gig-line/MEMORY.md

MEMORY.md

Shared working memory between sessions and collaborators. Short, dated entries with evidence: the command, the log line, or the file path.

Scope: facts and traps that cost time to establish. Planned work lives in TODO.md; development direction lives in ROADMAP.md; formal setup, pins, and licensing live in Docs/.

---

2026-11-08 — Documentation toolchain installed; asciidoc→markdown conversion

• pandoc 3.13 installed at ~/.local/bin/pandoc from the GitHub release tarball (statically linked, 161 MB). The Fedora repos carry only ghc-pandoc (Haskell library), not the CLI.
• asciidoctor 2.0.26 layered via rpm-ostree install rubygem-asciidoctor — pending reboot to activate. As a workaround, a distrobox Fedora container (distrobox enter arch-convert) runs asciidoctor without rebooting.
• prettier 3.9.10 installed globally via npm (Node v24.20.0). Project config: .prettierrc (proseWrap: never, tabWidth: 2), .prettierignore (skips Binaries/, *.uasset, .agents/).
• Conversion chain: .adoc → asciidoctor -b docbook5 → .xml → pandoc -f docbook -t gfm → .md → prettier --write. Pandoc's built-in asciidoc reader cannot handle Asciidoctor tables or attributes — it must go through docbook5.
• Post-processing fixes: glossary cols="e,2e" emphasis leaks italic into cell content (strip *...* wrapping), ® from the R key binding becomes (R), anchor IDs {#...} stripped, escaped quotes/brackets normalised.
• All .adoc files removed. Markdown is the canonical format. Architecture documentation lives in Docs/Architecture/ as README.md, twelve numbered section pages, and about-arc42.md.
• All .md in the repo passes prettier --check (23 files formatted, .agents/ excluded as vendored content).

2026-11-08 — QA branch merged; demo is walkable

• PR #7 squash-merged to main as 52b1212 — terrain collision fix, spawn correction, UGigLineAssetTools, HighResShot fix, -qaselftest harness.
• QA self-test results (from merged main): walk 750 cm / 2 s (Z steady 456), reset 0 cm from spawn, pause/unpause pass, transect 84 samples / 0 holes (ground Z 317..395).
• Traversal release check closed: the swept transect measures walkability rather than assuming it.
• gameplay-crossing-0.1.0 branch deleted (already merged); gig-line-judge worktree removed.
• qa-input-collision-fix-0.1.0 branch still exists locally/remote (squash merge has no shared ancestry, git branch -d refuses; delete with -D after release tag).
• GigLine SIM / GigLine GROUND TEMP diagnostic logging still in GigLineCharacter.cpp — must be removed before v0.1.0 tag.

2026-11-08 — Reset was returning to mid-air; QA self-test switch

• ResetToStart() restored a StartLocation captured in BeginPlay, which is the PlayerStart position above grade (Z 600) so the capsule never spawns embedded. Every reset therefore dropped the player 1.4 m back into the air. Fix: capture StartLocation on the first grounded frame (repeating 0.25 s timer, cleared once captured) — not a one-shot timer, which can fire after the player has already started walking and record a mid-stride position.
• -qaselftest drives the checks headlessly and logs GigLine QA: ... PASS/FAIL. It runs on a core ticker, not a world timer: pausing the game stops world timers, so a timer-driven test freezes exactly when it proves pause works.
• Current results: walk 746 cm / 2 s with Z steady, reset 0 cm from spawn, pause IsGamePaused() 1, unpause 0, transect 84 samples / 0 holes (ground Z 317..395).

2026-11-08 — The fall-through was an 8 m invisible box (QA branch)

• Symptom: player "spawned above ground but never fell", sat at Z 909, fall=0, mode=1 (Walking), no fall-through, view looked plausible so it was mistaken for working.
• Cause: Terrain_DEM had only a bounding-box simple collision proxy (Interchange default). A line trace down at every Y through the block hit Demo_Terrain_DEM at exactly 819 — the mesh bbox max Z. The whole street was under an invisible lid; the pawn stood on the box, 8 m above the road.
• Proof: a debug LineTraceSingleByChannel profile at Y −1912…+1447 reported top=819 actor=Demo_Terrain_DEM for the entire street. After the fix it reports Sidewalk_North 395, Road_Hollywood 366, Sidewalk_South 317.
• Fix: force complex-as-simple collision on the terrain. Python cannot set it (StaticMesh/BodySetup expose no such property; dir() shows only lod_for_collision). Added UGigLineAssetTools::SetComplexCollisionAsSimple in the C++ module (BodySetup->CollisionTraceFlag = CTF_UseComplexAsSimple, UE 5.8 no longer has bUseComplexAsSimpleCollision).
• Second gotcha: the flag lives on the asset, not the map. Setting it in a builder session and saving only the .umap loses it — EditorAssetLibrary.save_loaded_asset() is required, otherwise the next run reloads the bbox proxy.
• Third gotcha: HighResShot 1280 720 means multiplier 1280 (UE warns "the high resolution screenshot multiplier is too large for your system"). The correct form is HighResShot 1280x720.
• Resulting spawn: (-13336, -1150, 600) — inside the carved corridor, lands on the road at Z 456 (feet 368 on road top 366).

2026-11-07 — Publication audit

• MPFB 2.0.18's installed blender_manifest.toml identifies its code as GPL-3.0-or-later; its versioned LICENSE.md explicitly puts bundled mesh, texture, and rig assets under CC0 1.0. The earlier CC0 claim for the add-on code was wrong.
• Blender inspection found no packed images or linked libraries in the tracked scenes. The character scene references local MPFB textures and a temporary baked texture, so those paths are not portable to a fresh checkout.
• The 2026-11-07 publication rewrite removed the approved private documents and bundled skills from all reachable refs; local copies remain ignored. The original history backup at /tmp/gig-line-private-history-backup-42a29fe2 contains those private files and must not be pushed; /tmp may be cleared by the system.
• git ls-remote --refs origin succeeded on 2026-11-07 with no advertised refs; git ls-remote --exit-code origin HEAD returned 2 because no HEAD was advertised. This does not verify LFS upload support.
• Gitleaks 8.28.0 found an Android File Server token in historical Game/GigLine/Config/DefaultEngine.ini. The token was removed from reachable history and the unused server disabled; the old private backup and earlier temporary sanitized clone still contain it and must not be published.
• The first Gitea push uploaded 26 LFS objects. An independent clone fetched all 26 and git lfs fsck passed for the full history; the server also advertised LFS locking support.

2026-11-07 — Unreal Engine 5.8.4

• Extracted to /run/media/system/DATA/Engines/UE_5.8.4 (73 GiB, unzip exit 0). CL 58902052, ++UE5+Release-5.8.
• Gotcha: the editor needs a writable ~/.config/Epic. In the agent shell it crashed with signal 11 twice and looked like a broken install. The log named it: create dir('/home/zwitschi/.config/Epic/') failed: errno=30 (Read-only file system) → Fatal error: Failed to create Intermediate directory '…/.config/Epic/UnrealEngine/5.8/Intermediate/'. Rerunning with HOME=/tmp/uehome XDG_CONFIG_HOME=/tmp/uehome/.config reached LogInit: Display: Engine is initialized.
• Vulkan RHI is still unproven. The successful run used -NullRHI. Untested: UnrealEditor -vulkan from a desktop terminal.
• A packaged game does not need ~/.config/Epic, so this failure does not follow the build.

2026-11-07 — Agent shell limits

• The agent shell has a read-only home and read-only /run: sudo is blocked (no new privileges), flatpak run and podman cannot create instances, /dev/nvidia* is absent so nvidia-smi always fails there. None of this is a broken machine; verify GPU things from a normal terminal.
• Anything outside the session workspace (/run/media/system/DATA/Engines, /tmp) needs the wider sandbox with an approval prompt.

2026-11-07 — Reference photography

• Licensing is not a gate for local development. Reference images live in the gitignored Art/reference/ and are never committed. What each set establishes is written into Docs/reference-notes-hollywood.md.
• Commons text search runs dry; categories do not. Tools/gen/commons_fetch.py --category "Category:Hollywood Masonic Temple" surfaced 15 files a search never showed, including 4–5k px façades.
• upload.wikimedia.org throttles hard. Batches of six or fewer with --delay 3; a throttled file is skipped and recorded rather than killing the run (HTTP 429).
• Unqualified queries return junk. Hollywood Boulevard shops returned a Münster street, Hollywood Masonic Temple entrance a Seattle card index. Check the hit list before accepting a set.
• OpenRouter web search is expensive. One gpt-5.2 + openrouter:web_search call ran 17 searches and cost $0.38; two of its eight URLs were Commons files already held. Use it for a known gap, not a sweep. Tools/gen/page_image_fetch.py turns results into files by reading og:image.
• The inherited "eight stone steps, 12 m wide" is wrong. Photographs show three portals, each with its own short flight. Do not copy that number into geometry.

2026-11-06 — Blender and Blosm, headless

• flatpak run does not work here. Call the binary inside the app tree: B=/var/lib/flatpak/app/org.blender.Blender/x86_64/stable/active/files, then LD_LIBRARY_PATH="$B/blender/lib" "$B/blender/blender" -b … with BLENDER_USER_SCRIPTS="$HOME/.var/app/org.blender.Blender/config/blender/5.2/scripts".
• Blosm needs two headless workarounds, both already in Art/blender/blockout/blosm_blockout.py:
  1. it refuses to import until an asset directory is set, and the check fires before the mode is honoured;
  2. in background mode the default level table is empty, and the add-on then takes a path with a broken relative import (from gui import addDefaultLevels). Seeding one entry avoids it.
• Blosm projects on a sphere, not an ellipsoid. Its frame is about 0.4 % oversized and rotated by meridian convergence — roughly 2 m at the theatre. Fixed with a six-parameter affine into EPSG:26911 (0.0 cm residual).
• Transform roots only. 256 of 275 objects are children; transforming both parent and child applies the correction twice and reads as a 2 m placement error.
• Blosm quietly fetches from Overpass for relation members missing from the extract. loadMissingMembers is off by default in the tool so imports stay offline and reproducible.

2026-11-06 — Geo data traps

• The 3DEP export is a tiled TIFF, not strips. A strip reader silently scrambles rows and produces fake 0 m / 127 m bands. Geo/tools/dem_stats.py handles tiles.
• Sample walk routes through PROJ. Placing samples with the OSM spherical approximation put them metres off the UTM meshes and reported 91 phantom holes. /tmp is not shared between tool calls here either — regenerate instead of saving intermediate files there.
• 3DEP cannot resolve kerbs. Sharpest break near the kerb line is 0.09–0.12 m/m, a gutter transition, not a 0.15 m upstand. Kerb height stays a stated assumption.
• Building interiors are not ground. Points inside the hero footprint returned 114.5–117.1 m — structure, not terrain. Excluded from the relief figure.
• Two measurements independently agree on the carriageway: crossing length 20.9 m and gutter spacing −10 / +10…12 m.

2026-11-06 — Repository traps

• Without git lfs install the lfs filter is inactive and git add does not warn — it stores the file as a plain blob. Tools/check-lfs.sh is the guard.
• Never commit reference images or .env. .gitignore covers /Art/reference/, .env, *.env. When images slipped in early, history was rewritten to remove them.
• The project origin is Hollywood × Highland: E 376 515.389 / N 3 774 224.083 in EPSG:26911. Blender X east, Y north, Z up, metres, Z measured from the origin ground (NAVD88 115.911 m). The whole block sits at negative X.

