VF_AreConnected took the shortest coarse BFS route, re-walked it at full
resolution, and on hitting a wall returned false and gave up -- never
looking for another way round. That is a false NEGATIVE, and it was
inflating the failure count of a bug we were actively trying to fix.
On refutation it now excludes that specific coarse step and re-searches,
up to a bounded budget (16). Outcomes are distinguished honestly:
Connected a route survived full-resolution verification
NotConnectedAtThisResolution no coarse route remains after exclusions
CoarseLiedBudgetExhausted UNKNOWN -- routes kept being refuted
The third is deliberately NOT folded into either other verdict. Collapsing
"unknown" into "disconnected" is precisely what misled us.
arrival->departure 7/16 -> 14/16
All 7 former COARSE_LIED became CONNECTED.
arrival->spine 15/16 connected departure->spine 15/16 connected
per-archetype law: 6/6 applicable routes pass
Retries needed are small -- max 3 per query, 0 budget hits -- so 16 is
generous rather than load-bearing. Runtime slightly IMPROVED (45.3 -> 44.9 s
on the 16-seed sweep) because a satisfied query stops early.
The false-POSITIVE guard is untouched and is now stronger in meaning:
every Connected still requires a fully re-walked, air-verified route, and
it now says "this specific route is open" rather than "the first route I
tried was open". The controlled wall still returns BudgetExhausted after
exactly 16 retries, and a repeat query is bit-identical in verdict, snap
flags and retry count.
⚠️ REMAINING: seeds 6 and 14 are genuinely NOT_CONNECTED. Real or a
step-2 resolution limit is not yet settled -- refine before touching
generation again, per COMPOSER-NOTES §13.
No generation, placement or parameter changed. Verifier only.
Build clean, 20 succeeded / 0 warnings / 0 failed, confirmed in VoxelM.log.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VJiTcTSDNkKBtnNdiBtdaU
Fr0zkaandClaude Opus 5
44938f3f98