The adversarial review's three blocking findings, each fixed at root:
1. A vendor session now CLOSES when its entity retires (despawn,
death, ObjectDelete) and at teleport BEGIN
(HasPendingTeleportStart || IsTeleportActive at the existing
per-frame seam — both hosts funnel through
RuntimeWorldTransitState.TryQueueTeleportStart, which flips the
pending flag strictly before activation). The previous permissive
early-return stranded the session forever: panel pinned to a stale
guid, ActiveVendorId swallowing Use for the rest of the session.
2. VendorShopItem carries the desc's stack size, and
VendorPricing.PerUnitValue ports retail's stack-total division
(VendorProfile::VendorSellPrice 0x005D1B00: <= 0 guard, integer
division) — a stack of 50 arrows now prices per arrow, not at 50x.
3. VendorState.Close() guards its observer fanout with the
dispatcher's catch-and-log semantics — a throwing panel listener
can no longer propagate into the unprotected per-frame path.
Register honesty rides along: the 0.6 m UseRadius fallback was
acdream's invention (ACE's CheckClose has no fallback; retail passes
the raw authored radius) — removed, the watcher now uses the raw
radius and AP-160's citations are corrected and extended with the
accepted-position-snapshot cadence; AD-72 files VendorPricing's
double-vs-x87-extended narrowing (AD-33's class, bounded by the
±0.1 margin).
Nine tests added. Clean-room complete solution: 11,311 passed /
4 skipped / 0 failed.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
VendorState sits beside ExternalContainerState (contract decision 1)
with the same shape: private setters, Changed event, Reset with
AggregateException fanout; domain-shaped like ContainerContentEntry
since Core cannot reference Core.Net. No Runtime wiring, no UI — 5.3's
job.
VendorPricing ports ShopSystem::BuyPrice/SellPrice (0x006B6120/
0x006B6180) faithfully: retail's literal three-way branch survives,
including the unreachable-with-real-data negative -1 sentinel that
ACE's Math.Max(1, ...) collapse erases — equivalence for legitimate
inputs is hand-proven and documented rather than silently assumed.
Seven conformance tests with hand-derived golden values (float32
semantics verified independently), covering rate=1.0, fractional
rates, value=0, the rounding-sensitive halfway case, stack
multipliers, the ItemType rate-override branch, and the sentinel.
Clean-room complete solution with 5.1+5.2 in place: 11,291 passed /
4 skipped / 0 failed.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
