standalone: shape-world look round 2 -- shards gone, calmer layered relief, still 0.72 s / 1.6 s cold

- Removed the thin automatic room-support slabs that left floating caps (in-view detached thin
  sheets 9 -> 0; the audit now also fails on isolated small components).
- Bounded 16-voxel floor/wall blend (18 rejected: it created a sub-voxel sheet), 3-voxel
  strata/cave blend, C1 tunnel radius profile (0 radius/slope steps at joins).
- Surface relief: broad octave with amplitude modulation instead of uniform bumps.
- Cold, no cache: 64 m full detail 0.924 s (8 workers) / 0.715 s (16); whole view 2.016 / 1.605 s.
- 1-worker and 6-worker hashes match; 5,642/5,642 passage samples open.
Not solved: regular bands in the narrow passage shafts, a few slivers on tile boundaries, faint
wall/floor plates; close-view mean surface distance to the owner field got worse (see REPORT.md).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ

Standalone/Research/OwnerShapeWorld/round2/REPORT.md

@@ -1,57 +1,251 @@
-# OwnerShapeWorld round 2 — current evidence snapshot
+# Shape world round 2 — final measurements and visual status
 
-This report reflects the last source build in this folder. It is a **partial snapshot**, not a final sign-off: Windows stopped starting shell processes with exit status `-1073741502` (`0xC0000142`) during final capture/benchmark work. The latest source compiled successfully, and the audits below completed. Final captures, cold timings, owner comparison, and determinism have not yet been rerun against that exact build.
+## Status
 
-## Source changes
+The exact final source builds, meets both cold-generation targets, preserves all
+5,642 passage-centerline samples, and is deterministic across repeat runs and
+1/6 workers. The final mesh audit finds no sub-voxel sheets in the view and no
+isolated small components. Tunnel radius and slope values agree at shared
+joins.
 
-- Tapered tunnel and passage radii now match at shared control points after sub-voxel endpoint quantization. The measured worst mismatch fell from `0.0077` vox in the earlier `passage_fixed_after_extensions` audit to `0.000001` vox in the latest audit.
-- Native tunnel floor endpoints are checked independently; the current worst mismatch is `0.000000` vox. Smooth-union radius is zero, so it adds no tunnel inflation.
-- Authored radius samples are low-pass smoothed across nine control points before tube construction. This is the most recent source change; its final image captures are still pending.
-- Cave floor displacement uses a stronger relief multiplier (`2.40`), while cave wall relief uses `0.55`. Strata and solid-overlay relief remain `1.35`.
-- The packed interval and surface-band path remains the fast shape-world implementation.
+The visual target is **not fully met**. The floor-to-wall transition is softer,
+and the large floating shards from the earlier capture are gone, but the fixed
+chamber still has a few plate-like lips/slivers. The passage close-up still
+shows repeated annular bands. The final pictures are visibly different from
+the owner's field, especially in relief hierarchy and tunnel profile. The
+flat-patch and grid-stair tests also miss some of the visible chamber detail;
+their pass results are not a visual sign-off.
 
-## Latest completed audits
+## Final source and build
 
-Both latest audits used 8 workers, a 64 m full-detail clip, cache-free shape construction, and BelowNormal process priority on affinity `0xFF00`.
+All source/build changes are under `Standalone/`:
 
-| View | Shape stage | Route samples | Detail ray stddev | Axis-locked normals | Flat 16-voxel patch share | Result |
-|---|---:|---:|---:|---:|---:|---|
-| Passage floor | 725.807 ms | 5,642 / 5,642 open | 1.1335 vox | 17.8% | 0.0% | Pass |
-| Chamber | 947.796 ms | 5,642 / 5,642 open | 0.4545 vox | 2.2% | 0.1% | Pass |
+- `Source/StandaloneShapeWorld.h` evaluates hierarchical relief with a broad
+  32-voxel octave, a modulated 16-voxel octave, and a 6-voxel octave at 0.18
+  amplitude.
+  The coherent detail remains bounded to an 8-voxel surface band. Tunnel and
+  passage radii use a nine-control-point low-pass profile plus cubic Hermite
+  interpolation, matching radius and slope at shared control points. Floor
+  profiles are C1 blended; the floor-clip blend radius is 16 voxels. Thin
+  positive support overlays were removed.
+- `Source/StandaloneViewer.cpp` includes the surface, grid-stair, route, owner
+  comparison, and mesh-component/sheet audits. The temporary skirt-disable
+  probe switch was removed.
+- `Source/StreamingWorld.h` and `Build.ps1` allow 16 workers; the build uses
+  `/MP16`.
 
