Skip to content

Official Codebase of 3DCodeVerse

Python

placement: all five "missing planned contents" were the wrapper, not two of them

ziyao6f5acf18b035
diff --git a/harness/docs/DECISIONS.md b/harness/docs/DECISIONS.md
index b7c14aba..e560378b 100644
--- a/harness/docs/DECISIONS.md
+++ b/harness/docs/DECISIONS.md
@@ -910,6 +910,34 @@ written) that were accepted because the code works that way and the tests pin it
`medieval_cobblestone` first — against a judge complaint reading "no cobblestone
texture or material blending".)
+* **D59 A check that reads the scene graph by name walks the whole zone subtree
+ (2026-09-06).** All five `scene_placement` ERRORs left in the fixed arm were
+ "zone X is missing planned contents", and none of them was absent content. A zone
+ typically wraps what it builds in one group — floating_islands puts Windmill,
+ FloatingRock and SkyPine inside `IslandAssembly`; cozy_cabin's `PineSway_0` and
+ snowy_hut's `LanternPost1` are the ZONE's names for objects whose own roots, one level
+ further down, are `SnowyPineTree` and `HutPorchLantern`. The placement table lists
+ direct children of a zone, so the contract check saw one row per wrapper and called the
+ plan unmet. Each row now carries its named descendants (`inner`, capped at
+ `MAX_INNER_NAMES = 24`) and the check reads them; re-probing the three recorded
+ workspaces takes the contract from 5 ERRORs to 0. This is the third variant of one
+ defect — a glTF root called `Scene` (D56), a wrapper group, a behaviour-named parent —
+ so the rule is now: **a by-name check on the scene graph names a SUBTREE, never a
+ child.**
+
+* **D60 A harness failure is never handed to the agent as a repair (2026-09-06).**
+ `probes.probe_report` labels a driver that produced no output with
+ `harness_failure: True` and the hint "this is a harness/driver failure, not your code",
+ and `build_with_repair` did not read the flag. Measured on `bench/out/scene_textures`
+ (japanese_garden): three repairs against that message rewrote 5, then **14**, then 3
+ files — the 14 included `env.js` and every zone — and the fourth build passed on its
+ own. The 14-file rewrite deleted the texture use the arm existed to measure; that cell
+ scored 0.496 where the same prompt scored 0.636 without the detour. `BuildResult`
+ carries the flag now and the loop re-runs the BUILD instead, bounded by
+ `MAX_HARNESS_REBUILDS = 2` and costing no model call, since the node driver has already
+ retried once itself. Rare (1 cell of 12 in each of two arms) and expensive when it
+ fires.
+
## Rejected / deferred
* A versioned `Spec`/`RunRecord`/`RunState` load-normaliser (rejected 2026-08-30: of the seven
diff --git a/harness/tests/scene_gates/test_zone_contents_nesting.py b/harness/tests/scene_gates/test_zone_contents_nesting.py
index 7547eda4..9d9a0799 100644
--- a/harness/tests/scene_gates/test_zone_contents_nesting.py
+++ b/harness/tests/scene_gates/test_zone_contents_nesting.py
@@ -7,11 +7,15 @@ module built each one and named it exactly that — `windmill.name = 'Windmill'`
The placement table lists direct children of a zone, so the check saw one row named
`IslandAssembly` and called the plan unmet.
-All five of arm B's remaining `scene_placement` ERRORs were this family; the other two are
-the model naming the object after its behaviour (`PineSway_0..11` for a SnowyPineTree,
-`LanternPost1/2` for a HutPorchLantern). Those stay ERRORs — naming an object after the
-plan's content IS part of the contract, because that is how every gate and the judge find
-it — but the hint now says so.
+**All five** of arm B's remaining `scene_placement` ERRORs were this family, including the
+two that looked like the model renaming things. `PineSway_0` and `LanternPost1` are the
+names the ZONE gave the placed objects, but each contains the asset module's own root one
+level further down — `inner` reads `['SnowyPineTree', 'Trunk', 'Boughs', …]` and
+`['HutPorchLantern']`. Re-probing all three recorded workspaces takes the contract from
+5 ERRORs to **0**.
+
+The last test below still pins the rule for a subtree that genuinely does not carry the
+planned name; that shape did not occur in this corpus.
"""
from __future__ import annotations
@@ -52,9 +56,9 @@ def test_content_that_really_is_absent_is_still_reported():
assert _missing(out) == ["RopeBridge"]
-def test_an_object_named_after_its_behaviour_still_reads_as_missing():
- """`buildSnowyPineTree` was called and the objects were named `PineSway_N`. The tree is
- there, and the gate still says so — naming it for the plan is what makes it findable."""
+def test_a_subtree_that_carries_the_name_nowhere_still_reads_as_missing():
+ """The rule is "somewhere under this row", not "anywhere in the scene": a zone whose
+ whole subtree never mentions the planned content is still missing it."""
row = {"name": "Trees", "zone": "WindmillIsland", "inner": ["PineSway_0", "PineSway_1"]}
out = contract_findings(_census(row), _plan(["SnowyPineTree"]))
assert _missing(out) == ["SnowyPineTree"]