2026-11-07 — Unreal pipeline findings

• UE 5.8 game targets must use BuildSettingsVersion.V7 to share the installed editor's build environment; V5 is rejected with "modifies the values of properties … not allowed".
• UnrealEditor-Cmd -run=pythonscript swallows Python print() with -stdout in 5.8 — write results to a file instead (see Tools/unreal/ue_scale_test_import.py).
• unreal.Box has min/max, not origin/get_extent() in 5.8 Python; compute centre/size from the corners.
• FBX Blender→UE axis mapping is east→+X, north→−Y, up→+Z with preset ue_metric_v1; the Y mirror is inherent (RH→LH). Compass math in game code: north = −Y. Verified with the three-cube scale test.
• Headless asset import works fine with -nullrhi; unreal.AssetImportTask + FbxImportUI is the pattern (Interchange pipeline under the hood).
• Flatpak Blender headless needs XDG_CONFIG_HOME=$HOME/.var/app/org.blender.Blender/config for extensions: bpy.utils.user_resource('EXTENSIONS') otherwise resolves to ~/.config/blender/5.2/extensions (empty) and bl_ext.blender_org.mpfb is undiscoverable. Blosm unaffected (scripts/addons via BLENDER_USER_SCRIPTS). Found via Tools/addon_compat_probe.py; documented in Art/blender/README.md.
• UE 5.8 Interchange skips material/texture assets for skeletal FBX imports (static imports do get materials). Finish the chain in Python: import texture, build Material via MaterialEditingLibrary, assign with sm.set_editor_property('materials', [SkeletalMaterial]) — the struct field is material_interface, not material, and SkeletalMesh.set_material() does not exist. See Tools/unreal/ue_character_import.py.
• Blender 5.2 FBX exporter removed armature_nla (passes for 4.x recipes fail). Also: a baked image is only referenced by the FBX exporter when image.filepath (not just filepath_raw) is set and the exporter can map the node — MPFB's v2 skin node-group tree can't be mapped, so a flat export material is required.

2026-11-07 — Collaboration and scope

• Two sessions edit this repo at once. The other session created the Unreal project and the blank package while this one was documenting. Check git log and git status before committing, and never git add -A blindly: it once swept generated Game/GigLine/Build/** logs into a commit. Stages the source files deliberately.
• Build/ under a project is generated, not source — FileOpenOrder/*.log and packaging output. .gitignore now covers it alongside Binaries/, Intermediate/, Saved/, DerivedDataCache/.
• Scope boundary for 0.1.0: do not tag v0.1.0 without an explicit go-ahead from the owner. It is a minor release and carries the walking/driving demo promise; the board's release checks decide when it is ready, not the pace of work.

2026-11-07 — Screenshot pipeline

• -executePythonScript + editor is unreliable for captures: the editor quits right after a successful deferred script (async screenshot requests never render), and after load_map in a commandlet every spawn_actor call returns None — build scenes in the default world instead.
• Robust in-game capture recipe: commandlet builds a map (ue_build_review_map.py), then the game runs in -game mode and a tiny C++ actor (AGigLineShotActor) walks camera setups via SetViewTarget + FScreenshotRequest::RequestScreenshot. In 5.8 FScreenshotRequest has no resolution argument (window size via -ResX/-ResY) and no IsFinished() — poll !IsScreenshotRequested(). Lives in UnrealClient.h, not Misc/ScreenshotManager.h.
• Commandlet-only quirks: bulk render-target readback (read_render_target_raw, export_render_target) silently returns empty in commandlets; single-pixel read_render_target_uv works. Game targets must use BuildSettingsVersion.V7.

2026-11-07 — Enhanced Input from Python (UE 5.8)

• unreal.InputCoreTypes does not exist in the 5.8 Python bindings, and unreal.Key takes no constructor argument. Build a key with k = unreal.Key(); k.import_text("SpaceBar") — the export text is the key name (MouseX, Gamepad_Left2D).
• Factories are unreal.InputAction_Factory and unreal.InputMappingContext_Factory, not InputActionFactory. create_asset(..., factory=None) also works.
• Do not delete an asset and recreate it in the same commandlet run. create_asset returns None for a path deleted earlier in that process. Load the existing asset and overwrite it instead (Tools/unreal/ue_create_input_assets.py: fresh()).
• Modifiers on EnhancedActionKeyMapping do not survive saving. Assigning InputModifierNegate to the mapping's modifiers reads back correctly in memory, even after a gc.collect(), but the saved asset reads mods=[NoneType]. Keep-alive references, explicit flag setting, and in-place saves all failed to change it. Workaround: bind one action per direction (W and S as separate actions) so no modifier is needed.
• A read-only packed .umap silently blocks save_map. EditorLoadingAndSavingUtils.save_map returned False with no error until the file's mode was fixed (chmod u+w). Symptom: a rebuilt review map keeps three cameras after a fourth is added.
• -run=pythonscript still swallows print(), so write results to a file; /tmp is visible to the commandlets even though it is not shared between agent tool calls.

2026-11-07 — Hero facade build and review cameras

• The review floor was 5 cm thick and centred at Z -50 cm, so anything placed at Z 0 floated ~1.8 m. It is now a 90 m x 90 m x 10 cm slab with its top face at Z 0; a character's bounds should read Z 0..83 cm when it stands on it. The review floor is the only ground in the scene. Cameras 25 m from the building line look past its edge and see the black horizon band under the facade. Place facade cameras so both the facade and enough floor are in frame, or extend the floor.
• The facade's frontage faces Unreal -Y. The FBX preset maps Blender +Y (north) to Unreal -Y, so the elevation looks at negative Y and a straight-on camera must stand at Y < -1450 cm. A camera at positive Y looks at the back wall; cameras inside Y -1443..+250 are inside the building and render flat blue.
• The imported mesh is centre-origin in X but not in Y or Z. Bounds are x -1302..1302, y -7..1250, z 0..1450 cm, so to put the front plane on the building line the actor goes at Y = line + 7 cm, and X = east edge + half width.
• save_map returns False, without an error, when the packed .umap is read-only. The review map silently keeps its old camera list. chmod u+w the file before rebuilding; a umap that keeps rendering four shots after a sixth was added is this, not a caching problem.
• StaticMaterial is a struct: material.get_editor_property("material_interface"), no get_name() on the wrapper.
• Review screenshots are committed on the owner's instruction (2026-11-07). They are large PNGs, so a review run adds weight to the repo; keep the shot list short and delete retired shots rather than accumulating.

2026-11-07 — Sibling project reuse survey

• security-game holds 20 reusable street props with Blender sources (assets/models/blender_src/*.blend) and Godot GLBs: bollard, bench, trash bin, stanchion, crowd barrier, street lamp, traffic cone, fire hydrant, newspaper box, palm/street trees, sedan, marquee, doors, lanterns. About 727 KB and 11.8k triangles for the set. Audit: Tools/sibling_asset_audit.py, manifest at Art/props/reuse-manifest.json; the 50 .blend sources are copied to Art/props/blender_src/.
• Their building meshes are not reusable — they are Godot GLBs built to a different key set (hero facade scaled to 29.6 m, a figure this project measured at 25.75 m). Do not import their venue geometry.
• tools/generate_textures.py in that project is the OpenRouter image endpoint pattern (POST /api/v1/images, b64_json, optional reference_path for image-to-image on the chat endpoint) and it already has procedural brick_facade / seamless_concrete / asphalt generators. Worth porting for our facade materials.
• Its AI concept frames are restaged photograph references (.refgen.json sidecars point at photos with a "keep geometry fixed" suffix). Useful as art direction, not as texture sources.
• Blosm exposes setupScript (scene.blosm.setupScript → loadSetupScript) for custom building materials, and ships setup/neoclassical.py as an example — the hook for giving OSM buildings a per-building facade style instead of flat colours.

2026-11-07 — Geo tooling layout

• One conversion, one place. Geo/tools/osm_common.py::to_unreal_local is the only metres→centimetres conversion; every exporter calls it. os_common also needed an explicit import os at module level — a helper defined with a function-local import list crashed with NameError: os when called from a fresh interpreter.
• asset_metadata.py audit must recurse. Sidecars sit next to their .blend, which is one level below Art/blender/; a flat glob found nothing and reported "0 sidecars, 0 problems" — a green result that meant "did not look".
• Blender-side builders live in Art/blender/*/ (terrain, street, site, blockout) and reach helpers via os.path.join(HERE, "..","..","..","Geo","tools"). Moving them breaks the old "..", "Geo", "scripts" path silently until the script is run.
• Height provenance: 3 of the 15 block buildings carry no OSM height and Blosm renders them at a placeholder 6.4 m. Geo/tools/block_height_report.py lists them; they need heights measured off the photographs before the skyline is believable.

2026-11-07 — Blender rebuilds that silently did nothing

• A read-only .blend makes Blender skip the save without failing the run. Tools/blender/hero_facade_build.py printed its part list and "bounds" as normal, then went straight to the export: the .blend and the report stayed at their old timestamps, and the Unreal re-import brought back the old mesh. Symptom to watch for: the imported bounds do not change after a rebuild. Fix: chmod u+w the .blend first, then confirm the timestamp moved.
• Model bottom must land on Z 0. The hero portal jambs were sized PORTAL_H + 0.3 and centred at PORTAL_H / 2, so they hung 0.15 m below the pavement and the whole facade imported with min z = -15 cm. Always check the asset's bounds_m.min.z in the build report.
• Rebuilding a facade is not enough to make a detail shot work: doors modelled as thin panels on a solid plinth read as shading, not openings. Cutting the plinth into piers with a lintel band is what makes them legible from 8 m.

2026-11-07 — Demo map assembly traps

• Unreal signs, measured not assumed. The FBX preset flips north, so in the demo map the road spans Y -1382..+1047, the north sidewalk Y -1912..-1053, the south sidewalk Y +718..+1447, and the facade sits at Y -1450. A PlayerStart at +1600 is on the south sidewalk behind the building. Read the actor bounds before placing anything.
• Street meshes carry only PART of the sidewalk. In the source .blend, Sidewalk_South is the offset band (-14.47..-7.18 m) and Sidewalk_North is the kerb-to-wall band (10.53..19.12 m); the rest of the north walk comes from the terrain. So a sidewalk-looking strip of ground between kerb and wall is the terrain, not a gap.
• Do not mirror in the exporter. Sidewalk_South was already correct in the .blend; mirroring both sidewalks at export moved the correct one to the wrong side and produced an "everything on the wrong side" look that took three rebuilds to unpick. Check the source coordinates first.
• bpy.ops.mesh.normals_make_consistent needs a mesh context; in object mode it fails with poll() failed, context is incorrect. Flip winding by reversing polygon.vertices instead.
• StaticMesh has no trace-flag property in Python. Collision complexity is set through mesh.get_editor_property("body_setup").set_editor_property("collision_trace_flag", …); StaticMeshEditorSubsystem.get_collision_complexity reads it back but there is no setter.
• A -game run with a paused world still exits on timeout, so a smoke test that "hangs" may simply be running.

2026-11-07 — Documentation sweep

• Docs drift silently after a refactor. Moving the geo scripts to Geo/tools/ and the Blender builders into Art/blender/*/ left stale paths in six documents and two tool READMEs. Grep for old paths (.md and .py) after any move, not just the ones that error at runtime.
• Every new tool needs a row in Tools/README.md and Docs/build.md, or the next session will not know it exists. Both now list the full content pipeline in order.

2026-11-07 — Character diffuse bake

• img.save() writes black for an image created in memory even though Cycles bpy.ops.object.bake does update img.pixels correctly (verified in-memory px were warm). Writing the PNG ourselves from img.pixels is required.
• A per-pixel struct.pack loop over img.pixels is catastrophic in headless Blender (256² took 25 s). Blender ships numpy: read img.pixels[:] once into a float array, reshape (h,w,4), clip→uint8, reverse rows for PNG top-down, then zlib-compress. Whole 1024² encode ≈ 0.3 s.
• MPFB's v2-skin node tree does not resolve to a simple base colour, so a Cycles DIFFUSE bake returns black. Since the FBX export material replaces the node tree anyway, synthesize the diffuse directly (make_diffuse): warm skin body + a darker shorts-suggestive bottom band.
• A read-only .umap silently makes save_map return False with no error (already recorded); re-hit in the review pipeline — chmod u+w the maps before rebuilding, otherwise the new character material never appears in shots.

2026-11-07 — screenshot judge tool

• Tools/review/judge_screenshot.py — runs an OpenRouter vision model over a screenshot vs an expected-scene description. Reads OPENROUTER_API_KEY from the repo .env (ignored). Base64 data-URI image under image_url.url; OpenAI-compatible body → POST /api/v1/chat/completions. openai/gpt-4o-mini returns a {verdict, confidence, reason, missing} JSON. It flags genuinely missing content (e.g. main's pre-warm-skin character shot judged "no"); it is a pre-check, not the human verdict.
• Test one: Tools/review/judge_screenshot.py --image Screenshots/review-01-character.png --expect "<scene>" --verbose. Bulk: --shots (parses the README rows).

