1.25 replays the harness at N workers, M1.A on
+    # PERMANENT, not scaffolding. M1.1.21.1 replays the harness at N workers, M1.A on
     # a rebuilt scheduler DAG, and every shape added later moves the scenario —
     # each needs a regeneration with provenance.
     #

bench/forge_3d_character.zig

@@ -272,7 +272,6 @@ fn runBatch(gpa: std.mem.Allocator, scene: *Scene, mode: Mode, checksum: *f64) !
             },
             else => {
                 const r = try scene.chars.moveCharacter(
-                    gpa,
                     &scene.bp,
                     &scene.bm,
                     &scene.store,