-The chamber's largest flat patch was 56 samples. The floor probe tagged the directly hit cave surface as `floor_surface=yes`; its hit detail was `-0.4081` vox. All final audit records are in `profile9_floor_audit.stdout.log` and `profile9_chamber_audit.stdout.log`.
+The final exact-source build is `final_build_exact.log` and produced
+`bin/VoxelForgeStandalone.exe`. It succeeded. The compiler still emits
+the existing six C4838 double-to-float warnings in `VoxelStrateManager.cpp`;
+the setup script also prints that `vswhere.exe` is not recognized, but that
+diagnostic did not prevent the build.
 
-The latest tunnel continuity audit reports 280 shared joins in the chamber view, maximum endpoint radius mismatch `0.000001` vox, maximum floor-endpoint mismatch `0.000000` vox, and zero smooth-union growth. The older fixed-radius representation's maximum join step was `2.9156` vox; the final radius profile reports `1.7487` vox between adjacent samples in that view. These are geometric checks, not proof that every visible radius swell has disappeared.
+Timing, determinism, comparison, and capture processes used Normal priority
+and affinity `0xFFFF`, one heavy process at a time. Cold timing, determinism,
+and capture runs used `--force-generate`; timing runs report zero mesh-cache
+hits. The 8-worker numbers are for comparison with earlier rounds; the
+16-worker numbers use the full machine. Times below are shape-build stage plus
+uncached mesh-generation wall time.
 
-## Visual evidence currently available
+| Cold job | Workers | Local shapes | Signs/intervals | Surface band | Far parents | Shape stage | Mesh wall | Total |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| 64 m full-detail area, 960 tiles | 8 | 58.3 ms | 136.0 ms | 558.8 ms | — | 754.5 ms | 170 ms | **0.924 s** |
+| 64 m full-detail area, 960 tiles | 16 | 58.1 ms | 124.9 ms | 389.2 ms | — | 573.5 ms | 141 ms | **0.715 s** |
+| Whole view, 6,498 tiles through LOD6 | 8 | 58.5 ms | 133.3 ms | 538.7 ms | 188.7 ms | 923.7 ms | 1.092 s | **2.016 s** |
+| Whole view, 6,498 tiles through LOD6 | 16 | 57.3 ms | 126.0 ms | 407.8 ms | 190.0 ms | 785.8 ms | 819 ms | **1.605 s** |
 
-The owner reference images are under `../final/`. The four round-2 fixed views captured before the last radius-filter change are:
+Both targets are met: <=1 s for the full-detail area and <=5 s for the whole
+view. Peak process memory reported by `RunMachine.ps1` was 0.21/0.17 GB for the
+8/16-worker full-detail runs and 0.32/0.37 GB for the 8/16-worker whole-view
+runs. Fresh timing logs: `final_recheck_full_detail_w8.stdout.log`,
+`final_recheck_full_detail_w16.stdout.log`, `final_recheck_whole_view_w8.stdout.log`,
+and `final_recheck_whole_view_w16.stdout.log`. Total time is shape-stage time
+plus cold stream wall time; each run reported zero cache hits.
 
-- `chamber_after_seamfix.png`
-- `tunnel_after_seamfix.png`
-- `passage_after_seamfix.png`
-- `far_after_seamfix.png`
+## Defect checks
 
