erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases 6 Packages Wiki Activity Actions

fix(runtime): give the no-window host a post-merge canonical cell commit (D1, AD-60/AD-64, AP-146/#320)

Browse source 
C5b (735f0a72) made the steady-state accepted-Position merge stop writing
residency. That is retail-correct — HandleReceivedPosition @0x00453FD0 reads
the wire objcell_id into a local and never assigns the object's cell — and it
stays. What C5b did not account for is that its replacement writers both live
in AcDream.App: the OnPosition prologue rebucket (AD-60's W2) and the
post-routing wire-cell adopt (W3, AP-135).

The two hosts run parallel, non-shared inbound routes. LiveEntitySessionController
-> LiveEntityNetworkUpdateController.OnPosition is graphical-only;
RuntimeLiveEntitySessionController.OnPositionUpdated is the no-window route and
is constructed only at HeadlessSessionHost.cs:682. So AcDream.Headless had NO
post-merge cell writer at all. Every remote's FullCellId was written at
create/placement and then frozen for the session — and RuntimeEntityObjectViews
.Snapshot projects exactly that field as RuntimeEntitySnapshot.CellId, i.e. every
bot's entire world view. The local player lost one of AP-146's three refresh
edges, which matters beyond cosmetics: RuntimeSetPositionState
.IsAffectedCollisionResident reads FullCellId to pick which bodies a landblock
retirement parks, so a bot running A->B without teleporting would have retired A
while parking a body physically in B.

The fix, in three parts:

1. RuntimeEntityObjectLifetime.CommitWireCellRebucket — a new Runtime owner for
   the committed VALUE, extracted verbatim from LiveEntityRuntime
   .RebucketLiveEntity. This is also the root-cause fix for the layering
   inversion the review found: AD-60 was documenting its own correctness by
   naming an App class the Runtime assembly cannot reference. Behaviour on the
   graphical side is unchanged — record.FullCellId is a proxy for
   record.Canonical.FullCellId, which is the record the callee reads, and the
   commit is still CommitRebucket. Verified load-bearing for BOTH hosts:
   sabotaging the preserve branch reddens the graphical
   LiveEntityRuntimeTests.CanonicalOnlyRebucket_DoesNotOverwriteAuthoritativeFullCell
   as well as the new headless assertion.

2. RuntimeLiveEntitySessionController.TryCommitAcceptedWireCell — the no-window
   W2, under the same reachability rules the graphical route applies: Rejected
   writes nothing (the shape the App authority gate produces by returning false);
   a bound-projectile packet writes nothing (routed by the graphical host through
   the canonical projectile placement owner, which returns before W2); an active
   initial-create residence writes nothing (RebucketLiveEntity's own early
   return — while the lease is live the SetPosition conductor is the sole cell
   authority); a local ForcePosition writes only when the accepted-Position drive
   declined it (NotApplicable), because a handled force is
   placement-receipt-authoritative. W2/W3 themselves are untouched.

3. On the committed value (the landblock-vs-cell trap). RebucketLiveEntity's
   preserve branch fires on a LANDBLOCK-shaped id — low 16 bits 0xFFFF — and
   exists for LocalPlayerProjectionController.Project, the per-frame local
   movement caller that emits exactly that shape. An inbound wire objcell_id is
   never landblock-shaped, so on the accepted-Position route the branch is not
   taken and the exact wire cell is committed. That is what W2 commits today and
   what this now commits; the no-window host has no per-frame caller at all.

Ordering is matched, not improved on: the force drive submits its placement
before the commit, so its first submit still reads the pre-commit FullCellId —
AP-138's amended route-2 CurrentCellId measurement.

Bookkeeping in this commit:
- AD-60 corrected. Its surviving-channel enumeration presented "the local force
  path, the missile arm" as exhaustive; the entire no-window host belonged in it.
  23aa62f2's W2/W3-redundancy measurement is preserved verbatim.
- AP-146 and #320 amended the same way — their three-edge list was written from
  the graphical host and silently assumed both hosts shared it. The no-window
  host had two of three; it now has all three.
- AD-64 filed: the reachability decision is now expressed once per host. The
  value is single-sourced; the gate set is not.
- #324 filed: unifying the two session controllers is the genuinely correct fix
  and is campaign-sized (presentation recovery, hydration, the equipped-child
  renderer, and the remote/projectile routing arms only one host has). Not
  attempted here, per the fix brief.

Gates. Release build 0 errors. Complete suite 11,141 passed / 4 skipped /
0 failed, against the 11,134 / 4 / 0 baseline at 23aa62f2 — net +7, exactly the
7 tests added. Eight sabotages verified, each red on at least one discriminating
test and green when reverted: remote commit removed (2 Runtime + the end-to-end
Headless test); local ordinary commit removed; local NotApplicable-force commit
removed; force commit made unconditional; residence gate removed; missile gate
removed; Rejected gate removed; preserve branch broken (red on both hosts).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Erik 2026-08-05 22:36:31 +02:00
parent 23aa62f292
commit ff100cf33f
8 changed files with 746 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

7
docs/architecture/retail-divergence-register.md
View file

File diff suppressed because one or more lines are too long