2026-11-07 — Blender skill packs: discovery depth

• The 94 packs were never registered with the skill tool, only readable as files. The loader resolves <skills-root>/<pack>/SKILL.md one level down; .agents/skills/blender-skills/ is a container of 94 packs, so it was skipped silently (no warning, no error). Symptom: skill("archviz") unresolvable even though archviz/SKILL.md exists.
• Fix applied: 94 symlinks in ~/.agents/skills/<pack> → <repo>/.agents/skills/blender-skills/<pack>/. The catalog then lists all of them (verified by calling skill on collision-proxy: base dir resolves through the link). references/ stays unlinked — it is a shared folder with no SKILL.md, not a pack.
• The AGENTS.md "MCP-driven" caveat is overstated. 41 of 94 SKILL.md mention MCP, mostly only in the description line; the bodies are method guides (naming, proportions, validation checklists) that work headless. Do not skip a pack because it advertises MCP.
• Skill names in AGENTS.md/TODO.md all resolve except references, which is the folder above, not a pack.
• Symlinks are user-level: a fresh machine needs the same 94 links before the packs are loadable by name.
• A shared checkout can commit your uncommitted edit for you, or lose it. This entry was written but not yet committed when another session ran a git commit that included MEMORY.md; it landed inside their commit (54b6239) with a message that does not mention it. Had they overwritten the file instead, the work would have been gone with no reflog to recover from. On a shared worktree, either commit a shared file immediately or check git log -S for it before assuming the edit is missing.

2026-11-07 — Street visibility and the LFS lockable trap

• Every street surface (road, sidewalks, kerbs, markings) had 100 % downward face normals — strip()'s natural winding (rows along +s, cols along +normal) faces the ground. Invisible in engine for as long as the terrain covered them; the corridor cut exposed the black trench. Fixed by reversing the winding in strip(), kerb_faces() and crossing_band().
• *.blend and *.umap are LFS-lockable in .gitattributes — Git LFS checks them out read-only (444). bpy.ops.wm.save_as_mainfile then fails ("path is not writable") and EditorLoadingAndSavingUtils.save_map silently returns False. All Blender builders now chmod u+w before saving and treat a failed save as fatal; the UE map builders chmod the .umap first. A failed save used to leave downstream steps building from stale files — that produced a degenerate-mesh FBX (all-same vertex indices from an index-arithmetic bug) that cost an hour to trace.
• FBX SDK refuses to triangulate the crown-twisted crossing quads ("unable to triangulate mesh ... will be omitted") — emit triangles from the builder, and make sure per-stripe vertex indices advance (len(verts) must be captured before the loop, not per-corner).
• Demo-map coordinate note: UE Y is south-positive; the facade at Y −1450 and the player start at Y −3500 are on opposite sides of the road (Y −1382..+1047) — arrangement is correct; the Sidewalk_North/South mesh names are swapped relative to their physical sides (documented in export_site_fbx.py).

2026-11-07 — ReviewMap corruption: deprecated level calls + save-order bug

• EditorLevelLibrary.load_level/new_level are deprecated no-ops in -nullrhi commandlets (they return without switching worlds, so staged actors land in whatever world is current and get saved over the wrong map). Use the LevelEditorSubsystem: unreal.get_editor_subsystem(unreal.LevelEditorSubsystem).load_level(path) actually switches. Symptom was ReviewMap containing Demo actors + ShotSequencer.
• Map-save ordering bug in Tools/unreal/ue_build_review_map.py: the script saved ReviewMap after the DemoMap section, but the switch back to ReviewMap fails silently because the map only exists in memory at that point — so the final save_map wrote the demo world over ReviewMap. Fix: save ReviewMap first, then stage the demo cameras, and never switch back. Both maps verified by probe (ReviewMap lists Review actors; DemoMap lists Demo actors).
• Rebuilt ReviewMap after the fixes: character, both cubes (A grounded, B floating at 2 m), pawn, NPC prefab, hero facade all present; user re-approved the shots. README cube naming corrected (README had A/B swapped vs the script).

2026-11-07 — Materials and lighting in the demo map

• Blender flat diffuse colours do not survive FBX -> Unreal. Every site mesh (road, sidewalks, kerbs, crossing paint) imported with a material asset that had a default white base colour, and the terrain imported with WorldGridMaterial (UE's bright grey grid). That — not the sun — is what made the demo look blown out and washed white. Fix: create the colours in UE (M_Asphalt, M_Concrete, M_Ground, M_RoadPaint via Constant3Vector) and assign them per actor in Tools/unreal/ue_build_demo_map.py.
• Auto-exposure makes review shots non-comparable: the same road rendered washed-out in one framing and near-black in another because each camera saw a different amount of sky. Fix: an unbound PostProcessVolume with AEM_HISTOGRAM and auto_exposure_min_brightness == auto_exposure_max_brightness (fixed exposure). Do not use AEM_MANUAL + auto_exposure_bias — a bias of 11 blew the whole frame out.
• DirectionalLightComponent.light_color takes a Color struct, not LinearColor (unreal.Color(r=255,g=245,b=230,a=255)); passing LinearColor raises a nativize error.
• One band per object loses geometry downstream. crossing_band() created Crossing_Markings each call, so Blender renamed the second and third to .001/.002; site_assembly.py appends by exact name, so only one crossing ever reached the game. Build all bands into a single mesh instead of one object per band.
• Street arc length -> UE X: s=0 is the west end of the block (UE X −27966 cm) and s increases eastward, so the designated mid-block crossing at s=130.1 m is UE X ≈ −14956, not the −24291 used by an earlier camera (that is the s=36.6 crossing).
• Camera placement in the demo map needs the mesh bounds, not intuition. The hero façade body spans UE Y −1457..−200, so a "street level" camera at Y −400 is inside the building and renders its unlit back face. Read get_bounding_box() on the staged meshes before choosing a camera.
• "LIGHTING NEEDS TO BE REBUILT" (Build Lighting Only / BuildLighting) appears whenever a map has Stationary/Static lights and no baked lightmap. For a fully dynamic scene the clean fix is r.AllowStaticLighting=False in Config/DefaultEngine.ini plus set_mobility(ComponentMobility.MOVABLE) on the directional light, sky light and height fog (and MOVABLE on the staged mesh components). Nothing to build, warning gone. Note the cvar is read at startup, so each new commandlet/game run picks it up.
• The Android File Server SecurityToken kept reappearing in Config/DefaultEngine.ini; the plugin is disabled and excluded from shipping, so the value is cleared and must stay empty — it is a credential and must not be committed.

2026-11-07 — Worktrees and the material look pass

• A fresh git worktree has no compiled game binaries. Binaries/ is gitignored, so UnrealEditor-Cmd on a new worktree exits "abnormally (error code: 1)" before running the Python script, with no Python error in the log. Run Engine/Build/BatchFiles/Linux/Build.sh GigLineEditor Linux Development -project=<worktree>/Game/GigLine/GigLine.uproject first.
• /tmp/gigline/*.txt result files are shared across worktrees. Reading /tmp/gigline/demo-map.txt after a run can show another worktree's stale output — check the file mtime, or use a per-worktree result path.
• Flat Constant3Vector materials read as plastic. Giving each surface a roughness plus a world-space-noise blend between two shades (WorldPosition → Multiply → Noise → Lerp between two Constant3Vectors → BaseColor) is enough to stop the blockout looking like a toy; no textures required.

2026-11-07 — The Y-side bug (found when the frontage was added)

• UE convention in this project: south = +Y, north = −Y. Ground truth is the street mesh: Sidewalk_North is Y −1912..−1053 and Sidewalk_South Y +718..+1447.
• The hero facade had been placed at Y −1450 (north side) and X −12024..−9420 — 26 m east of the 6840 footprint and on the wrong side of the street. It only looked right because it stood alone in an empty field; the approved screenshots were of a building with no context.
• The two north-side buildings (TCL, Tussaud's) were similarly on the south side: export_north_buildings.py mirrored Y, which put them where the frontage belongs.
• Fix: hero at X −13336 / Y +1307 (the frontage gap between 6902 and 6838); frontage and north buildings exported without a Y mirror (the FBX preset already maps Blender −Y south to UE +Y south).
• Lesson: place a building only against a neighbour or the street mesh, never in isolation. A single object in an empty map hides any number of placement errors; adding the adjacent frontage immediately exposed this one.

2026-11-07 — walkable demo QA bugs (fix branch qa-input-collision-fix-0.1.0)

• "Stuck" window = wrong default map. A plain ./GigLine.sh opened /Engine/Maps/Templates/OpenWorld (empty engine template) instead of the demo, so nothing was interactive. GameDefaultMap was in DefaultGame.ini and not applied; it must live in DefaultEngine.ini[/Script/EngineSettings.GameMapsSettings] with GlobalDefaultGameMode. Verified: plain launch now loads /Game/Hollywood/DemoMap. Evidence: LogLoad: LoadMap: /Engine/Maps/Templates/OpenWorld.
• Input assets were pruned from the pak. The Enhanced Input assets under /Game/Input were only soft-referenced at runtime, so the cooker excluded them (cook ReferencedSet.txt had zero /Game/Input lines -> all inputs nil at runtime). Fix: hard-reference via ConstructorHelpers::FObjectFinder in the character ctor (cooker follows hard refs) + +DirectoriesToAlwaysCook. Packaged log then showed ctx=ok move=ok ... jump=ok.
• Character fell through the site: site meshes staged MOVABLE + PlayerStart far past the north sidewalk. Fix: STATIC mobility + explicit collision (QUERY_AND_PHYSICS/WorldStatic/BlockAll) and PlayerStart at Y -1200 on the sidewalk. Packaged run: grounded=1 movement=Walking.
• Diagnose packaged runtime via the engine log (.../Linux/GigLine/Saved/Logs/GigLine.log); a headless -nullrhi commandlet cannot validate physics traces (even a spawned cube returns NO HIT there).

2026-11-08 — reference photography gaps closed (local-only)

• Closed the two missing reference angles for the hero façade, both from Wikimedia Commons (gitignored deliverables under Art/reference/, provenance in sidecars):
  • el-capitan-entrance/ — File:Hollywood El Capitan entrance.jpg (+ detail), Andreas Praefcke, CC BY 3.0. This is the isolated hero (6840) entrance/door recess that the notes doc had claimed Commons didn't hold. Query that found it: --query "El Capitan entrance".
  • el-capitan-night/ — File:El Capitan theater with Tron marquee.jpg (CC BY-SA 3.0) + West Side Story premiere (Adam Jones, CC BY 2.0). The theatre is the building NEXT to 6840; treat as proximity evidence only, not the hero.
• Commons throttles hard: HTTP 429 after a batch. Fetch fewer files with --delay/separate calls and wait ~15-20s between runs (queue.py tool default 1 req/s is insufficient once rate limited).
• Task was claimed on branch ref-photo-0.1.2; delivers only the TODO closure (all actual reference files are gitignored, so no code branch content). Reported to coordinator for merge/cleanup.

2026-11-08 — security-game scan: concepts copied, licensing ignored (owner)

• Scanned /run/media/system/DATA/Documents/06-games/security-game for reference material vs our Art/reference/. Result: its photo set (docs/screens/refs/ hollywood/, 9 files) is already fully duplicated in our Art/reference/hollywood/ (with SIDECAR/BOARD copies) — nothing new there.
• What we lacked: its AI concept renders (docs/screens/concepts/ hollywood_boulevard/, 12 PNGs + .refgen.json). Copied into Art/reference/el-capitan-concepts/ with a folder README. Owner explicitly said ignore licensing completely — they are gitignored local development reference, never committed or shipped; .refgen.json still records the model and source photo per file for traceability.
• Use concepts for mood/lighting intent only; geometry truth stays with the photo refs (facade, entrance, night marquee, TCL/Tussauds, Masonic 1922, etc.).
• The security-game realism/ A/B screenshots and concepts/joke_shop/ are its own engine/venue renders — not relevant to our 6840 modeling.

