fix #483: restore far terrain and close OVERHAUL with #484 deferred

docs/research/2026-09-01-overhaul/2026-09-06-campaign-closeout.md

@@ -0,0 +1,89 @@
+# Campaign OVERHAUL v2 — owner-directed closeout, 2026-09-06
+
+**CLOSED — OWNER ACCEPTED WITH CARRIED ISSUES.** After the connected testing
+and the #483 distance repair, the owner directed: "Ok defer that as an issue.
+I want to close out the campain, commit and merge this worktree with main and
+then push main to remote gita". This is the final merge-gate disposition:
+accept the campaign with #484 deferred, not a claim that every remaining G4
+checklist row was individually run or passed. Unreported rows remain
+unreported; known failed evidence remains FAIL.
+
+## Shipped scope
+
+S1-S5 provide the geometry and cross-cell membership products, single retail
+frame walk, depth/alpha ordering, and consumer/material cutovers. Gate-found
+repairs cover outdoor buildings, portal reveal, dungeon lighting, held items,
+cathedral exterior collision and finally missing Far-terrain registration.
+FW's predecessor ledger remains historical; its old clip model is not the
+shipping architecture. M4 as a whole is not declared complete.
+
+The last bounded fix, #483, carries the existing height-table terrain bounds
+through Near/Far construction and all Far conversions, registers terrain
+independently of buildings, retains it during demotion and removes it during
+full retirement. It changes no radius, fog, LOD or retail walk algorithm.
+The owner used the repaired build; the subsequently reported mountain flicker
+is separately retained as #484, with no confirmed cause or speculative fix.
+
+## Final candidate verification
+
+- Release App/test build: PASS, zero warnings/errors.
+- Narrow production-path regression suite: 106 passed / 0 failed / 0 skipped,
+  using the explicit recipe10 package. The initial 105/1 run selected the
+  obsolete default recipe8 package; both reports are retained under
+  `artifacts/483-far-terrain/`. No source change masked that mismatch.
+- Final no-build App hermetic-filter regression: **7,161 passed / 0 failed /
+  0 skipped**. Report:
+  `artifacts/483-closeout-app-hermetic/483-closeout-app-hermetic.trx`.
+  Filter excludes Lane InstalledDat, Linux, Manual, Timing, Live and
+  PreparedPackage. This is an App result, not a new complete-solution run.
+- Post-owner independent retail review: **PASS**, no blocking finding.
+  Independent architecture/lifetime review: **PASS**, no blocking finding.
+  Both reviewed the six production/two test files; no further fix round.
+  Review scope and named/paired retail evidence are in
+  `483-far-terrain-registration-contract.md`; neither review certifies #484.
+- App SHA256:
+  `869BA14CB2C610E006C80BDFFCBD75629339A3486B8B98D7708B23050700FF34`.
+  Core SHA256:
+  `B0E676482A45ED77C2AB050BBBC988139D1BF6D99308C42ABB0CFF440D39A0BC`.
+  Package `acdream-v10-s5-c5.pak` SHA256:
+  `4E0FA6C589ECE12E13099A56385077964961EC52F888BC9161C6F5D34829B59B`.
+- Owner launches: `logs/owner-g4-20260906-r2/` (PID17856) and
+  `logs/owner-g4-20260906-r3/` (PID24476), same repaired build.
+  No new automated visual/PNG PASS is invented for this closeout.
+
+## Preserved evidence and carried work
+
+The previous integrated source `eea83ac884d0b0fd74d86a17a16d7d5d93f64d94`
+passed Release, 17,094 complete-solution hermetic tests, repeated unchanged
+geometry/membership and Signature8 checks, lifecycle/reconnect and R6.
+Those results are historical integrated evidence, not a full-suite run on
+the final #483 delta.
+
+Canonical InstalledDat remains **376 pass / 8 known failures / 1 known skip**.
+The five final visual controls retain 51 inspected PNGs and their narrow,
+provisional verdicts. Four controls exited0; the wall control exited
+**0xC0000374 (process FAIL)** despite graceful network logout. No native stack
+establishes its cause or equates it with the AMD notice. See
+`c1c-integrated-verification.md`,
+`s5-final-lifecycle-stability-verification.md` and
+`s5-final-visual-controls.md` for exact limitations, paths and sealed manifests.
+
+Carried without claiming repair: **#484** distant silhouette flicker (explicit
+owner deferral); **#422** native shutdown symptom; **#477** AMD crash notice /
+unconfirmed initiating cause; the known InstalledDat failures/skip; the known
+retail cathedral exterior-slab leak; owner-deferred FPS/C2; and the existing
+divergence-register residuals including **AP-149** (Far static/building
+closure). #483 adds no intentional retail deviation and retires none of
+those residuals. Remaining unreported G4 matrix rows are not fabricated PASS
+rows; the owner's closeout instruction supersedes waiting for that session.
+
+## Integration authorization
+
+Commit the owned candidate and documentation, merge the campaign branch into
+main, then push **origin/main** to Gitea
+`https://git.snakedesert.se/erik/acdream.git`. Do not push the GitHub remote.
+Preserve main's unrelated untracked files and WorldBuilder submodule checkout;
+no reset, clean, stash, history rewrite or force push. Actual Git completion
+is reported after the operations succeed, not asserted by this pre-commit note.
+Retail is untouched. Its presence does not block a build by the owner's
+explicit override; a running ACDream still must not have its binaries replaced.