-The closest available passage-floor and bend images are `softprofile_floor_closeup.png` and `tunnel_bend_closeup_after_seamfix.png`. The fixed-view images are from the endpoint-calibration build; the floor close-up is from the five-tap radius filter. The last build changed only the filter to nine taps, so these are diagnostic references, not exact after-build captures.
+| Check | Before | Final source | Result and limitation |
+| --- | --- | --- | --- |
+| In-view disconnected/sheet candidates | 9 sheets, 149 triangles under 1 voxel; failed | 0 sheets, 0 sheet triangles, 0 isolated small components; 5 small components touch tile boundaries | Pass. The audit now exits nonzero for either a thin sheet or an isolated small component. Boundary components remain separately reported because per-tile meshes do not establish whether they connect across neighbors. |
+| Tunnel radius joins | Earlier extension/profile trial reached 0.0077-voxel endpoint mismatch; legacy control-radius step was 2.9156 voxels | 0.000000-voxel endpoint delta and 0.00000000 slope delta at 280 shared joins; smoothed adjacent-control step 1.7487 voxels | Numeric join check passes. The annular bands in the close-up show that C1 continuity alone does not remove the visible profile rhythm. |
+| Tunnel grid-stair index | Correct baseline tunnel scan without `--round-tunnels`: relative join step 0.434; it fails at the final 0.400 threshold | 0.028 at threshold 0.400 | Pass on the fixed tunnel view. |
+| Chamber grid-stair index | 0.041; below the 0.400 threshold | 0.003 at threshold 0.400 | Numeric improvement, but the baseline already passed. The audit does not catch every plate visible in the chamber image. |
+| Passage grid-stair index | 0.276; below the 0.400 threshold | 0.205 at threshold 0.400 | Numeric improvement, but the baseline already passed. |
+| Detail/flat-patch audit, chamber | Detail standard deviation 1.0056 voxels; flat-patch share 0.1% in the baseline audit | Detail standard deviation 0.6683 voxels; flat-patch share 5.0%, largest patch 433 samples | The audit reports PASS, but these results do not establish that the visible join plates were removed. |
+| Passage routes | 5,642/5,642 open | 5,642/5,642 open in chamber, tunnel, and passage audits | Pass. This is a centerline sign test, not a character-capsule collision test. |
 
-Visual review: the chamber now has detail across the cave wall and floor, but its broad relief still looks lumpier than the owner's chamber. The passage tube is connected and routeable, but its long profile still reads as gently repeated radius bands in the close-up. Those bands are continuous by the endpoint audit, though visual closeness to the owner is not yet satisfactory enough to call complete. The fixed passage view also includes an authored landing plane; it should not be counted as a missing cave-floor detail patch.
+The final full-view sheet check is `final_recheck_whole_view_shard_audit_v2.stdout.log`.
+The selected 16-voxel floor blend came after an 18-voxel probe that introduced
+one boundary sheet (20 triangles, estimated thickness 0.497 voxel) and failed
+with exit 13. That probe is recorded in
+`final18_whole_view_shard_audit_w16.stdout.log`; radius 18 was reverted.
 
-## Missing final evidence
+The grid-stair detector does not satisfy the original request on every view:
+it fails on the pre-fix tunnel view and passes after, but the pre-fix chamber
+and passage metrics already passed its threshold. The images still show why
+that check needs better targeting.
 
-- Cold full-detail area and whole-view timings on the latest nine-tap build, with per-stage timings and cache counts.
-- Final lattice-sign difference and surface distance to the owner field.
-- Repeated-run and 1-worker versus 6-worker determinism hashes on the latest build.
-- New fixed-view and close-up captures from the latest nine-tap build, followed by a final side-by-side visual review.
-- A same-camera before/after numeric flat-patch and grid-stair comparison. The baseline captures show the defects, but the current audit instrumentation was added after those captures.
+## Owner-field comparison
 
-Earlier successful timings and hashes are in `benchmark_full_detail_final.stdout.log`, `benchmark_full_clipmap_final.stdout.log`, and `determinism_w*_final.stdout.log`; they predate the final radius-profile filter and must not be treated as current measurements.
+Before and after use the same scan settings: 8 workers, 960 full-resolution
+tiles for chamber/tunnel/passage, 4-sample mesh-vertex stride, and the same
+fixed cameras. The far comparison uses 4,770 LOD6 tiles. “Changed signs” is the
+share of lattice samples where the shape field differs from the owner field;
+surface distances are sampled symmetric nearest-vertex distances, not a
+continuous Hausdorff distance. Unmatched samples are listed separately.
 
-## Failed attempts to disclose
+| View | Lattice samples | Changed signs before → after | Mean surface distance before → after | Max distance before → after | Unmatched samples before → after |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Chamber | 35,103,600 | 2,098,617 (5.978%) → 2,069,214 (5.895%) | 1.0325 m → 1.0826 m | 6.4677 m → 6.4840 m | 9,393 → 8,832 |
+| Tunnel | 35,103,600 | 2,098,617 (5.978%) → 2,069,214 (5.895%) | 1.0325 m → 1.0826 m | 6.4677 m → 6.4840 m | 9,393 → 8,832 |
+| Passage | 37,240,000 | 1,232,883 (3.311%) → 1,230,337 (3.304%) | 0.9814 m → 1.0292 m | 6.3422 m → 6.3750 m | 12,602 → 12,520 |
+| Far view, LOD6 | 9,470,835 | 856,305 (9.041%) → 856,305 (9.041%) | 2.1667 m → 2.1588 m | 5.3070 m → 5.3070 m | 783,976 → 783,967 |
 