2026-11-08 — ODbL review closed; remaining active TODO items finished

• Wrote Docs/odbl-review.md (decision 2026-11-08): the shipped block geometry (streets/footprints/sidewalks) is a Produced Work of the OpenStreetMap database (ODbL 1.0). Obligations held: on-screen attribution via AGigLineHUD ("© OpenStreetMap contributors" + "Data © OSM, ODbL 1.0 — openstreetmap.org/copyright"), ODbL link, and the raw extract committed in Geo/ with a manifest (date, bbox, SHA-256). Share-alike is limited to the data layer; project code/art remain CC0 (Docs/licensing.md decision). 3DEP DEM is US Government public domain, unrelated to ODbL.
• Closed both [A] items in TODO: (1) sidewalks/roads/collision — collision verified; remaining kerb radii/aprons/kerb-height explicitly tracked as 0.1.1 patch tasks (Sidewalk outer edge, Crossing ramp gradient, Kerb height); (2) ODbL review — recorded. The [D] Blosm-import (blockout cleaning) item stays deferred (tracked separately).
• Branch: odbl-review-0.1.0 (docs + TODO closure only).

2026-11-08 — placement release check; north-buildings camera was 9.5 km off

• Release check 2 (placement vs references): rendered demo, probed DemoMap world positions, compared to Geo references. Matches: road 24.3 m (ref crossings 20.9-25.9), sidewalk centreline offsets 12.5/13.2 m (ref 12.6/13.6), hero frontage 26.0 m / setback ~13 m (ref 25.75/14.5), six south-frontage buildings
  • TCL/Tussaud's in verified OSM positions.
• Bug found: the demo-north-buildings camera sat at X -22800 (the buildings' own X), aiming at its own position -> 85%-sky screenshot. Probed real geometry: TCL bbox X -22889..-18442, Y -9431..-1802; Tussaud's X -27071..-22799, Y -7586..-1888; hero at X -13336, Y +1307 (frontage 26 m). Camera now at (X -22750, Y +1600) looking north; sky fraction 85% -> 41%.
• Getting actor world bounds: spawn at loc(0,0,0) means static mesh bbox IS the world footprint (FBX carries the geo offset). Probe by loading the level explicitly (load_level) — a fresh commandlet starts on Untitled otherwise.

2026-11-08 — collision/seam audit done; QA self-test extended

• -qaselftest transect widened from 84 samples (3 lane centres) to 1456 samples / 16 rows including kerb/road/sidewalk seam Ys (probed: N side -1912..-1053, kerb -1382..-1053, road -1382..+1047, kerb +718..+1047, S side +718..+1447). 0 holes.
• New wall probe: ground-referenced horizontal 30 m trace at +160 cm. First attempt used a fixed Z 160 and started INSIDE the earth (street is at Z ~366) -> every row reported a dist-0 blocker. Always ground the probe first.
• Guard: GigLineShotActor must skip rounds under -qaselftest, else bQuitWhenDone closes the game before the transect phase (observed: game exited at t=2s with "ShotActor: done (4 shots)" + RequestExit).
• Results on packaged build: walk 761cm/2s, reset 0cm, pause/unpause, transect 0 holes, wall 0/35 -> all PASS. Evidence recorded in Docs/release-checks.md.

2026-11-08 — setup logs gated behind a cvar (diagnostics item)

• The remaining spawn/input setup logs (settle-timer "start captured", "GigLine IN:" pushed/bound/errors) are now gated behind gigline.SetupLogs (TAutoConsoleVariable, default 0 via DefaultEngine.ini [ConsoleVariables]). Re-enable at runtime: gigline.SetupLogs 1.
• The -qaselftest QA logs stay ungated (deliberate release-check feature per Docs/release-checks.md). Verified packaged build: 0 setup-log lines, QA suite still all PASS (walk/reset/pause/transect 0 holes/wall 0-35).
• Branch qa-diagnostics-0.1.0. SIM/GROUND probes were removed earlier in the DefaultKeyMappings fix (PR #8).

2026-11-08 — v0.1.0 released

• All six 0.1.0 release checks pass and are recorded in Docs/release-checks.md: traversal (seam transect 0 holes), placement vs references, collision+packaging (wall probe 0/35), 60 fps at 1440p (2.42 ms avg; TSR on, VSync off, no FG), fresh-checkout build, packaged Linux build with controls shown in-game.
• CHANGELOG.md gets the 0.1.0 entry; the completed 0.1.0 section was removed from TODO.md (goal stays in ROADMAP.md); tag v0.1.0 cut on main.
• Owner approved the regenerated screenshot set (PR #15) on 2026-11-08.
• THIRD_PARTY_NOTICES switched to checkbox statuses ([x]/[P]/[ ]); all tools incl. Blosm, QGIS and Git LFS are now verified as tool-only.
• Next work: the 0.1.1 collision-detail patch (sidewalk outer edge 15.8 cm step, crossing ramp gradient 10 %, kerb height assumption, traversal regression).

2026-11-08 — v0.1.1 released (collision patch)

• Sidewalk upstand now tapers full-at-kerb → zero at the building edge (removes the 15.8 cm outer ledge). Crossing ramp run derives from --max-ramp-grade (8 % ⇒ 1.63 m) instead of a fixed 1.5 m/8.7 %. Street + site rebuilt, site re-imported, DemoMap rebuilt, packaged build verified.
• Pre-check tool fixes: coverage no longer shrinks with --sample (it used range(int(block_len))), and it reports grade + location of the worst step.
• Residual, documented not hidden: 13.4 cm surface step at s≈198 (south) / 8.0 cm at s≈251 (north) — a strip discontinuity near a centreline vertex, below the 45 cm MaxStepHeight, so not blocking.
• Kerb height finalised at 0.15 m modelled / 0.13 m built (6 in LA standard; DEM and photos can't resolve finer).

2026-11-09 — Mapbox building data as a second height source

• Geo/tools/mapbox_buildings.py (new, stdlib urllib) Tilequeries mapbox.mapbox-streets-v8 layer building at every building centroid in the extract and merges the answer with our OSM height + source. Token comes from MAPBOX_ACCESS_TOKEN in the ignored .env; responses cache to Geo/raw/mapbox-cache/.
• Mapbox terms restrict persisting/redistributing their data, so .gitignore now covers /Geo/raw/mapbox-cache/ and /Geo/verification/*mapbox*. The tool also refuses to cache a response that contains the token.
• First run (72 buildings): 58 match, 5 close, 3 disagree, 6 non-building.
  • It independently classifies our roof/canopy/shelter volumes as roof/ shelter — corroborating the 0.1.2 finding that they are not buildings.
  • Remaining disagreements are untyped building=yes/apartments rear structures where our fallback guess (2 × 3.2 m = 6.4 m) meets Mapbox's 3.0 m.
  • Every "close" case is a building:levels × 3.2 m building where Mapbox is 0.6 m lower per 6 levels → Mapbox implies 3.1 m per level, not 3.2. Worth revisiting LEVEL_HEIGHT if a measured storey height ever lands.
  • Its query API surface: height, min_height, type, extrude per polygon.

2026-11-09 — Mapbox calibration adopted into the blockout

• Owner decision: adopt Mapbox. Applied:
  • LEVEL_HEIGHT 3.2 → 3.1 m in Geo/tools/block_height_report.py, and the Blosm --level-height default in Art/blender/blockout/blosm_blockout.py.
  • HEIGHT_OVERRIDES in block_height_report: ways 1490910971, 1492058875, 1492747608 → 3.0 m (Mapbox single-storey; the 2-level guess gave 6.4 m).
  • classify_height(tags, level_height, way_id) now takes the way id so overrides apply; mapbox_buildings.py passes it, so the cross-check agrees with the height table.
• Re-running the cross-check: 66 match, 6 non-building, 0 close/disagree.
• Geometry: blockout regenerated at 3.1 (object names preserved — the frontage exporter maps element.020/element/element.019/named buildings, all still present, 6/6 exported), frontage re-imported, DemoMap rebuilt, packaged build -qaselftest all PASS.
• Blosm adds roof mass on top of levels × levelHeight: 6922 39.1 m vs table 37.2, El Capitan Theatre 20.5 m vs 18.6 — both ~+1.9 m. That is a roof, not a height error; the table is the tagged height, the mesh is storey + roof.
• Gotcha: regenerating the blockout is safe for names, but any future change to the import must re-verify the 6/6 frontage export and the element.* mapping.

2026-11-09 — hero colonnade proportion fixed (47.6 % → 53.1 %)

• The hero's colonnade (columns 2.00–7.95 + capitals 7.95–8.90) was 6.90 m of the 14.50 m elevation = 47.6 %, below the photographed 50–60 %, because the entablature (3.40 m) and parapet (2.20 m) used 5.60 m of the height.
• Thinned ENTABLATURE_H 3.40 → 3.00 and PARAPET_H 2.20 → 1.80 in Tools/blender/hero_facade_build.py; the colonnade grows to 7.70 m = 53.1 % while the envelope, motto band and medallions keep their relationships (all positioned relative to those two constants).
• Gotcha (cost a bad build): hero_facade_build.py takes --out for the FBX and --save-blend for the .blend. Passing --out <file>.blend writes FBX bytes over the blend and Blender then reports "File format is not supported". Correct invocation: --out /tmp/gigline/hero-facade.fbx --save-blend Art/...blend --report Art/...json.
• Rebuilt hero → re-imported (SM_HeroFacade, bounds 2604 × 1257 × 1450 cm) → DemoMap rebuilt → screenshots re-rendered. Note ue_hero_facade_import.py saves the map, so chmod u+w the .umap first or the import silently does not persist.

2026-11-09 — AI detail plates from reference photos; floating windows fixed

• Tools/gen/facade_detail_concepts.py generates architectural detail plates from one local reference photograph via OpenRouter image models (default google/gemini-2.5-flash-image; the sibling project uses the same route in tools/generate_textures.py — chat/completions with a text + image_url content part, image returned as message.images[].image_url.url). Eight studies, --list free, --all generates, -vN never overwrites, .refgen.json sidecar per image, all output under gitignored Art/reference/facade-details/.
• Prompt gotcha: the model refuses when the prompt describes features the reference photograph does not show ("the image you sent appears to be a decorative architectural detail, not an entrance with steps"). Prompts must describe the reference subject and only restage light/people/surface.
• Output is real: 1184×864 to 1344×768 PNGs, ~2 MB each.
• Measured defect found and fixed: hero window panes sat at Y −1.98 while the mass front is Y −0.50 and the columns reach −1.02 — i.e. glazing floating 1.44 m in front of the wall and ahead of the colonnade. Now WINDOW_Y = −0.56 (6 cm proud of the wall, behind the columns). Visible as dark openings in review-hero-facade-front.
• Reminder: .blend is LFS read-only on checkout — chmod u+w or Blender fails with "Cannot save blend file … not writable" (this bit again here).

2026-11-09 — UE export hygiene: smoothing groups + facade collision

• All our FBX exporters used mesh_smooth_type="OFF", which is why every Unreal import logged "No smoothing group information was found for this mesh". The hero now exports "FACE" and its round parts are shaded with shade_smooth_by_angle(35°) before export: cylinders read as round instead of faceted, and the import warning is gone (6 → 0).
• The hero façade had no simple collision (complex-only), so line traces and capsule sweeps passed straight through it — the player could walk through the wall of the demo. Caught by a new -qaselftest facade occlusion probe (40 m sideways traces from the road centre at walking height).
• Auto-generated convex collision on a 95-part 26 m façade is unusable: it produced coarse hulls that bulged into the street and the wall probe went 0/35 → 2/35. Rejected in favour of UseComplexAsSimple via UGigLineAssetTools.set_complex_collision_as_simple (same reason the terrain uses it), saved on the asset.
• Probe expectations must mirror the staged buildings: the facade probe now carries explicit X spans per frontage (and the TCL/Tussaud's spans on the north side) so a gap between buildings is not reported as a defect.