docs/research/2026-09-01-overhaul/483-far-terrain-registration-contract.md

@@ -5,6 +5,50 @@ Owner order: **implement, owner tests, then review**. This overrides the normal
 pre-owner review order for this bounded fix; no claim of review closure follows
 from implementation or successful compilation.
 
+**Latest owner safety correction (2026-09-06):** retail being open does NOT
+block building/testing/launching ACDream. The owner explicitly directed
+"Just ignore if retail is up or not, does not affect our building. Now build
+and launch." Leave retail untouched. Only a running ACDream blocks overwriting
+its build or launching a duplicate. This supersedes all older both-client
+absence requirements below; those describe the earlier handoff, not current
+permission. The requested launch is manual, with review after the owner's test.
+
+## Final disposition — CLOSED, 2026-09-06
+
+The owner directed campaign closeout and integration, with subsequent mountain
+flicker deferred as #484. After the owner session, independent retail and
+architecture/lifetime reviews both **PASS**, no blocking findings and no new
+fix round. Retail review verified unchanged get_land_limits bounds, independent
+Far terrain admission, all three metadata-preserving conversion paths and
+unchanged walk/LOD policy against named retail and the paired evidence below.
+Lifetime review verified the sole owner, retained publication receipt,
+demotion/full-removal split, late-clear non-resurrection and production-route
+regressions. Neither review certifies #484 or closes AP-149.
+
+Final no-build App hermetic-filter regression: **7,161/0/0**; report
+`artifacts/483-closeout-app-hermetic/483-closeout-app-hermetic.trx`.
+The Release build and 106 narrow passes below cover the same product candidate.
+See `2026-09-06-campaign-closeout.md` for the owner's merge-gate disposition
+and retained failures. No new individual visual PASS is asserted.
+
+## Historical execution — owner client launched
+
+After the explicit owner override, Release App/test build passed with zero
+warnings/errors. Initial narrow run: 105 pass / 1 fail because it selected the
+old default acdream.pak (recipe 8, expected 10). No product change was made for
+that environment failure. Rerun with ACDREAM_PAK_PATH explicitly selecting the
+same acdream-v10-s5-c5.pak as the client: **106 pass / 0 fail / 0 skip**.
+Both TRX reports remain under artifacts/483-far-terrain/.
+
+The fixed uncommitted candidate launched once: PID17856 at
+2026-09-06T07:37:36.5010165Z, enteredWorld07:37:40.5272305Z; logs are
+logs/owner-g4-20260906-r2/. App SHA256
+869BA14CB2C610E006C80BDFFCBD75629339A3486B8B98D7708B23050700FF34;
+Core B0E676482A45ED77C2AB050BBBC988139D1BF6D99308C42ABB0CFF440D39A0BC.
+Retail untouched, no automated route or automatic close. Owner testing and
+post-owner review were pending at this launch checkpoint; the final disposition
+above supersedes it and the earlier unbuilt handoff statements below.
+
 ## Confirmed cause
 
 Owner-g4-20260906-r1 logged NearRadius4/FarRadius25, unchanged maximum preference.
@@ -29,7 +73,8 @@ preserves this known gap instead of asserting a usable far draw product.
    both tiers. Keep exactly one landscape owner and existing walk ordering/LOD.
    Near publication supplies buildings separately; far entries have no buildings.
 3. Forward the terrain metadata through both PublishAsFar conversions and all
-   normal load/promotion routes. No extra DAT or collision/static read for Far.
+   normal load/promotion routes, including LandblockStreamer's unconditional
+   LoadFar near-payload stripping. No extra DAT or collision/static read for Far.
    Respect the retained publication receipt and existing retry/budget semantics.
 4. Near-to-far retirement clears building references but retains the terrain
    entry/bounds. Full terrain removal removes its landscape entry. Repeated
@@ -50,7 +95,8 @@ do not claim it retired because distant terrain now draws.
 
