.nornir/handover-2026-09-29-oden.md

Handover 2026-09-29 evening — dwarf's `claude` container → oden

nornir:class=design · written on dwarf at the move to oden. Continues handover-2026-09-29-navgraph-gpu.md (read that first for the owner orders; they all still hold).

Owner orders added today

Use up to 3 agents (owner, 2026-09-29 afternoon; the morning order was 2).
Merge to main when tests pass — CPU suite on the build box AND the GPU-only tests on a real GPU. A failure that also fails on main is main debt and does not block; a new one blocks.
Second wasm bundle: yes. docview, manual and model3d must not grow the default bundle (12.64 MB → 16.44 MB measured). Grouping: yes — groups of facets are declared in the build thing (facett-build-thing code): one wasm row per group, the routing table emitted from the recipe, never a hand list. Groups today: core (default, byte-identical to main's), docs (docview + manual), model3d.
The t14s (owner's laptop) is a GPU box for tests: ssh rickard@100.67.148.64, Radeon 890M (RADV STRIX1), glibc 2.43, Chrome. The owner works on its screen: headless Chrome only: --headless=new --enable-unsafe-webgpu --enable-gpu --ignore-gpu-blocklist --disable-software-rasterizer --enable-features=Vulkan --use-angle=vulkan --render-node-override=/dev/dri/renderD128 (without --enable-gpu/--use-angle=vulkan headless gets SwiftShader, which does not count; --use-gl=angle or --ozone-platform=headless crash the GPU process). Never compile there. One test binary at a time, --test-threads=1, VK_LOADER_DRIVERS_SELECT=*radeon* (hides llvmpipe), 30 s pause before heavy binaries, stop if Tctl > 95 °C, never all-pairs layout at 100k on it (it caused one amdgpu ring reset at 10:06:53). Clean ~/runs and /tmp there afterwards.
loki slept all day; the RTX 5080 rows are still owed (stage-int/RUNBOOK.txt on dwarf; the runbook text is at the end of .nornir/facett-benchmarks.md).

On main (`728c4fe`)

merge	what
`e03f2df`	lanes A + B of the morning handover: navgraph live read-back, osm2d adapter preflight + retry, GPU cloud lanes for graph2d/geomap, 0 undeclared atoms, 45 → 48 registry entries, GridLane parity, graphview flake fixed (`gputurn::hold`), graphnav GPU layout at any size, 100k benches (t14s numbers in `facett-benchmarks.md`), knut dependency restored
`daddd0a`	gallery_probe uses browser_bench's headless flags + software-adapter check (one copy)
`537ea35`	CPU main debt: road_rank motorway tier, map3d houses fixture, one MSAA source, graph3d mds asserts the fallback
`728c4fe`	xtask build is a thin door onto facett-build-thing (`.nornir/xtask-thin-door-audit.md`)

Open branches (all pushed; check `git ls-remote` for the final shas)

branch	state	next
`lane-b/main-debt-gpu`	3 GPU main-debt fixes: geomap osm_paint_order fixture, map3d item6 cull pad, ortho depth bug in `facett-core/src/render/view.rs` (plan view drew no ground). CPU suite on `check/main-debt-gpu` (= main 537ea35 + this) clean: 3562 passed, every failure "no GPU adapter". Plus `67a3dfb` grid snapshot tolerance (may be on `lane-b/grid-tolerance-unproven`). Per-binary GPU results: `.nornir/gpu-gate-2026-09-29.md` on that branch if the gate got that far.	Finish the GPU gate on the Radeon at `check/main-debt-gpu`, then merge together with `lane-a/ortho-depth-test`
`lane-a/ortho-depth-test`	`3b6ab20`: CPU test for the ortho fix; proven red on main's view.rs (depth exactly 1.0), green with the fix	merge with lane B's fix
`lane-a/last-seams` (+ maybe `lane-a/bundle-groups`)	docview + manual + model3d data seams (`1f9a52d`, `3bf5044`); feature gating for the bundle split in progress (WIP)	finish grouping (facet-wrapped side: entry → group, per-bundle manifest, loader routing from the recipe table, refusal naming the right bundle); gallery probe headless on the Radeon
`lane-c/bundle-groups`	recipe side of grouping (WIP)	`BUNDLE_GROUPS` in facett-build-thing, one row per group, routing table emitted, wheel carries all dists
`lane-c/container-road`	law: build in a container, never on the host (WIP)	`build_variant_in_container` + `DraupnirCargoRunner`, pinned toolchain digest `2cf9b012…`
`check/main-debt-gpu`, `integration/0929`	merge-check branches	delete once merged