2026-11-09 — one FBX export preset instead of six copies

• The bpy.ops.export_scene.fbx(...) argument block was duplicated in six scripts (hero, block frontage, north buildings, site, scale test, MPFB character), and all six carried mesh_smooth_type="OFF" — hence the "No smoothing group information was found" warning on every import.
• Now Tools/blender/fbx_export_preset.py holds ue_metric_v1 and ue_character_v1: preset_kwargs() (bpy-free, unit-tested) plus export_static() / export_character() wrappers that import bpy lazily. Every caller imports it and passes selection=True/False; the seven duplicated blocks are gone.
• Callers in Tools/ resolve the module with sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))); the Blender-side Art/blender/site/export_site_fbx.py computes the repo root (.., .., ..) and joins Tools. Getting that wrong silently inserts Tools/Tools — check the import path when adding a caller.
• Verified end to end: scale test, hero, frontage (6/6), site (7/7) and north buildings all re-exported through the preset, re-imported with 0 smoothing warnings, DemoMap rebuilt and -qaselftest all PASS (walk, transect 0 holes, wall probe 0/35, facade probe 0/9).
• mpfb_character_export.py was not re-run: it needs the MPFB add-on on BLENDER_USER_SCRIPTS ("Add-on not loaded: bl_ext.blender_org.mpfb"), which is pre-existing and unrelated; the character is not in 0.1.x gameplay.

2026-11-09 — v0.1.2 released

• Cornice finished: projecting corona over a 52-dentil course (0.18 × 0.30 × 0.22 m at 0.45 m centres, z 12.12–12.34, 0.25 m clear of the corona). Hero is now 3 044 tris across 148 parts, still 14.50 m. Verified: packaged -qaselftest all PASS (transect 0 holes, wall probe 0/35, facade probe 0/9).
• CHANGELOG.md gets the 0.1.2 entry (added/changed/fixed/verified/limitations), the 0.1.2 section is removed from TODO.md, ROADMAP.md status updated, tag v0.1.2 cut after the merge.
• Assumptions recorded rather than hidden: dentil spacing, base and abacus dimensions are not resolvable in our reference photography.

2026-11-09 — elevation cameras levelled (straight verticals)

• archviz MUST: an elevation must not lean. hero-facade-front looked up ~4° from 4.2 m at a 7.6 m target; demo-spawn-view and demo-street-level were pitched too. All three now put the camera at the target's height, so FindLookAtRotation yields pitch 0 and roll 0 and verticals are parallel by construction.
• Do not try to measure verticality from the rendered PNG with a naive edge scan: banners, windows and the parapet produce different edge sets at different heights and the pairwise match is meaningless (it reported a 418 px "drift"). Verify it geometrically in the camera list instead.

2026-11-09 — 0.1.3 packaging/setup fixes

• Tools/unreal/ue_common.py: make_writable() clears the read-only bit an LFS .umap checks out with (otherwise save_map() returns False with no error and the staged actors vanish), and require_saved() makes that failure visible. ue_hero_facade_import.py and ue_import_north_buildings.py never did this — which is why importing the hero appeared to succeed while saving nothing. Verified by forcing chmod u-w on DemoMap first: the run logged map writable: chmod u+w … (was 0o444), then save_map: True, exit 0.
• Gotcha while adding it: ue_hero_facade_import.py had no import os/sys (only unreal), so the sys.path line raised NameError: name 'sys' is not defined inside the commandlet — an exit code of 255 with no obvious cause.
• Docs: build.md gained the current pipeline table (7 site FBXs, frontage, north, hero, map), the Builds/GigLineQA archive, the packaged-log path, -qaselftest usage and a traps table; setup.md marks the fresh-checkout build verified and documents the two optional API keys; architecture 07 + combined doc corrected the build-output path.

2026-11-09 — 0.1.4 materials + architecture-camera pass

• Material graphs: ensure_material() now supports masonry coursing (courses_m) and roughness variation. UE's WorldPosition is in CENTIMETRES, so a course height in metres must be ×100 — the first attempt divided by 0.55 and produced coursing 100× too fine to see.
• connect_material_expressions returns False instead of raising, so an unknown pin name leaves an input unconnected and the material silently compiles with a default (a mask of 0 is invisible). All links now go through a link() helper that logs failures. There is also no MaterialExpressionLess; use Frac → Divide → OneMinus → Clamp for a soft threshold band, or MaterialExpressionStep/SmoothStep if a hard one is wanted.
• Material assignment was slot-0-only, which on the hero is the banner — the stone, shadowed stone and glazing kept their flat Blender materials. Assignment is now per slot, keyed on the imported slot material's name (glass/banner/stoneshadow/stone/roof/wall). Hero: 4/4 slots.
• New materials: M_FacadeStoneShadow (same coursing, a stop darker), M_FacadeGlass (roughness 0.14, specular 0.62 ≈ IOR 1.5 expressed without a refraction pass), M_FacadeBanner (the one saturated surface).
• Camera FOV: levelling the elevation cameras left the building small in frame (39 % sky). Added optional per-shot ShotFieldOfView to AGigLineShotActor and a 4th element on the shot tuples; a long lens (42–55°) fills the frame from across the street while the camera stays level, so verticals stay parallel. review-hero-facade-front sky 39 → 20 %, demo-spawn-view 23 → 16 %.

2026-11-09 — facade texture maps (0.1.4)

• Tools/gen/generate_facade_textures.py writes deterministic 512² base-colour and roughness maps (ashlar stone, stucco, asphalt, concrete) with a pure-zlib PNG encoder and hash-based value noise — no numpy/Pillow, ~14 s for all eight, and byte-identical on a re-run (checked by md5sum).
• ue_build_demo_map.py imports them to /Game/Hollywood/Textures and samples them world-aligned: WorldPosition (cm) → ×(1/tile_cm) → ComponentMask(R,G) → TextureSample UVs. tile_cm per material: stone 440, stucco 260, asphalt 900, concrete 250.
• Gotchas: MaterialExpressionTextureSample's UV input pin is "UVs", and the output pin for colour is "RGB" / roughness "R". ComponentMask pins are R/G/B/A booleans. A textured material skips the graph coursing on purpose — overlaying both would double the course pattern.
• The script needs ROOT in ue_build_demo_map.py and sys imported; forgetting either fails with NameError and an exit code of 0 from the commandlet (always read the result file under /tmp/gigline).

2026-11-09 — façade module kit (0.1.4)

• Tools/blender/facade_kit_build.py builds the kit at the measured 4.96 m bay and 3.10 m storey: pilaster, sill, lintel, window bay, shopfront bay, door bay, cornice, parapet + coping = 16 meshes, exported per-object with fbx_export_preset plus a JSON part report (dims + evidence per piece).
• Tools/unreal/ue_import_facade_kit.py → /Game/Hollywood/Kit/*, complex-as-simple collision (a convex hull would fence off the shopfront glazing).
• Tools/unreal/ue_apply_facade_kit.py lays the kit out from each building's own bbox: bays = round(width / 496 cm), storeys from height minus the 310 cm ground floor. 638 pieces over six frontage buildings + both north buildings.
• Bug worth remembering: the first window piece was modelled entirely at +Y (into the wall), so every window sat inside the solid blockout wall and was invisible in every render while the placement maths looked right. The kit now projects the surround 3 cm proud with the glazing 6 cm behind its face — a solid-wall blockout cannot carry a real recess, so the piece has to project.
• Blender gotcha: clearing the scene including materials invalidates handles already taken ("StructRNA of type Material has been removed"); clear objects and orphan meshes only, and create the material after the clear.
• The kit meshes are spawned at the world origin with their position baked into the mesh (like the frontage exports), so placement uses the bbox directly and must not add the actor location.

2026-11-09 — CRITICAL: blockout must be regenerated with --origin

• Regenerating the blockout in 0.1.2 (for the 3.1 m storey height) was run without --origin, so Blosm centred on the extract instead of fitting the project frame. The whole frontage shifted about 119 m in X and 2.4 m in Y, and the hero ended up outside the 6902/6838 gap — the "verified placement" was undone silently, and every render still looked like a plausible block.
• Correct invocation (README documents it, and the affine fit prints frame correction residual: max 0.0 cm): blosm_blockout.py --osm … --bounds 34.1006,-118.3422,34.1020,-118.3370 --level-height 3.1 --origin 34.101545,-118.338718 --assets-dir … --out …
• After re-export + re-import, the frontage spans are back to the verified values: 6922 −24276..−17700, 6902 −17749..−14591, hero −14644..−12028 between them, theatre −12082..−8210, and the north buildings at −22889..−18442 / −27071..−22799.
• ue_import_north_buildings.py also lacked import sys, so its sys.path line raised NameError and the import silently did nothing (exit 255). Same trap as the hero importer: these scripts imported only unreal.
• Lesson: after any blockout regeneration, probe the staged buildings' world X spans and compare them with Geo/verification/2026-11-06-el-capitan-block.md before trusting a render. A shift of this size is invisible in a screenshot.

2026-11-09 — kit naming + topology validation (blender-modeler checklist)

• Kit assets now follow the Unreal convention: the Blender object keeps its readable name (Kit_Pilaster), the exported asset and file are SM_Kit_Pilaster. The kit loader strips the SM_ prefix so the layout code still addresses pieces by short name.
• facade_kit_build.py::validate() runs the blender-modeler cleanup checklist on every piece before handoff — non-manifold edges, loose vertices, triangle count — and records the result in the part report. Current kit: 16 pieces, all manifold, no loose geometry, 12 tris each.
• ue_import_facade_kit.py now deletes kit assets that the current build no longer produces, so a rename cannot leave two generations of the same piece in the folder (the rename from Kit_* to SM_Kit_* did exactly that until it did).

2026-11-09 — visual comparison loop; screenshots are demo-only

• Tools/review/visual_loop.py closes the reference → render → feedback loop. --verify gates captures against Tools/review/visual_targets.json (blank, tone-less, wrong-size fail); --compare measures luma, RMS contrast, edge density, palette distance and a 3x3 luma grid against the reference and prints a 0–1 score; --iteration N writes Art/reference/iterations/N/{metrics.json,feedback.md}. The feedback names the skill to consult, so the next modelling pass starts from a measurement rather than an opinion.
• Iterations are gitignored (/Art/reference/iterations/); only confirmed shots are promoted into Screenshots/. A test asserts the ignore rule and that every target is a demo-map shot.
• Screenshots/ now holds demo-map captures only — the review staging map is no longer rendered for review. render_screenshots.sh --verify renders the demo map, collects the shots and runs the verification gate in one step.
• PIL gotchas: FIND_EDGES leaves its own border ring at 128 (236/3600 px on a 60x60 frame = 6.5% "detail" on a perfectly flat image), so crop after filtering — cropping before does not help, the new border gets the same treatment. Palette shares must divide by the full-image area, not the cropped one.
• First comparison (iteration 2): spawn view 0.70, street level 0.35, crossing 0.41, north buildings 0.34. Edge density is the dominant gap everywhere (0.04–0.11 against 0.20–0.32 in the references) — the render still lacks the relief the photographs carry, which is exactly what the 0.1.4 façade work is for.

2026-11-09 — theatre marquee + blade sign

• Tools/blender/theatre_trim_build.py builds the El Capitan Theatre's signature features in world coordinates: marquee canopy 12.0 × 2.60 m at a 4.60 m soffit, illuminated fascia, recessed soffit, vertical blade sign 1.30 × 5.20 m with frame and bracket. Tools/unreal/ue_stage_theatre_trim.py imports to /Game/Hollywood/Theatre and stages one actor per piece (single spawn at the origin — no placement arithmetic).
• Two placement bugs, both invisible in a render:
  1. Axis sign. Blender +Y is Unreal −Y, so a face at Unreal Y +1001 is Blender Y −10.01. Using +10.01 put the whole marquee on the far side of the boulevard.
  2. Absolute Z. Trim heights were absolute while the pavement at the theatre is 4.56 m above the project datum, so the canopy hung 5 cm above the pavement — the QA wall probe caught it as a knee-height blocker. GROUND_Z now offsets every height.
• Wall probe: a scan row at Y 1000 sat ~1 cm from the frontage building line and grazed the theatre wall once the kit's proud shopfronts went in. Rows now sample the street envelope only (−1400, −167, 800, −1382, 718).