 - Far build keeps EnvCells null and performs only its existing heightmap read,
   but carries exact non-flat authored vertical bounds.
-- Production publisher + real landscape/walk product produces a nonempty
+- Actual LoadFar worker -> DrainCompletions -> production publisher + real
+  landscape/walk product preserves the bounds and produces a nonempty
   terrain turn for an in-view far-only block beyond near radius4, with no
   building/object population. The old missing-registration implementation must
   fail this assertion; a renderer-slot count alone is insufficient.
@@ -64,11 +110,17 @@ do not claim it retired because distant terrain now draws.
 ## Ownership, safety and handoff
 
 One OpenAI implementer owns only LandblockBuild.cs, LandblockBuildFactory.cs,
-LandblockRenderPublisher.cs, LandblockPresentationPipeline.cs,
+LandblockRenderPublisher.cs, LandblockPresentationPipeline.cs, LandblockStreamer.cs,
 Rendering/Walk/WalkLandscapeAssembler.cs and their directly affected tests.
 Small related declaration/comment edits may be proposed, not broadened silently.
 Lead owns issue/ledger/architecture/register and this contract. No shader/PAK edit.
 
+Scope amendment: the implementer found that the real LoadFar worker reconstructs
+LandblockBuild while stripping near payload, losing separately carried terrain
+bounds. The lead authorized only that metadata forwarding in the sixth production
+file and an actual worker-to-walk regression. This implements requirement 3;
+no scheduling, loading policy or new retail deviation is authorized.
+
 **Retail PID25988 is owner-held: DO NOT TOUCH.** No close, input, debugger attach,
 build, tests or launch while either client is running. Code edits may proceed.
 Do not invoke Claude or start review agents. Return the uncommitted scoped diff
@@ -76,3 +128,47 @@ and exact unrun test commands; implementation is NOT yet built or tested.
 After the owner closes retail, lead does the green Release build and narrow
 regression run, launches one manual client, and leaves review until after the
 owner's test. No main merge and no G4 acceptance are implied.
+
+## Lead static retail check — 2026-09-06
+
+Named pseudo-C identifies the exact helper as
+`CLandBlock::get_land_limits0052F1D0`, not calc_lighting. The existing local
+Ghidra endpoint8081 is offline; no new tool or retail process was launched.
+Instead the lead read the paired file `C:\Turbine\Asheron's Call\acclient.exe`
+as data: timestamp2013-09-06, PDB GUID9e847e2f-777c-4bd9-886c-22256bb87f32,
+age1, SHA256006ffeadc5d679c871497112a5bd1f87714d0e273e2166bae5052dde369297b1.
+The separately named acclient_codex_oracle.exe is2015 and was excluded.
+
+Direct PE byte inspection agrees with the named source:
+
+- 0050634F onward loads mid_width from this+4, squares it, and walks the
+  complete block_draw_list backwards. Non-null and block+FC in-view tests
+  precede the virtual DrawBlock call. There is no building-presence gate.
+- 0052F281 onward loads the maximum byte's height-table value, adds
+  float[007C8F1C]=200.0 (bytes00004843) and stores this+E0; the minimum-byte
+  value subtracts float[007928B0]=1.0 (bytes0000803F), stored at this+E4.
+  The existing compute helper's formula is therefore preserved, not replaced
+  by the GPU AABB's different margins.
+
+This is factual input verification for the repair, not the postponed code
+review or a test PASS. Retail remained open and untouched throughout.
+
+## Candidate handoff — 2026-09-06
+
+The implementer returned the six production files plus
+LandblockBuildFactoryTests.cs and LandblockConcretePresentationPipelineTests.cs
+uncommitted. The candidate and regression tests are written; ownership is back
+with the lead. No implementation agent remains active on this repair.
+Only whitespace validation (`git diff --check`) ran. The candidate is
+**UNBUILT, UNTESTED and UNREVIEWED**; the previous eea83ac88 binary is unchanged.
+Retail PID25988 remains owner-held, ACDream is absent, ACE UDP9000/PID2296 is up.
+No build, test, client launch, debugger attach or client close was performed.
+
+When both clients are absent, build the App test project in Release (which
+builds its App dependencies) and run the narrow lane below before owner launch.
+The test run must not be reported as a visual gate or the deferred review.
+
+```powershell
+dotnet build tests/AcDream.App.Tests/AcDream.App.Tests.csproj -c Release --no-restore -m:1 -p:UseSharedCompilation=false
+dotnet test tests/AcDream.App.Tests/AcDream.App.Tests.csproj -c Release --no-build --no-restore --filter "FullyQualifiedName~LandblockBuildFactoryTests|FullyQualifiedName~LandblockConcretePresentationPipelineTests|FullyQualifiedName~LandblockPresentationPipelineTests|FullyQualifiedName~LandblockRenderPublisherTests|FullyQualifiedName~WalkLandscapeAssemblerTests|FullyQualifiedName~RetailFrameWalkTests|FullyQualifiedName~WalkLandCellOrderTests" --logger "trx;LogFileName=483-terrain-regression.trx" --results-directory artifacts/483-far-terrain
+```