-- `build_hardjoin.stdout.log`: failed because it referenced `standalone.rsp` from the wrong working directory. Retrying with `Build\standalone.rsp` succeeded.
-- `final_passage_floor_audit.stdout.log`: failed its detail probe because the camera was placed about 11 cm above the landing. This was a bad probe; the raised natural-floor probe passed.
-- Several early image captures failed because their relative output path was resolved from the repository root. Captures were subsequently rerun from `Standalone\` and succeeded.
-- An owner-floor capture requested the full clipmap and was stopped after its full-detail phase exceeded 19 seconds. The later full-detail-only owner probe completed, but its staircase-heavy image is not a valid speed comparison.
-- `passage_worldchain.stdout.log` and `passage_compare_worldchain.stdout.log`: an intermediate world-chain extraction left 412 of 5,642 route samples solid; that extraction was discarded, and the current route audit passes all samples.
+The changed-sign share improves slightly in the close views; surface distance
+gets worse on average there. The far field changes almost not at all. Taken
+together, these numbers do not show a material increase in owner similarity.
+Final logs are `final16_compare_baseline_{chamber,tunnel,passage}_w8`,
+`fine18_compare_after_{chamber,tunnel,passage}_w8`,
+`final16_compare_baseline_far_lod6_w8`, and
+`fine18_compare_after_far_lod6_w8`.
+
+## Determinism
+
+Two cold 6-worker runs and one cold 1-worker run at the passage camera produced
+the same hashes, with zero cache hits and 5,642/5,642 route samples open:
+
+| Field | All three runs |
+| --- | --- |
+| Full detail | `0x0c985b329e1aaf79` |
+| LOD0 | `0x778716fa053056c1` |
+| LOD1 | `0x49e560dbbca29f45` |
+
+Fresh matching runs are `final_recheck_determinism_w6_a.stdout.log`,
+`final_recheck_determinism_w6_b.stdout.log`, and
+`final_recheck_determinism_w1.stdout.log`. The exact final-source rebuild is
+recorded in `final_build_exact.log`.
+
+## Captures and visual review
+
+The final after captures were made from the exact 16-voxel floor-blend and
+0.18-fine-amplitude source with fresh
+generation. The `before_*.png` captures use the saved pre-fix executable; owner
+references are the fixed captures under `../final/`.
+
+| View | Camera | Before | Final after | Owner |
+| --- | --- | --- | --- | --- |
+| Chamber | `(-11.75,-123.25,-19.85)`, yaw 0° | `before_chamber.png` | `final_after_chamber.png` | `../final/chamber_fixed_owner.png` |
+| Tunnel | `(-11.75,-123.25,-19.85)`, yaw 90° | `before_tunnel.png` | `final_after_tunnel.png` | `../final/tunnel_fixed_owner.png` |
+| Passage | `(-39.19,6.27,-28.51)`, yaw 63° | `before_passage.png` | `final_after_passage.png` | `../final/passage_fixed_owner.png` |
+| Far view | `(-11.61244,-118.2212,-19.8452)`, yaw 0° | `before_far.png` | `final_after_far.png` | `../final/far_owner.png` |
+| Passage-floor close-up | `(-39.19,6.27,-25.51)`, yaw 63°, pitch -25° | `before_passage_floor_closeup.png` | `final_after_passage_floor_closeup.png` | No separate owner close-up was supplied |
+| Tunnel bend | `(-198.97,-99.57,-45.58)`, yaw 33.59°, pitch 23.17° | `before_tunnel_bend.png` | `final_after_tunnel_bend.png` | No separate owner close-up was supplied |
+
+I reviewed the four fixed after/owner pairs and both close-ups. The chamber
+transition is softer than before, but small plate-like lips and dark slivers
+remain. The tunnel still has a visible wedge and its forms repeat more
+regularly than the owner's. The passage view has broad horizontal shelves and
+large dark cutouts; its landing alone is present in the owner image, but the
+surrounding shape differs. The close-up shows conspicuous concentric bands.
+The far view reads as cloudy relief, while the owner image has larger,
+connected cave forms with finer detail. Overall, this is a fast coherent
+alternate field, but the final images are **not visibly close to the owner's
+field** and the wall/floor plates and tunnel bands remain.
+
+## Failed and rejected builds/runs
+
+The failures below are retained in the logs; none is presented as a successful
+final check.
+
+- `build_hardjoin.stderr.log` failed with MSVC D8022 because it looked for
+  `standalone.rsp` in the wrong working directory. `build_hardjoin_retry` and
+  the final exact-source build succeeded.
+- `resume2_before_tunnel_audit.stdout.log` is the invalid baseline audit that
+  included `--round-tunnels`; it failed the stair audit at 0.768. Keep it as a
+  failed run, not as the baseline measurement. The corrected run without that
+  option is `resume2_before_tunnel_audit_correct.stdout.log` (0.434 at the
+  older 0.450 threshold); the same corrected baseline scan under the tightened
+  0.400 threshold is `resume2_before_tunnel_diagnostics.stdout.log` and fails
+  at 0.434.
+- Baseline mesh sheet audit `resume2_before_mesh_shard_audit_8.stdout.log`
+  failed with 9 sheets/149 triangles. Earlier candidates also failed:
+  `resume2_trial_mesh_shard_audit_8.stdout.log` (1 sheet/8 triangles) and