2026-11-09 — 0.1.4 signed off; 0.1.5 created for the carried geometry

• Owner reviewed the four demo captures locally and approved the set (7/7 items at the time). The signoff lived in Docs/reviews/0.1.4-review.md, which was removed on 2026-11-09; the per-shot Approved column of Screenshots/README.md is now the only record and is authoritative.
• The three geometry items that 0.1.4 did not finish — retail neighbours from their own references, hero fluting/motto band, blockout cleaning — move to a new 0.1.5 Façade finishing patch rather than growing this one, plus a much-needed 1440p performance re-measure (625 kit actors have never been measured; 0.1.0 measured 414 fps).
• Decision recorded because it sets a precedent: finish and sign off a version when its stated scope is done, and carry leftover scope into the next patch — do not let a patch accumulate unmeasured geometry.

2026-11-09 — quality priority flipped: realism first, performance last

• Owner decision: realism leads; performance closes the quality list. Applied in Docs/Architecture/01_introduction_and_goals.md, 10_quality_requirements.md, 04_solution_strategy.md, Docs/release-checks.md (performance section now last, with the reason), ROADMAP.md (a "Quality priority" section) and AGENTS.md (the tie-break rule next to the scope-cutting rule).
• Ordering used: 1 visual realism, 2 geographic fidelity, 3 walkability/usability, 4 build reproducibility/portability, 5 provenance, 6 performance.
• Both revisions carry the same guard so the order cannot be used as an excuse: performance is a floor, not a preference — 60 FPS at 1440p without frame generation must be met at every release, and if realism cannot fit, the realism scope is cut deliberately and recorded in CHANGELOG.md, never the frame rate. Walkability outranks realism in one direction: a block the player falls through is not a demo.

2026-11-09 — 0.1.5 start: retail reference + kit layout maths

• Commons holds exactly one usable reference for the retail run: Art/reference/retail-neighbours/01-lee-drug-and-neighbor.jpg (Gb321, CC BY-SA 4.0, 1920×953) — 6800 Lee Drug and its neighbour at street level, shopfronts and parapets visible (vision pre-check: yes). Nothing for 6822 or 6806–6820, so those two façades still have no reference and stay blocked.
• Kit layout was wrong for tall/wide buildings, not just ugly: bays were round(width / 496), so 6922 (65.8 m wide, 37.1 m tall) got 13 bays — a window every 4.96 m across the whole frontage — and the storey count came out at 11 × 310 cm = 34.1 m against a 35.7 m window zone, leaving bare wall under the cornice. Fixed: MAX_BAYS = 9 (module widens instead of subdividing), and the storey spread is derived from the available height (usable / storeys) so the top band always lands inside the building. Result: 6922 9 × 11, theatre 8 × 6, 6806 7 × 3, 6800 4 × 3, TCL 9 × 5.

2026-11-09 — hero fluting + motto band; and a silent placement divergence

• Fluting: 20 grooves per column as thin shadow-tone cylinders set into the shaft (120 pieces) rather than boolean cuts — the eye reads them in silhouette and grazing light, booleans would cost far more for that. Motto band is now a real recess: surround 6 cm proud, panel 12 cm behind (was an 8 cm dark stripe). Hero 148 parts/3 044 tris -> 271/5 492, still 14.50 m.
• The real find: ue_hero_facade_import.py had its own hero position (-12024.5, -1450, 0) — the pre-0.1.0 value, before the frontage placement fix moved the hero to (-13336, +1307). Running the importer on its own therefore teleported the hero 13 m east and onto the north side of the boulevard, and nothing complained: a displaced building still renders. The only symptom was the reference score dropping 0.69 -> 0.56 with the 3x3 composition correlation flipping +0.54 -> -0.08.
• Fix: Tools/unreal/ue_placement.py is now the single source of truth (hero position, verified frontage spans, span_error/describe_span_failure). Both ue_hero_facade_import.py and ue_build_demo_map.py import it, the import asserts the staged span and logs PLACEMENT ERROR when it does not match, and tests assert the stale literal cannot come back. Spawn-view score back to 0.72 (best yet).
• Lesson to keep: duplicated placement constants drift silently. The metric that caught it was the composition correlation, not a visual review.

2026-11-09 — 1440p re-measured with the kit: 363 fps, ~12% cost

• 2560×1440, Epic preset, TSR, VSync off, no frame generation: 2.76 ms average (363 fps), p50 2.73, p95 3.17, p99 3.64, max 4.62 — against 0.1.0's 2.42 ms (414 fps). The 625 kit actors + 8 generated textures cost about 12 % and leave ~6× the 60 fps floor.
• Variance trap confirmed a third time: runs reporting ~105 fps (9.5 ms) were two overlapping game instances — an earlier launch had not exited. pgrep -af Binaries/Linux/GigLine should show one wrapper + one binary; 363 fps reproduced twice on a clean run. Always count instances before believing a perf number.
• Series committed at Docs/perf/0.1.5-kit-walk-frametimes.csv; comparison table in Docs/release-checks.md.

2026-11-09 — 604 unmaterialed actors: material rebuild was delete+recreate

• The "blockout cleaning pass" turned out to need material and placement hygiene, not more geometry. Tools/unreal/ue_audit_scene.py (new) found 604 of 620 mesh actors with no material: kit pieces and theatre trim all untextured.
• Cause: ensure_material in ue_build_demo_map.py deleted and recreated each /Game/Hollywood/Site/M_* asset every run ("so a look change cannot be masked by a stale asset"). That orphane every reference: actors staged by other tools pointed at deleted assets and fell back to the default material. The map looked only slightly flat — no error, no warning.
• Fix: update the material in place — load it, MaterialEditingLibrary. delete_all_material_expressions(mat), rebuild the graph, recompile, save. Note Material.expressions is protected from Python; use the library's deleter.
• Ordering rule now documented and enforced: the demo builder rebuilds materials, so kit and theatre staging must run after it. render_screenshots.sh now runs demo map → kit assembly → theatre trim → scene audit → shot cameras → capture, and fails on a non-zero audit count.
• Second find: 53 duplicated kit labels (both shop piers of a bay shared one name).
• Scene audit after the fixes: 0 failures (all materialed, all spans verified, no duplicate labels). Packaged QA still PASS; spawn-view reference score 0.72.

2026-11-09 — v0.1.5 released

• All actionable 0.1.5 work is done: hero fluting + recessed motto band, kit layout for wide/tall buildings, placement single-source-of-truth, 1440p re-measure (363 fps), and the material/placement hygiene that the "blockout cleaning" task actually needed (604 unmaterialed actors, 53 duplicate labels) with a new scene audit gate.
• Only item carried forward is the retail frontage (6822 / 6806–6820 / 6800), blocked because Commons holds no usable reference for two of the three buildings and substituting the theatre's photographs would misattribute the façades. It now lives in a 0.1.6 Retail frontage patch.
• Precedent applied again: tag when the stated scope is done, carry blocked scope into the next patch rather than letting a release wait indefinitely on missing reference photography.

2026-11-09 — Retail frontage and two pipeline traps

• Tools/review/render_screenshots.sh called ./Engine/... for the capture step while every commandlet above went through run_cmdlet's cd "$ENGINE". With set -e the script died at exit 127 after rebuilding the maps, so the PNGs silently stayed a day old and --verify never ran. Fixed to "$ENGINE/Engine/Binaries/...". Symptom to remember: maps rebuild, screenshots do not move.
• The generic façade kit is applied to every actor labelled Frontage_* or North_*, so staging a hand-modelled façade under those labels gets the kit pasted over it — two contradicting façade languages on one wall. ue_apply_facade_kit.py now skips MODELLED labels, and retail trim stages as RetailTrim_*.
• A Blender builder cannot be imported by a unit test (import bpy at module scope). Pure layout arithmetic lives in Tools/_common.py; Tools/tests/test_retail_frontage.py tests it there, and the builder delegates. Same trick fbx_export_preset.py uses for its preset dicts.
• Joining 85 parts into one mesh per building before export cut the Unreal import from 85 assets to 2, and the staging step to two spawns. Record per-part dimensions before the join: after it only the joined bounds exist.

2026-11-09 — Third-person body renders in bind pose (open defect)

• Symptom: the pawn's body is visible and skinned, the run logs GigLine: playing body clip Character_AnimHuman_rig_GigLine_Idle, and the figure still stands in the A-pose bind. Two captures were taken before the call was proven to succeed — the camera, then the material, were blamed first and neither was the cause.
• The FBX is not the problem. Re-importing /tmp/gigline/character-anim.fbx into Blender gives two actions, GigLine_Idle (frames 1–60) and GigLine_Walk (1–30), each with 1379 f-curves over 137 bones. Animation is baked and exported correctly, so the defect is on the Unreal side.
• The material was a real but separate defect: M_Character_Export had no SkeletalMesh usage flag, so Unreal substituted the default grey with only a LogMaterial warning. Fixed in place (used_with_skeletal_mesh = True).
• PlayAnimation alone is not enough. USkeletalMeshComponent::PlayAnimation needs the component already in AnimationSingleNode mode; the body is hidden at construction (SetHiddenInGame(true)) and only unhidden by the camera toggle, so the first call can land on a component that is not yet in the right mode and never retries — the guard Wanted == PlayingAnimation then returns forever and the pose stays at bind.
• What to log next time: animation mode and GetSingleNodeInstance() together with the clip name. A clip name alone shows the intent, not the effect.
• Fixed 2026-11-09, two independent causes. (1) PlayAnimation needs AnimationSingleNode plus a live single-node instance; the body is hidden at construction and only unhidden by the camera toggle, so the first call could land before the component was ready and the Wanted == PlayingAnimation guard then returned forever. UpdateBodyAnimation now sets the mode, calls SetAnimation + Play, and re-asserts while GetSingleNodeInstance() is null. (2) The arms-down offsets lived only in the clip keys — the exporter writes clip data relative to the rest pose, so they arrived in Unreal as a ±20° swing around an A-posed bind. apply_base_pose now bakes them into the rest pose before any clip is authored, which is what mpfb_character_export.py --pose idle already did.
• The base pose needs inverted signs when baked to rest. Applied as written, the FBX axis mapping produced arms raised in a V; negating the table gives arms down. Both states were verified on the imported asset, not assumed.
• Unreal bone names use underscores: upperarm01_L, not Blender's upperarm01.L. GetBoneIndex(TEXT("upperarm01.L")) returns INDEX_NONE, GetBoneTransform then returns identity, and a probe reading it reports a constant 0.0000 — which looks like "animation not applied" and is actually "bone not found". Check AnimPose.get_bone_names() before trusting any bone probe.
• Log verbosity does not reach the capture log, Warning does. Two diagnostic runs produced zero lines because the messages were UE_LOG(LogTemp, Log, ...); the same lines at Warning appeared. Instrument with Warning when the evidence has to survive into /tmp/gigline/render-*.log.

2026-11-09 — Unreal silently refuses to replace an imported asset

• The real reason the character pose never changed across a dozen rebuilds: the FBX importer reported success every time, but Character_Anim and its two AnimSequences kept their original data. Proof: after changing the arm base from 1.55 rad to 0.55 rad and rebuilding, AnimPose.get_bone_pose("upperarm01_L") returned the identical quaternion (+0.006, +0.002, -0.179, -0.984) as before. An asset that does not change when its source does is not being written.
• AssetImportTask.replace_existing = True was set, and EditorAssetLibrary.delete_asset returned False for both clips, even after chmod u+w (the .uasset files are LFS-lockable and check out read-only). Deleting is blocked because the pawn's C++ constructor hard-references the clips.
• Workaround that works: import under a new name. Tools/unreal/ue_character_anim_import.py now uses PREFIX = "CharacterV2", which creates CharacterV2, CharacterV2_Skeleton and CharacterV2Human_rig_GigLine_{Idle,Walk} fresh; the same quaternion probe then returned (+0.439, +0.898), i.e. the 0.55 rad arm rotation had finally landed.
• Check the asset, not the importer. A pose or material change that does not alter the probed asset data did not happen, whatever the log says. Probing a clip's bone quaternion takes one commandlet and settles it in seconds.
• The arms-down rotation belongs in the clips, not in the bind. Every bind experiment (bake pose to rest, clear shape keys, apply the armature modifier, re-bind) was either impossible on this shape-keyed MPFB mesh or left Unreal showing the old bind. Tools/blender/character_anim_build.py now carries a measured base offset in the clip keys (BASE: 0.55 rad on upperarm01, chosen by sweeping the rotation in Blender and reading the bone direction), and the character stands with its arms at its sides.