docs/research/2026-09-01-overhaul/484-distant-terrain-flicker-investigation.md

@@ -0,0 +1,53 @@
+# #484 — distant mountain silhouette flicker
+
+**OPEN — OWNER-DEFERRED, 2026-09-06.** The owner explicitly requested filing
+this issue and closing/merging the campaign. Investigation stops here; exact
+pose and consecutive-frame reproduction are future issue work, not a pending
+campaign gate. Cause remains unconfirmed. No production fix or PASS claimed.
+
+2026-09-06; owner report on the uncommitted #483 candidate based on e0ba3e7f2.
+Owner says flicker continues with character/camera still; turning reveals more
+mountain at affected spots. User image:
+`C:/Users/erikn/AppData/Local/Temp/codex-clipboard-d6976e08-527f-47b9-8f00-79429065ea13.png`.
+A still does not establish the temporal mechanism. No fix or visual PASS yet.
+
+## Read-only findings
+
+- #483's launched App hash remains
+  869BA14CB2C610E006C80BDFFCBD75629339A3486B8B98D7708B23050700FF34;
+  PID17856, logs/owner-g4-20260906-r2/. Retail was untouched.
+- `TerrainModernRenderer.AppendCellIndexRuns` covers all 384 mesh indices for
+  side1/cell0; other LOD cell grids expand disjoint complete row ranges. This
+  is draw granularity, not a changing geometry LOD. No obvious stale terrain
+  list accumulation found in driver/renderer clear paths.
+- Terrain indirect commands use a per-frame upload-ring allocation. Fog only
+  blends colour; the terrain shader does not discard fogged fragments.
+- CPU walk uses CY/near and view-edge planes, while the world camera projection
+  has .1 m near / 5000 m far. A radius25 square can reach roughly 6.8 km
+  diagonally. This permits hardware clipping of CPU-admitted blocks, but that
+  is NOT sufficient evidence of this fault or justification for extending far.
+- Named retail also uses finite far clipping: zfar0081EC88 defaults to4000;
+  PrimD3DRender::SetFOVInternal0059AB40 builds the projection from znear/zfar;
+  SetDefaultD3DStates sets clipping state0x88 to1 at005A3B30. Terrain submits
+  XYZ through DrawPrimitiveUP, not a projected-Z clamp/far bypass. Only sky
+  temporarily multiplies zfar by4, then restores it.
+- Lead verified those three load-bearing facts from the matching 2013 PE as
+  file data (no attach): SHA256
+  006ffeadc5d679c871497112a5bd1f87714d0e273e2166bae5052dde369297b1.
+  VA0081EC88 bytes00007A45 encode4000; 0059AB40 body loads globals81EC88/81EC84
+  before the perspective call; 005A3B32 has push1/push88 before SetRenderState.
+  The far-clipping hypothesis remains unproven; no projection change was made.
+
+## Next discriminating observation
+
+Owner was asked to leave the client at an affected viewpoint and provide /loc.
+The one window observation instead caught arrival in Neftet, not the reported
+mountain silhouette. No input was sent to either client. Need that pose before
+capturing consecutive stable frames and camera/terrain admission/clip facts.
+Distinguish disappearing block/triangle coverage from changing shaded pixels;
+then write the bounded fix contract, add a discriminating regression, verify
+and gate the actual repair. Do not turn an unproven far-plane theory into a fix.
+No new code, build, test, client close/launch, review, or G4 acceptance this turn.
+
+Current safety: retail remaining open does not block build/test/launch (explicit
+owner override). Leave retail untouched; do not overwrite a running ACDream.