+  `resume2_bounded_mesh_shard_audit_16.stdout.log` (1 sheet/20 triangles).
+  The separate radius-18 trial failed with 1 sheet/20 triangles and was
+  reverted; radius 16 passes.
+- `resume2_curvefloor_tunnel_audit_8.stdout.log` is the rejected whole-segment
+  curve trial: only 3,279/5,642 route samples stayed open. The current route
+  test passes 5,642/5,642.
+- Failed floor/detail trial gates include
+  `resume2_after_tunnel_audit_8.stdout.log` (smoothed floor-slope delta 1.082
+  and flat-patch share 21.8%), `resume2_floorrelief_tunnel_audit_16.stdout.log`
+  (slope delta 2.383), the `resume2_familyblend12_*` logs (0.807), the
+  `resume2_profile_baked_relief_*` logs (0.807/2.373), and
+  `resume2_trial_smoothlow_capture.stdout.log` (slope delta 1.082). The
+  no-floor-bridge trial `resume2_no_floorbridges_tunnel_audit_16.stdout.log`
+  failed its flat-patch gate at 20.2%. These variants were discarded.
+- Earlier join trials in `passage_fixed_after_extensions.stdout.log` and
+  `tunnel_overlap_floor_audit.stdout.log` reported 0.0077-voxel endpoint
+  mismatch against a 0.005 threshold. The final radius-profile join audit is
+  0.000000 voxels.
+- The first 18-voxel capture used a repository-root-relative output path while
+  the runner's working directory was `Standalone/`. Capture failed repeatedly;
+  I stopped only that viewer process and reran with the correct path. The
+  corrected 18-voxel captures succeeded, but the 18-voxel mesh audit failed
+  as described above.
+- The 13-control-point radius-filter probe built and captured successfully,
+  but the close-up and bend images showed no visible improvement over the
+  nine-point filter, so the source was restored and rebuilt.
+- The 6-voxel smooth-union probe did not remove the shaft bands; the
+  12-voxel union added lumpy ledges. Both were rejected. Disabling skirts also
+  produced no visible change and its temporary CLI switch was removed.
+- The first floor close-up camera probe was below the surface and was rejected
+  as invalid framing; `final_after_passage_floor_closeup.png` is the raised
+  replacement.
+- During this resume, `candidate_chamber.png` and `candidate_tunnel.png`
+  tested a continuous combined SDF for roughness with a weaker floor-detail
+  boost. The wall became too smooth, so the candidate was rejected.
+- Candidate 2 used broader, quieter relief and a reduced floor boost.
+  `candidate2_whole_view_shard_audit.stdout.log` failed with one 48-triangle
+  sheet at 0.459 voxel thickness; its chamber flat-patch share rose to 13.7%.
+- Candidate 3 reduced only the floor boost. Its grid and sheet checks passed,
+  but the fixed captures stayed too quiet and still showed the floor join.
+  Candidate 4 increased the broad octave; its one-voxel sheet check passed,
+  but the tunnel capture still showed floating plates and boundary components.
+- Candidate 5 added an independent broad amplitude envelope.
+  `candidate5_whole_view_shard_audit.stdout.log` failed with one 18-triangle
+  sheet at 0.678 voxel thickness. The fixed captures still showed plates and
+  shards.
+- Candidate 6 moved floor clipping after the primitive union.
+  `candidate6_chamber_audit.stdout.log` and
+  `candidate6_tunnel_audit.stdout.log` failed the grid-surface audit;
+  `candidate6_whole_view_shard_audit.stdout.log` found one isolated small
+  component, and `candidate6_chamber.png` / `candidate6_tunnel.png` showed
+  severe lattice stairs and shards. Its first shard-audit run printed `status=FAIL`
+  but exited 0 because the exit condition ignored isolated components. The
+  final source corrects that exit condition; its shard audit passes with zero
+  isolated components and zero sub-voxel sheets.
+
+Other retained `GridStairAudit status=FAIL` logs from the exploratory sequence
+are `after_chamber_grid`, `after_csg_chamber_grid`, `after_csg_tunnel_grid`,
+`after_fade12_chamber_grid`, `after_floor_metadata_chamber_grid`,
+`after_floor_metadata_tunnel_grid`, `after_floorblend_chamber_grid`,
+`after_floorblend_tunnel_grid`, `after_scan16_chamber_grid`,
+`after_tunnel_grid`, `critical_shapecontains_tunnel_audit`,
+`final_compare_baseline_passage_w8`, `final16_compare_baseline_passage_w8`,
+`final_grid_tunnel`,
+`floor_profile_tunnel_audit`, `floor_slope_tunnel_audit`,
+`proxy_fade_tunnel_audit`, `resume2_familyblend12_tunnel_capture`,
+`resume2_no_roomclips_chamber_capture`,
+`resume2_no_roomclips_tunnel_capture`,
+`smooth_capsule_union_tunnel_audit`, and `union_width3_tunnel_audit`. They
+record failed or rejected variants/baselines, not the final 16-voxel source.
+
+The corrected grid detector still does not meet the requested before/after
+criterion on every fixed view: the baseline chamber (0.041) and passage
+(0.276) metrics were already below the 0.400 threshold, while the baseline
+tunnel (0.434) fails and the final tunnel (0.028) passes. The final captures
+also still show wall/floor plates and repeated tunnel bands, so these checks
+and images do not justify a visual-completion claim.
 
