acdream/docs/research/2026-07-25-slice-i7-closeout.md

Slice I7 — flat collision closeout

Result

Modern Runtime Slice I is complete. Production collision consumes the immutable prepared flat representation exclusively, per-resolve scratch is retained and reset-complete, and no parsed DAT collision graph remains in stable gameplay state.

The two production cutover commits are:

• flat-authoritative traversal: 068a06518dd710ca5e7f166754cbb7789ca1ed0f;
• parsed-graph removal: 82f8d4f82e24ad85042ac3e25c4f6464bebce758.

Exact rollback commands:

git revert 82f8d4f82e24ad85042ac3e25c4f6464bebce758
git revert 068a06518dd710ca5e7f166754cbb7789ca1ed0f

Revert graph removal first. Do not revert the I3 prepared-package, I4 flat-traversal, or I5 publication work unless evidence implicates those layers.

Production ownership after removal

• PhysicsDataCache publishes immutable flat GfxObj, Setup, CellStruct, and EnvCell topology records directly.
• Stable GpuWorldState strips temporary PhysicsDatBundle source material.
• Near LandblockBuild payloads discard Environment/GfxObj raw collision graphs after their immutable flat closure is built.
• Live objects publish through the same strict prepared source.
• Graph constructors remain only as explicit test, bake, and equivalence oracles. Production has no graph fallback and no sampled referee.
• The single validated acdream.pak mmap remains the production render and collision source.

This is a storage/ownership change only. Collision comparisons, traversal order, source float bits, retail thresholds, cell semantics, and authoritative game behavior did not change.

The retail divergence register was audited. No row was added or retired: Slice I changes the representation and lifetime of already ported collision data without introducing a behavior deviation.

Automated evidence

• Release build: pass with zero errors and zero warnings.
• Complete Release solution: 8,413 passed / 5 skipped.
• Focused Core collision, doors, doorway, cellar/ramp, jump, projectile, and allocation gate: 154 passed.
• Focused App camera/projectile gate: 47 passed.
• Player, grounded walkable publication, remote, projectile, and camera profiles: 0 B/resolve.
• Direct flat CellSurface, walkable-plane, behaviorless-cell, stable-state stripping, and production graph-residency fixtures pass.

Connected lifecycle/reconnect gate

Report: logs/connected-world-gate-20260725-182242/report.json in the exact-binary gate worktree.

• Source and binary: 82f8d4f82e24ad85042ac3e25c4f6464bebce758.
• Result: PASS; two graceful exits; zero failures.
• Six capped checkpoints plus one fresh-process uncapped reconnect passed.
• Parsed collision graph counts were 0/0/0 at every checkpoint.
• Flat GfxObj/Setup/cell residency remained populated, including:
  • capped login: 168/112/71;
  • Rynthid: 384/262/790;
  • Facility Hub: 436/272/279;
  • Holtburg after dungeon: 761/415/531;
  • fresh uncapped reconnect: 167/103/59.
• Publication, retirement, upload, warmup, and class backlogs converged.

The 25 world-edge misses are the route's existing expected diagnostic, not a new collision failure.

Connected canonical nine-stop gate

Report: logs/connected-r6-soak-20260725-182823.report.json in the exact-binary gate worktree.

• Result: PASS; graceful exit; zero failures.
• All nine portal checkpoints materialized and stabilized.
• Parsed collision graph counts were 0/0/0 at all nine checkpoints.
• Flat cell residency covered Caul, Sawato, Rynthid, Aerlinthe, and Holtburg, including 790 cells at Rynthid and 531 at Holtburg.
• Movement, jump/fall, combat/projectile, portal, dungeon, revisit, and authoritative server-echo checks passed.
• Final publication, retirement, worker, destination, control, unload, near/far, staged-upload, and composite-warmup backlogs were zero.

The known route warnings were limited to a Caul visible-workload delta of two, an eleven-particle plateau delta, 32 optional missing-DAT VFX diagnostics, and the 25 expected world-edge misses.

Measurement boundary

The user accepted collision correctness under RDP. The RDP nine-stop route also showed a bounded final plateau near 1.93 GiB working set while carrying observer/automation state, below the earlier dual-representation diagnostic route's roughly 2.06 GiB final Caul working set. Those are not controlled ordinary-client measurements, so no exact performance percentage is claimed.

Absolute post-removal CPU, GPU, frame-time, and ordinary-process memory comparison remains deferred to the next physical-display performance pass. That deferral does not weaken the deterministic, connected, ownership, or resource-lifetime acceptance above.

Next execution point

Slice J1 begins the borrowed-view, typed-command, ordered-event, clock, and lifecycle contract layer in AcDream.Runtime. It must adapt the current canonical owners first; it may not create a second gameplay world or add a frame of graphical input latency.