The margin is exonerated. Every earlier run excluded 8 voxels at each
strate boundary, and voxels 4-8 of those are ordinary interior, not seal
-- the same window mistake that produced "FlatPlain is 99.95% void" at
margin 32. So the query was re-run at margin 0, 2, 4 and 8. All four stay
negative; margin 0 includes the seal itself and still reports negative.
(The implementation had clamped explicit margins to >= 1, which made the
margin-0 check impossible; Codex found and fixed that.)
Diagnosis -- this is not a near-miss wall:
arrival LowerPoint (65.8, 95.8, -1047.25)
departure UpperPoint (88.4, -104.6, -991.58)
XY separation 201.7 voxels (~50 m), Z delta 55.7
arrival component 22,531 cells (22.0% of air)
departure component 21,621 cells (21.1% of air)
nearest approach between the two components: 135-179 voxels
Two large, healthy, entirely separate shaft networks half a strate apart.
Nothing to breach; the archetype has no lateral connection at all.
16-seed sweep: 1 pass, 15 fail. Not one unlucky world -- structural.
So: descend into a VerticalShafts strate and you probably cannot reach
its exit. This is the "live hole in the shipping game" COMPOSER-NOTES
Tier 1 predicted, now measured instead of assumed.
Stated precisely, because it names the fix: VF_SuggestLandingPoint makes
each mouth land in open space, but NOTHING makes the two mouths of one
strate land in the SAME open space.
New: VF_DiagnoseConnectivity returns component sizes and the spatial
separation between the two mouths' components, so a near-miss and a
never-meant-to-connect read differently instead of both being "false".
No generation change, no threshold tuning. The fix is a design decision.
Build clean, 19 succeeded / 0 warnings / 0 failed. Sweeps 3.4 s + 6.9 s.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bp7FxoQcSoiuhxjuLr6aBM
Fr0zkaandClaude Opus 5
0bbfdeb7a7