2026-11-09 — Textures: ComfyUI works, Stable Diffusion fails for flat swatches

• ComfyUI is reachable at 127.0.0.1:8188 (0.44.0) and its API is the clean way in: POST /prompt with an API-format graph ([node_id, output_index] references — the UI's workflow JSON cannot be posted), then poll /history/<id> and fetch /view?filename=.... Tools/gen/comfyui_generate.py does this and writes a sidecar with model, prompt, seed and the full graph. Installed checkpoints: juggernautXL_ragnarok.safetensors, epicrealismXL_pureFix.safetensors, sd-v1-5-inpainting.ckpt, plus inpainting variants.
• Do not ask Stable Diffusion for a flat texture swatch. Three attempts at "seamless skin texture" produced an anatomical close-up twice and a leather-scale surface once. SDXL has no reliable notion of a material sample without anatomy, so every seed needs manual rejection — worse than noise we control. Procedural generation won for skin pores (Tools/gen/generate_character_textures.py, tileable octave noise) and a strand sheet for hair.
• Unreal collapses the second material slot on a skeletal-mesh import. A joined hair cap (Blender shows two slots, hair faces on index 1) arrives as a single slot named after the body material, so the hair renders with the skin material. unreal.SkeletalMesh has no set_material, so it cannot be re-slotted after import. A separate skeletal mesh attached as a second component is the way around it.

2026-11-09 — Hair as a separate mesh, and a material-editing crash

• Hair joined into the body mesh loses its material. Blender shows two slots (hair faces on index 1) and the FBX exports both, but Unreal's skeletal-mesh import arrives with one slot named after the body material, so the hair renders as skin. unreal.SkeletalMesh has no set_material, so it cannot be re-slotted afterwards. A separate skeletal mesh imported on its own keeps its slot and imports cleanly (CharacterHair → M_Character_Hair).
• A second skeletal component follows the body with SetLeaderPoseComponent(GetMesh()) — no second rig, no second animation, and it hides with the body in first person.
• unreal.MaterialEditingLibrary segfaults in -run=pythonscript when it clears and rebuilds a material graph. The tool ran its texture imports, then crashed the commandlet with no Python traceback. Graph editing needs an editor session, or a material created from scratch rather than edited in place.

2026-11-09 — Character blueprints and how to compare a render to one

• A SDXL character sheet is reliable: "full body front view, neutral pose, plain grey background, character sheet" gives a usable turnaround with fixed seeds. Four views (front/side/back/head) from one description is enough to evaluate against; the sheet and its JSON live in Art/characters/blueprints/.
• Segmenting the character out of a review shot does not work by colour. The sky band, the pale wall and the dark pavement are each larger and more saturated than the body, so every size-based rule picked a backdrop blob and the first report measured the whole frame. Tools/blender/character_match.py now takes an explicit --render-box; the proper fix is a background plate rendered without the character.
• The report has to separate the failures it measures. Aspect, fill, coverage and two colour distances are reported apart on purpose: an absolute similarity number between a photoreal sheet and a low-detail body would mean nothing. The first real comparison (aspect Δ1.24, mean ΔE 31.9) was dominated by the model being unclothed while the blueprint wears a tee and jeans.
• SetMaterial(0, ...) in the pawn constructor does not visibly land on the body mesh: M_Character_Skin is built and assigned, yet the render stays neutral grey. Either the slot does not exist yet at that point or the assignment is dropped; BeginPlay is the next thing to try.
• MPFB vertex groups are polluted by helper geometry. The body mesh weights its helper cubes to groups named Left, Right, HelperGeometry and JointCubes at 1.0, so a "strongest group" lookup returns a helper for most vertices - the first clothing build selected zero faces. Filter helper names out before asking which bone a vertex belongs to.
• Blender refuses to apply a modifier to a mesh with shape keys. The body carries ten; a garment copy must clear them (shape_key_remove(all=True)) before solidify can be applied.
• Edit-mode selection reverts to vertex selection. Setting polygon.select in object mode and then entering edit mode re-derives selection from the vertices, which were all still selected - deleting "the unselected faces" wiped the entire garment. Filter with bmesh.ops.delete(context="FACES") instead.
• ConstructorHelpers::FObjectFinder outside a constructor is a fatal error, not a warning: the game dies in Tick with "FObjectFinders can't be used outside of constructors". Use LoadObject<T>(nullptr, Path) for runtime loads and keep FObjectFinder to the constructor.
• A 1 cm solidify shell lets the body poke through. Garments built by duplicating body faces needed 2 cm of solidify plus an 8 mm fatten along their own normals before the cloth read as continuous at shoulders, hips and knees.
• Warm skin sampling as blue-grey was lighting, not material. The body stood in building shadow and was lit almost entirely by the blue Demo_SkyLight (intensity 2.6): face (99, 98, 100), arms (67, 74, 81). Raising the sun to pitch -74 moved the face to (131, 121, 117). A deliberate red test colour reaching the render first proved the material and the slot assignment were fine, which is what stopped a long hunt in the wrong place.
• The Unreal commandlet drops unknown command-line switches. A --unlit-test flag passed after -script never reached Python; an environment variable does get through (and even that proved unnecessary once the red test settled the question).

2026-11-09 — The sidewalk does not block, and why

• The staged Sidewalk_North mesh does not stop a trace or a capsule. It carries QueryAndPhysics, its bbox tops out at Z 380, and Tools/unreal/ue_import_site.py sets CTF_USE_SIMPLE_AND_COMPLEX with auto_generate_collision=True — yet both a downward trace and a falling capsule pass straight through to Road_Hollywood at 366.
• Likely cause: the auto-generated simple collision for a 279 m x 8.6 m x 3.8 m slab is degenerate. A convex hull over that aspect ratio either collapses or ends up far from the surface; the complex side is one-sided and the mesh may be wound the wrong way. bOneSidedGeometry no longer exists on BodySetup in 5.8, so the flag cannot be flipped from Python.
• Working stopgap: an invisible 279 x 8.6 x 0.08 m blocking cube with its top at Z 380 in Tools/unreal/ue_build_demo_map.py. Verified: top surface under player: Z=380 mesh=Cube, foot: actor Z 470, feet 382.
• The real fix belongs in the site export: split the sidewalk into segments of a few metres, or export explicit UCX_ boxes, so the collision is a set of sane hulls rather than one degenerate one.

2026-11-09 — Slicing the sidewalk did not fix its collision

• The 20 m slice hypothesis is dead. The site export now splits sidewalks and kerbs into ~20 m pieces (Art/blender/site/export_site_fbx.py, SEGMENTED, SEGMENT_M), so each has an honest per-slice bounding box. The slices still do not block: with the invisible stopgap retired, a downward trace goes straight to Road_Hollywood at 366 and the capsule lands there.
• So the slab's 279 m aspect ratio was not the cause of the missing collision — a 20 m slice of the same mesh fails identically. The mesh itself imports with collision that does not block, regardless of size.
• What does collide: Road_Hollywood (366) and the engine cube stopgap. What does not: every Sidewalk_* and Kerb_* slice.
• Next place to look: the sidewalk mesh's geometry - an open surface with no volume, inverted winding, or zero-thickness triangles - rather than the import flags or the size.

2026-11-09 — Fast probe loop, and the real heights

• -probes runs the ground measurements and quits: a measurement run went from minutes to 13.7 s. Without the switch the game keeps running for the capture pipeline.
• The sidewalks' normals are correct (2790 faces, all pointing up) and they have no loose edges, so winding was not the reason they fail to block.
• set_complex_collision_as_simple on sidewalks and kerbs did not make them block either - the terrain needs it, they do not respond to it.
• The heights at the player, corrected once more: the road is 366 and the sidewalk 380 in the shipped map; the 158 cm figure came from the site blend's own Z datum, which is ~2.2 m below the map's because the terrain was added after that blend was written. The float is therefore about 14 cm - the character stands on the road because the sidewalk does not collide - not the metre it looks like against the kerb.

2026-11-09 — The slices were nested, and the site surfaces disagree with each other

• Bug in the slice export: the far cut used clear_outer=True, which kept the wrong half, so every "slice" still spanned to the end of the street (-27950..-43) and its bbox topped out at the same far-away maximum. Fixed by making both cuts keep the middle (clear_inner=True with the normal pointing at the half to keep). The slices are now disjoint and their tops read 380, 340, 309, 284, 254, 231, 205, 169, 138, 104 cm along the street.
• The surfaces themselves disagree: at the player's X (-133 m) the sidewalk top is 169 cm while Road_Hollywood is 366 cm - the road sits ~2 m ABOVE the sidewalk. The player therefore lands on the road, and no sidewalk collision trick will change that, because the sidewalk is not where the road is.
• That is a site-data problem, not a placement or collision one: the road and the sidewalks were generated from different passes and their heights no longer agree along this stretch.
• Next: regenerate the road and sidewalks from one elevation pass in the site assembly so their surfaces meet at the kerb, then re-export and re-import (one export + one import + one 14 s probe).

2026-11-09 — No duplicates, so the road mesh itself is at 366 there

• Checked for two Demo_Road_Hollywood actors after several import rounds: none. One actor, at the origin, identity scale, mesh Z spanning -44..366 cm.
• So the 366 cm the trace reports at the player's XY is the road mesh's own local height there, not a stale duplicate and not a far-side hit. The sidewalk slice at the same X tops at 169 cm.
• In the source blend at the same street position both surfaces sit around 1.4-1.6 m and agree within a kerb height, so the ~2 m gap appears after the mesh leaves Blender - which points at the import path (unit handling, or the Y mirror noted in export_site_fbx.py) rather than the strip generation.

2026-11-09 — The road is FLAT at 366 while the sidewalks slope: that is the gap

• Ground profile along Hollywood Blvd, traced every 20 m at the player's Y (-probes -profile):

x=-14336 z=366 Road_Hollywood     x=-2336 z=366 Road_Hollywood
x=-12336 z=366 Road_Hollywood     x= -336 z=302 Road_Hollywood
x=-10336 z=366 Road_Hollywood     x= 1664 z=  5 Terrain_DEM

• The road is a flat slab at 366 cm across the whole demo block, while the sidewalk slices follow the DEM and descend 380 -> 104 cm. The gap is therefore in the street generator: the road is not sampling the elevation per row the way the sidewalks do, or it was written with a single Z.
• That also explains why every sidewalk fix failed - the street the player actually stands on is the flat road, and the sloped sidewalk is metres below it.
• Next: inspect how rows is built in Art/blender/street/street_mesh.py - the road's road_z reads dem_z(x, y) from the same rows as the sidewalk, so a flat result means dem_z is returning the same value for every row (origin offset or a lookup that fails and falls back), and the fix is in that sampler.

2026-11-09 — Source truth vs what the engine reports

Sampled the regenerated street blend at the player's X (x = -133.36 m):

Road_Hollywood   verts at y = -9.0..-9.1, z = 0.87..0.94 m
Sidewalk_North   verts at y = +11.9..+12.0, z = 1.51..1.59 m

So the road is the south band at ~0.9 m and the north sidewalk is at ~1.55 m, exactly a kerb above it. The source geometry is correct and consistent.

In the engine at the player's position (UE Y -1150 cm, i.e. the north sidewalk band) the trace reports Road_Hollywood at 366 cm - the road is neither in that Y band nor at that height in the source. So the mismatch is introduced on the way in: either the FBX export's Y mirror documented in export_site_fbx.py, or the importer's transform. Until that is settled, no screenshot reflects the source, and a visual signoff would be approving a defect.

2026-11-09 — Why the road appears 2 m above the sidewalk at the same XY

• street_mesh.py builds the road across SOUTH_GUTTER -10.5 .. NORTH_GUTTER +10.5 m and the north sidewalk +10.5 .. +15.8 m, so on the straight the two bands meet exactly at the gutter.
• The strips are generated along a curved centreline (strip() places vertices along each row's normal). Where the centreline bends, the fan-out pushes the road's outer edge past its nominal band: the imported road's Y spans -13.8 .. +10.5 m while the sidewalk occupies -17.3 .. -11.8 m, so they overlap by about 2 m near the west end.
• In that overlap the road surface is at 366 cm and the sidewalk at 158 cm - which is why a trace at the player's Y kept finding the road, and why the scene read as the road floating above the pavement.
• Fix belongs in the strip builder: clamp each row's outer offsets to the nominal gutter so the road cannot fan past the sidewalk, or shorten the road strip at the ends. Both are one-function changes in strip() / main(); the demo currently reads correctly because the player and car stand on the sidewalk.