Final shas at the move (verified with ls-remote)

branch	sha	note
`lane-a/last-seams`	`59b8767`	WIP. Core bundle byte-identical to main (sha256 `d1806046…`). Groups live in a NEW package `facet-wrapped-group` (`[lib] path` = facet-wrapped's `src/lib.rs`, lib name `facet_wrapped_group`, features `group-only`, `docs`, `model3d`) because ANY `[features]` table on facet-wrapped — even `docs = []` — changes 14 517 bytes of the release wasm. Group dists `facet-wrapped-group/dist-docs` (9.03 MB wasm) and `dist-model3d` (7.25 MB), files `facet_wrapped_group{.js,_bg.wasm}` + `manifest.json`. `src/groups.rs`: `GROUP_ENTRIES`, `group_of`, `THIS_BUNDLE`, `wrong_bundle`. Loader routing NOT done (index.html loads core only).
`lane-c/bundle-groups`	`e4d2c1a`	WIP, recipe side: `BundleGroup { name, variant, features, no_default_features, dist_dir, wheel_dir }`, `BUNDLE_GROUPS`, `routing_json()`/`write_routing()` → `pkg-facet-wrapped/src/facet_wrapped_web/assets/bundles.json`, rows `facet-wrapped-wasm` / `facet-wrapped-docs-wasm` / `facet-wrapped-model3d-wasm`. edda allows a package one `from_wasm` producer, so it is `None` now (edda needs a list).
`lane-c/container-road`	`74e5bd9`	WIP. `cargo xtask build` = `build_variant_in_container` + `DraupnirCargoRunner`, host road removed. Toolchain pin moved `2cf9b012…` → `3d5115ba…` (the old pin was never built; `3d5115ba…` = edda's and holger's pin, SOURCE_DATE_EPOCH 1783900800). facet-wrapped-wasm built in the container in 421 s: manifest identical to host, wasm differs (rustc 1.97.1 in image vs 1.98.1 host; 27 `/home/rickard` paths in the host wasm, 0 in the container's). wasm stages (bindgen, manifest) still on the host. Full `cargo test -p xtask` not re-run. dwarf: rootless podman 5.7.0.

| lane-b/main-debt-gpu | 56ee32f | 14b1c6e + WIP results file .nornir/gpu-gate-2026-09-29.md. GPU gate NOT run: 0 of 82 binaries (the copy to the t14s was stopped at the move). Staged only on dwarf (stage-gpugate/, 15 GB, run.sh with the heat rules) — rebuild on oden rather than copying. | | lane-b/grid-tolerance-unproven | 67a3dfb | assert_golden_within, facett-grid snapshot tolerance 2. Not proven on a GPU; look at the diff image (anti-aliasing only?) before fast-forwarding main-debt-gpu to it. |

Next on oden, in order: (1) the GPU gate at check/main-debt-gpu (+ ortho-depth-test 3b6ab20's facett-core lib): every GPU-only test binary on a real GPU, one at a time; a red that main 537ea35 also has is debt, a new red is bisected; then merge lane-b/main-debt-gpu + lane-a/ortho-depth-test. (2) Grouping: align lane C's rows to lane A's facet-wrapped-group package, loader routing from bundles.json, headless gallery probe on the GPU, merge. (3) Container road: full cargo test -p xtask, rebase over grouping, merge.

⚠ The two grouping halves do not match yet. Lane C's rows assume facet-wrapped declares features docs/model3d and writes facet-wrapped/dist-<g> with facet_wrapped* file names; lane A proved facet-wrapped cannot carry features and built a separate package facet-wrapped-group with facet_wrapped_group* files. The recipe must follow lane A's package (row = -p facet-wrapped-group --features <g>, dist facet-wrapped-group/dist-<g>, lib name facet_wrapped_group). lane-c/bundle-groups and lane-c/container-road both touch facett-build-thing/src/lib.rs, xtask/src/door.rs and the emitted toml: the second to merge needs a rebase.

Still owed, needs the owner

korp navgraph-overlay @ 76b7e4b, dwarves navgraph-overlay @ 0763014: not compiled — they need ../tunnr and ../gunnar, which codeberg nordisk/ answers "not found" for.
cargo xtask test-demo: needs a holger pip index on :8455; holger needs edda/crates/appliance-grow, which edda main does not have.
xtask still has 11 deps (law: one) — the extra ones serve verbs that build nothing; panel_wasm, edda's nornir_pages::demo_wasm and publish-pip still write their own argv (audit §5).

How it ran on dwarf (for the next box)

Fresh clones side by side under ~/runs/facett-0929/; one worktree + one CARGO_TARGET_DIR per lane; nice -n 10 cargo … -j 4. Worktrees get Git-LFS pointer stubs: fetch with inline -c filter.lfs.* settings (no config written); git status reads clean either way, and facett-docview fails to build on a stub font. Codeberg rejects pushes often ("Internal Server Connection Error" / "pre-receive hook declined"): retry and verify with git ls-remote, never by grepping push output (the rejection line also says main -> main). A shared target dir between worktrees gave byte-identical binaries for different trees once — use one target dir per tree.

Handover 2026-09-29 evening — dwarf's claude container → oden