docs/research/2026-09-01-overhaul/g4-owner-checklist.md

@@ -1,11 +1,14 @@
-# G4 owner checklist — final candidate, 2026-09-05
+# G4 owner checklist — retained matrix; closeout 2026-09-06
 
-Status: **NOT RUN / NOT ACCEPTED**. This is the remaining owner session,
-not another implementation or review round. Binding scope is the campaign
-plan's S5 gate and section10 scene matrix. The automatic checks do not replace
-the owner's judgment. Do not merge or mark the campaign closed from this card.
+Status: **OWNER-DIRECTED CLOSEOUT ACCEPTED WITH CARRIED ISSUES.** On 2026-09-06
+the owner directed deferral of #484, campaign closeout, commit, merge to main
+and push to Gitea. This supersedes waiting for the remaining matrix session;
+it does not assert individual PASS results for unreported rows or erase FAILs.
+See `2026-09-06-campaign-closeout.md` for the merge-gate disposition and exact
+final #483 candidate. The earlier checklist and evidence below are retained
+history, not a new launch request.
 
-## Candidate and launch
+## Historical candidate and launch (2026-09-05)
 
 Use the existing Release binaries built at
 `eea83ac884d0b0fd74d86a17a16d7d5d93f64d94`; documentation checkpoint
@@ -17,7 +20,9 @@ Exact App/Core/package hashes and green build evidence are in
 Before launch, recheck binary/package hashes, both client processes and ACE
 UDP9000; record the launch commit, flags, profile, process and log directory.
 Keep the saved High/MSAA4, near4/far25 profile for the normal candidate run.
-No rebuild beside either client; no code change after the accepted binary.
+The original no-rebuild-beside-either-client rule was superseded on 2026-09-06:
+retail stays untouched but does not block building. Do not overwrite running
+ACDream binaries or launch duplicates.
 Owner session has no automatic route or automatic close. A retail comparator
 is needed for the two-client row and known cathedral slab comparison; launch
 each client in its own command. Close owned clients gracefully when agreed.
@@ -42,7 +47,8 @@ each client in its own command. Close owned clients gracefully when agreed.
 ## Owner session
 
 For every row retain PASS/FAIL, actual `/loc`, symptom if any, and PNG/log
-paths. All rows below remain **PENDING** on this final binary. The starting
+paths. Rows without a reported individual verdict remain **UNREPORTED**, not
+new PASS results; waiting for them was superseded by owner closeout. The starting
 locations are navigation aids, not a substitute for movement or camera work.
 
 | Row | Check | Start / navigation |
@@ -60,13 +66,14 @@ network logout or a client-authored code0 does not override the OS exit.
 Do not retry just to replace a failed record. The lead reads local crash
 reports/Windows evidence without changing registry, driver or dump settings.
 
-## Acceptance and remaining closeout
+## Historical acceptance requirements and final disposition
 
 Owner PASS must identify the tested rows and any explicit carried exception.
 The historical acceptance of#422 in Campaign VM is not automatically a new
 OVERHAUL acceptance. A good visual session does not prove the original AMD
 reset's initiating cause, remove the installed-data failures, or convert
-the recorded shutdown failure to PASS. Their disposition remains explicit.
+the recorded shutdown failure to PASS. Their failures remain explicit in the
+closeout report.
 
 FPS/C2 performance work is owner-deferred. Retail's known cathedral leak is
 unchanged; its proposed post-campaign improvement is not part of this gate.
@@ -74,7 +81,7 @@ unchanged; its proposed post-campaign improvement is not part of this gate.
 After the owner results, reconcile the shipping documentation before closure:
 architecture, WorldBuilder inventory, divergence register, launch options,
 FW ledger (superseded, not deleted), roadmap/milestones and campaign memory.
-Keep the existing failure records. Only when G4 and that closeout actually
-agree with the code can the campaign be marked CLOSED — USER ACCEPTED and
-the authorized merge considered. Do not rebuild the accepted product for
-documentation-only changes.
+Keep the existing failure records. The owner's 2026-09-06 closeout instruction
+now authorizes integration with carried issues, rather than requiring another
+matrix session. Documentation records that disposition without inventing
+per-row passes. Do not rebuild the accepted product for documentation-only changes.