 No commit, push, or stash was made.

Standalone/Research/OwnerShapeWorld/round2/RunMachine.ps1

@@ -0,0 +1,36 @@
+[CmdletBinding()]
+param(
+    [Parameter(Mandatory=$true)][string]$ExecutableName,
+    [Parameter(Mandatory=$true)][string[]]$ExecutableArguments,
+    [Parameter(Mandatory=$true)][string]$LogStem
+)
+
+$ErrorActionPreference = 'Stop'
+$RoundDir = [IO.Path]::GetFullPath($PSScriptRoot)
+$StandaloneRoot = [IO.Path]::GetFullPath((Join-Path $PSScriptRoot '..\..\..'))
+if ([IO.Path]::GetFileName($ExecutableName) -ne $ExecutableName) { throw 'ExecutableName must be a filename.' }
+if ([IO.Path]::GetFileName($LogStem) -ne $LogStem) { throw 'LogStem must be a filename stem.' }
+$ExecutablePath = Join-Path (Join-Path $StandaloneRoot 'bin') $ExecutableName
+if (-not (Test-Path -LiteralPath $ExecutablePath -PathType Leaf)) { throw "Executable not found: $ExecutablePath" }
+$StdoutPath = Join-Path $RoundDir ($LogStem + '.stdout.log')
+$StderrPath = Join-Path $RoundDir ($LogStem + '.stderr.log')
+
+$HostProcess = Get-Process -Id $PID
+$HostProcess.PriorityClass = 'Normal'
+$HostProcess.ProcessorAffinity = [IntPtr]0xFFFF
+$Child = Start-Process -FilePath $ExecutablePath -ArgumentList $ExecutableArguments `
+    -WorkingDirectory $StandaloneRoot -WindowStyle Hidden -PassThru `
+    -RedirectStandardOutput $StdoutPath -RedirectStandardError $StderrPath
+$PeakWorkingSet = 0L
+while (-not $Child.HasExited) {
+    $Child.Refresh()
+    if ($Child.HasExited) { break }
+    $Child.PriorityClass = 'Normal'
+    $Child.ProcessorAffinity = [IntPtr]0xFFFF
+    $PeakWorkingSet = [Math]::Max($PeakWorkingSet,$Child.WorkingSet64)
+    Start-Sleep -Milliseconds 500
+}
+$Child.WaitForExit()
+$Child.Refresh()
+Write-Output "RUN_EXIT pid=$($Child.Id) priority=Normal affinity=0xFFFF peak_process_gb=$([math]::Round($PeakWorkingSet/1GB,2)) exit_code=$($Child.ExitCode)"
+exit $Child.ExitCode