CellTransit.FindTransitCellsBox ports CEnvCell::find_transit_cells'
part-array overload @0x0052cae0 line-for-line: per-portal x per-part
order, the sphere cheap-reject at F_EPSILON+radius, the box admit whose
"Straddle or crossing-side" rule is exactly retail's `eax != side` under
the PDB Sidedness enum, leads-outside placed AFTER the admit, the
unconditional unloaded-neighbour hint without the sphere overload's
re-test, the destination box_intersects_cell gate with its deliberate
no-break, and add_all_outside_cells after the loop. The box-vs-cell BSP
traversal lands in BOTH representations behind the flat-authoritative
dispatcher with a graph referee whose 20,000 installed comparisons are
pinned by assertion (review F5), zero mismatch.
Dual Opus review: PASS on both lenses. The mandatory D0 pseudocode pass
caught that the contract's own supplementary note misattributed the box
block to the sphere overload — it belongs to a SECOND
check_building_transit overload @0x0052c680, whose portal-side
convention is INVERTED and whose admit differs; the pseudocode doc now
records that trap plus two byte confirmations made at review:
which_side @0x00444720 is strictly > eps for POSITIVE, and
intersect_box's in-plane early exit returns CROSSING(3)
(jp @0x005aa1bc -> mov eax,3), settling review items b1/b2 for the
future bridge porter. The bridge itself stays unported as AP-159's
explicit remainder.
The review also retired #335's severity premise honestly: "over-
inclusive only, never a missed one" is wrong at production shape ratios,
where the box (whole-vertex AABB) legitimately exceeds the sphere
(physics-polygon root sphere). Measured, both populations: rigged
(box << sphere) — 1,520 placements, 978 cells removed, 0 added;
production-ratio (box >= sphere) — 950 placements, 20 removed, 1 ADDED
through the loaded-neighbour gate, which is retail's direction, not a
defect. The no-op guard (review F4) asserts removal is nonzero so an
unwired admit cannot pass silently.
Process note: the implementer authored against this session's worktree
at bec5c69d, 25 commits stale — the recorded worktree-base class. All
six files were byte-identical between bases, the diff transplanted
losslessly, and every verdict-bearing run (referee, direction sweeps,
this clean-room) was re-executed on current main. S2's uncommitted
phase-1 edits were stashed for this landing so the suite verdicts
exactly one changeset.
Also untracks 341-slope-capture.jsonl (an accidental add) and
gitignores it.
Clean-room suite: 11,248 passed / 6 skipped / 0 failed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Make prepared flat BSP data authoritative for gameplay while retaining the parsed graph only as an exact sampled referee. Fail production publication when collision package data is genuinely absent, keep idempotent already-cached publication valid, and move cell membership, floor lookup, camera diagnostics, and live/static shape bounds onto the flat representation.
Validated by 8,402 Release tests, a strict dense-Arwic connected gate with 46,309/46,309 exact referee matches, and graceful shutdown.
Co-authored-by: Codex <codex@openai.com>
Carry one immutable prepared collision closure with each accepted near-tier generation and install graph plus flat views through the same retained publication receipt. Apply the same strict package-only rule to live entities, add exact sampled graph-authoritative comparison artifacts and lifecycle counters, and prove cancellation, demotion, rehydrate, revisit, teardown, reconnect, and the nine-stop route with 14,064 zero-mismatch samples.
Co-authored-by: OpenAI Codex <codex@openai.com>
