# What a relocated assembly tree does in a browser

Status: recorded product evidence; closes the browser gate ADR-0017 left open.

[ADR-0017](../../../docs/adr/0017-relocated-hierarchy-sidecar.md) added
`--relocate-hierarchy-nodes`, which writes a federation's mesh-less nodes into
a package sidecar (`hierarchy.json` + `hierarchy.bin`,
`naru.package-hierarchy.1`) instead of into `scene.gltf`. The offline record
([`artifacts/compiler/hierarchy-relocation`](../../compiler/hierarchy-relocation/README.md))
measured the bytes: the document loses about 22%, the whole package only about
2.6%, because the tree is carried, not deleted. That left the question the ADR
stayed Proposed on — whether a smaller document is *faster to open* at
federation scale, and what the sidecar costs to fetch — for a browser to
answer.

This record answers it on the sixty5 federation (78,173 occurrences).

## Method

Two packages compiled from the same retained Scene IR split differ in one
thing: whether the mesh-less nodes live in the document or in the sidecar.
`scene.bin`, `coarse.bin`, `properties.json`, and `properties.bin` are
**byte-identical** between them — the validator checks each digest — so a
difference measured here is relocation and nothing else.

Each sample spawns the committed browser recorder
(`scripts/record-ifc-browser-evidence.mjs`) in a fresh process: fresh Vite,
fresh headed Chrome 151.0.7922.139, viewport 1320x1000. Three runs per arm,
interleaved, so host drift cannot accumulate against one arm. Medians are the
lower median of three, so every figure below is one of the runs.

## What it changes

| Median of 3 | In place | Relocated | Change |
|---|---:|---:|---:|
| Hierarchy ready | 2,313 ms | 2,290 ms | −23 ms (−0.99%) |
| **First coarse frame** | **4,702 ms** | **3,966 ms** | **−736 ms (−15.65%)** |
| Budget-limited ready | 9,953 ms | 9,834 ms | −119 ms (−1.20%) |
| Peak JS heap | 856,279,408 B | 697,327,573 B | −158,951,835 B (−18.56%) |

Per run (hierarchy / first frame / ready, ms):

- in place — 2,365 / 4,702 / 10,427 · 2,313 / 4,808 / 9,953 · 2,261 / 4,586 / 9,243
- relocated — 2,270 / 3,966 / 9,834 · 2,341 / 4,045 / 9,868 · 2,290 / 3,915 / 8,383

Both arms were re-captured on 2026-09-07, after the Studio camera fix and the
view cube, because the committed screenshots still showed the mirrored
from-below rendering. Every counter in *What it must not change* reproduced;
the wall-clock figures above are the new capture's, and the first-frame
conclusion is unchanged.

The two arms' spreads do not overlap on first frame or on heap. Hierarchy-ready
barely moves, which is the honest reading: the tree still has to be read before
the panel can list it, and reading it from a sidecar costs about what reading it
from the document did.

## What it costs

| | In place | Relocated | Change |
|---|---:|---:|---:|
| `scene.gltf` | 448,823,852 B | 347,731,160 B | −101,092,692 B (−22.52%) |
| Sidecar (`hierarchy.json` + `hierarchy.bin`) | — | 46,253,688 B | +46,253,688 B |
| Package | 657,116,508 B | 602,277,504 B | −54,839,004 B (−8.35%) |

The document delta plus the sidecar equals the package delta exactly; the
validator asserts that arithmetic, so no other resource moved under cover of
the comparison.

`hierarchy.bin` is fetched **once per run**, as response 3 of 118, ahead of
`coarse.bin` — the Studio reads the tree on the main thread while the Worker is
still decoding geometry, which is why a 46 MB extra fetch costs 23 ms rather
than a round trip. The in-place arm requests no sidecar at all. Both facts are
derived per run from the response stream, not asserted.

## What it must not change

Every run of both arms reaches the same endpoint over **17 counters**: status
text (`Residency budget reached · 24326 surface batches retained · 78173
renderable occurrences`), 42,435 prototypes, 78,173 occurrences, 2,255,235
triangles, 12 edges, 111 of 234 chunks resident, 66,686,508 decoded /
66,783,808 GPU bytes against the 67,108,864 B budget, 111 scheduler requests +
123 skips, 113 ranged `scene.bin` responses, the same picked occurrence
(`occurrence:ifc:facade-a9a1b20214da:52355`) and its 44 IFC2X3 property
entries. Zero console issues in all six runs.

One thing deliberately differs: the picked node **index** (74387 → 37247). A
relocated document keeps only the nodes that draw, so they are renumbered. The
occurrence, its identity, and its properties are unchanged.

## Host locality

The package digests (`a2d6c72a6e93` in place, `b821e4316a5b` relocated) are
**host-local**: this Windows host's IFC adapter emits a Scene IR split a few
bytes different from the macOS host's. The validator says so, and they must not
be retargeted to make a re-record pass. Timings are held as bounds, not pins.

## Reproducing

```
pnpm spatial:package --input output/ifc/sixty5 --output output/ifc/sixty5-relocated --no-spatial-index --relocate-hierarchy-nodes
pnpm hierarchy:browser:evidence
pnpm hierarchy:browser:check
```

The in-place arm is `output/ifc/sixty5-prb`, the package the committed
[`sixty5-first-frame`](../sixty5-first-frame/README.md) record uses. Six headed
browser runs take roughly thirteen minutes on this host.