2026-11-09 — The 2 m lid was a convex hull, not geometry

• A 279 m road strip imports with ONE auto-generated convex hull, and a convex hull's top is the strip's maximum height. A downward trace anywhere on the road therefore landed on that hull at 366 cm - a flat lid ~2 m above the visible surface, which is at about 152 cm there. The road's vertices were never wrong.
• Slicing the road fixed it exactly as it had for the sidewalks: after adding Road_Hollywood to SEGMENTED, the cross-section is continuous -

Terrain_DEM      154
Sidewalk_North_07 158
Road_Hollywood_07 158   (was 366)

• Two things had to change together: the road had to be exported in slices, and the map had to stop staging the single Road_Hollywood asset, which still sat on top of the slices and won every trace.
• General lesson: a long walkable mesh needs slicing in the source, because Unreal's simple collision is a convex hull of the whole mesh - the per-slice hull is the only way its collision follows its surface.

2026-11-10 — chmod u+w on the Blender files, and what re-rendering does not do

• Every .blend under Art/ checked out read-only; a Blender save then fails with RuntimeError: Error: Cannot save blend file ... is not writable, and because the export happens after the save in Tools/blender/hero_facade_build.py, the FBX is never written and the old one is reused silently. All 20 .blend/.blend1 files were made writable; find Art -name '*.blend' ! -perm -u+w now returns 0.
• Re-rendering does not pick up .blend changes. Tools/review/render_screenshots.sh renders the imported UE assets, so a rebuild is needed in between: Blender build → Tools/ue_*_import.py → render. Chmod alone cannot change a pixel; the hero captures are byte-identical before and after.

2026-11-10 — skills: one common copy, project-specific stays local

• Symlinked packs were the wrong fix. The 2026-11-06 fix put 94 symlinks in ~/.agents/skills/<pack> → the repo, while ~/.pi/agent/skills/blender-skills/ held a full second copy of the same packs. Pi keeps the first discovery and warns on every start: 94 "collision" lines. ~/.agents/skills/blender-skills/ (the upstream git checkout) also contains 94 packs under .claude/skills/ and .cursor/skills/ — inert, because dot-directories are not scanned.
• blender-director/SKILL.md never loaded. It declared role: twice in metadata (specialist then orchestrator); duplicate YAML keys are fatal, so the pack was dropped silently while the other 94 loaded.
• Layout now — shared, one level down and flat so it is discovered: ~/.pi/agent/skills/<93 packs>/, ~/.pi/agent/skills/references/, ~/.pi/agent/skills/LICENSE. Project-specific and versioned: .agents/skills/blender-skills/blender-director/, whose ../references/ links were rewritten to ~/.pi/agent/skills/references/… because it no longer sits beside them. The 94 ~/.agents/skills symlinks are gone.
• Verified by walking the three discovery roots (~/.pi/agent/skills, ~/.agents/skills, project .agents/skills, dot-dirs skipped), parsing every frontmatter: 97 skills loaded, 0 malformed, 0 collisions. Backups of the old trees: /tmp/agent-skills-backup.tgz, /tmp/pi-blender-skills-dupe (now consumed).
• Rule: a skill exists in one place. Two roots with the same name is a warning on every start, and the loser is invisible — check the startup list, not the file tree.

2026-11-10 — Tools restructured, and the four traps it set

The sixty scripts had grown into one flat directory; they are now grouped by what they talk to (and therefore by sandbox): Tools/unreal/ (commandlets, both map builders, ue_common.py, ue_placement.py), Tools/blender/ (builders, exporters, FBX preset, MPFB pipeline, blender_screenshots.py), Tools/review/ (render/judge/compare/e2e/visual_targets.json), Tools/gen/ (ComfyUI driver, procedural textures, reference fetching), Tools/input/ (OS input injection). Root keeps _common.py, check-lfs.sh, addon_compat_probe.py, sibling_asset_audit.py, tests/.

Four things broke, each with the same shape — a path that used to resolve and now does not:

• Sibling imports. import ue_placement / import ue_common failed inside every Unreal commandlet (ModuleNotFoundError), because the rewrite put only the tools root on sys.path while those modules are the script's siblings. Both paths are inserted: sys.path.insert(0, dirname(_HERE)) then sys.path.insert(0, _HERE).
• Shell scripts resolve REPO relative to themselves. Tools/review/render_screenshots.sh computed the root as one level up, so every commandlet path became Tools/Tools/... and the capture step died in 0 s. Any script whose own depth changes needs its REPO="$(cd "$(dirname ...)/.." && pwd)" re-checked.
• Stale result files pass as success. The e2e driver accepted /tmp/gigline/demo-map.txt left by an earlier run, so a 10-second "demo build" reported PASS without building. Result files must be unlinked before a step runs, and child output must stream to a per-step log file — a captured pipe hides fifteen minutes of a Unreal build and deadlocks if the engine outlives it.
• Two editors hang, they do not fail. Unreal does not refuse a second instance on the same project; it waits. The driver now refuses to start while any UnrealEditor-Cmd is alive. pkill -9 -f <name> in the same shell also matches the shell's own command line — kill in a separate invocation, or by PID.

2026-11-10 — LevelEditorSubsystem.new_level() cannot replace a map in one session

Measured on a throwaway path:

new_level fresh         -> True
delete_asset            -> True
disk file still present -> True     (delete_asset leaves the .umap)
new_level again         -> False

So "delete the map and recreate it" does not work in a commandlet: the package stays in memory and new_level refuses the path even after removing the file from disk. What works — and is what the builders actually need — is load, destroy every actor, save, and verify the count is 0. That is what Tools/unreal/ue_recreate_maps.py does now.

The related lesson: a prefix-based clear (clear(("Demo_", "PlayerStart", ...))) only removes what the builder knows about. Stale actors from experiments survive every rebuild — which is how a stale road mesh and a duplicated vehicle actor were both saved back into the map. Clearing the whole level turned the old 600-actor map into a 409-actor one that audits clean (0 failure(s), no duplicate labels): 72 site meshes, 310 facade-kit pieces, six frontage buildings, lights, cameras, player start. A falling actor count was the fix, not a regression.

2026-11-10 — Every image is judged by two models; the disagreement is the signal

Tools/review/judge_screenshot.py runs openai/gpt-4o-mini and qwen/qwen3.7-flash by default and reconciles them: agreement carries the verdict with the mean confidence, disagreement reports uncertain and keeps both reasons rather than letting the more confident model win. --model forces a single judgement.

• The second model had to be replaced once already: anthropic/claude-3.5-haiku returned 404, no longer served by any provider. Take the model list from https://openrouter.ai/api/v1/models rather than trusting a remembered id.
• On this content the judge is consistently pessimistic: it reports 0/11 across review shots the owner has approved, and it has been wrong in both directions (missed a real float, disputed correct frames). Treat it as an advisory pre-check, and treat a model disagreement as the thing worth reading.
• Tools/gen/comfyui_generate.py applies the same rule to generation: two checkpoints by default, same prompt and seed, scored by mean absolute Laplacian (catches blurry/flat output), winner takes --out, loser is kept beside it. The score ranks candidates; only the judge understands content.

2026-11-10 — Local ComfyUI, and the workflows worth using

• The server at http://127.0.0.1:8188 is a Stability Matrix install and nothing in this repository starts it. If a tool cannot connect, start Stability Matrix, launch ComfyUI, re-run. The install documents 141 GB of models in /home/zwitschi/StabilityMatrix/Data/README.md.
• The documented texture graphs are ComfyUI UI format, which /prompt cannot accept; a tool must rebuild the equivalent API graph. Tools/gen/comfyui_generate.py texture does that with their exact settings: albedobase (30 steps, CFG 6, dpmpp_2m/karras), krea2 (8 steps, CFG 1, euler/simple, ConditioningZeroOut, texture-fix VAE), qwen21 (25 steps, CFG 1, TextEncodeQwenImage21, texture-fix VAE).
• Two limits the graphs carry and a tool must not hide: they produce albedo only (normals/roughness come from a material tool downstream), and ComfyUI 0.44 core has no seamless sampler, so "tileable" in a prompt is a hint — offset the image 50 % and check the seam.
• Licences differ per workflow and matter before shipping: AlbedoBase allows commercial use with credit, Krea-2 has a community licence, Qwen-Image 2.1 is research-only. THIRD_PARTY_NOTICES.md carries the row.

2026-11-10 — Hero facade: what actually blocked it

• Detail modelled inside the wall. Every applied part sat behind the mass's front face (Hero_Banner_1 at Y −1.02 against a wall at −0.50), so nothing rendered but the columns and cornice. seat_on_front() now states the plane once (FRONT_Y) and seats each part at FRONT_Y + offset; it reports every part it moves.
• The hero's materials are not the kit's. M_FacadeStone in /Game/Hollywood/Site belongs to the facade kit; the hero uses /Game/Hollywood/ElCapitan/M_Facade_*, created by the FBX import. Warming the Site material changed nothing for three rebuilds.
• Unreal keeps a material asset on re-import, so a colour change in the Blender builder never arrived. Deleting the materials before the import broke the mesh's slots (one colour, no banners, no windows). What works: import normally, then author one plain material per slot with the reference colour and assign it.
• Save the material itself. Saving only the mesh left the new materials unsaved; after a reload the slots pointed at nothing and the whole elevation fell back to one default colour.
• unreal.Paths.project_dir() is Game/GigLine, not the repository root — textures under Art/ must be resolved from the script's own location.
• Still open: the generated ashlar albedo is imported and the material contains TextureSampleParameter2D → Multiply → Base Color with UV tiling, but the render still shows the flat tint. Next probe is the compiled Base Color pin, not the node list.

2026-11-10 — Render-to-import chain, in one order

A visual change needs build → import → render, never just a render: render_screenshots.sh renders the imported Unreal assets. Tools/review/e2e_asset_check.py runs the whole thing (Blender renders → empty maps → both builds → captures → two-model judge → README) and is the command to use instead of remembering four.

2026-11-10 — Hero attic: the medallions were buried, not just flat

• seat_on_front offsets are relative to FRONT_Y (the wall plane, Y −0.5), not to the parapet face. The parapet is 1.2 m deep with its front at Y −0.15, so any applied detail with offset < 0.35 sits inside the parapet and never renders. The medallions (offset 0.08 → front −0.42) and the motto band + surround (offset 0.06 → front −0.44) were all buried; only the inscription (moved to 0.40 → front −0.10 in the previous commit) read.
• Medallion fix: stepped relief with per-layer prefixes so seat_on_front gives each its own stand-off — MedallionRim 0.36, MedallionField 0.43, MedallionBoss 0.50. One shared prefix (Hero_Medallion) would seat every disc at the same offset and flatten the relief, and it also matched the more specific prefixes via startswith, so it had to be removed entirely.
• Medallions were also at ±6.2, inside the inscription text (spans ±6.87). Moved to ±9.8.
• Tools/review/hero_iterate.sh still had the pre-move cd "$(dirname "$0")/.." from the tools restructure — it landed in Tools/ and failed find Art with "no such file". Fixed to ../.., matching render_screenshots.sh.

2026-11-10 — Hero motto band + surround: same burial, plus dark-on-dark

• The motto band and moulded surround shared the medallion burial (offset 0.06 → front Y −0.44, inside the parapet). Fixed the same way: band recessed at 0.34 (Y −0.16), surround proud at 0.43 (Y −0.07), inscription stays 0.40 (Y −0.10).
• The band had to go light (stone, was stone_dark) at the same time. A dark band at the parapet face would have sat directly behind the dark inscription marks and swallowed them; the reference is dark incised lettering on a pale stone band.
• The seat_on_front table had a dead duplicate — Hero_MottoBand appeared at both 0.06 and 0.40, and the first match (0.06) won, so the 0.40 was silently ignored. That is why "the band is buried" looked fixed in the table while the build kept burying it.