Repair deterministic Occupy test transitions and card staging
This commit is contained in:
1 parent
38b1ba3419
commit
f2fd274502
5 files changed
+61
-18
No files matched your search
+11
-7
@@ -215,13 +215,17 @@ reveals before that companion moves, in both modes. A repeated-Karama
|
||||
cancellation of an already-canceled compulsory cache lot remains an explicit
|
||||
open boundary recorded in the decision index. Two intermittently failing
|
||||
Moritani skill-runtime tests were measured at the same 3/30 rate on both
|
||||
revisions, so they are pre-existing nondeterminism rather than a regression; the
|
||||
same holds for `tests/ecaz-occupy-options.test.ts`, whose Easy Occupy plan
|
||||
candidate is occasionally absent (2/20 failing runs at `4ed6beb` versus 3/20
|
||||
now), for the paired Moritani skill-runtime tests, and for
|
||||
`tests/ecaz-occupy-cards-runtime.test.ts` (one intermittent Harass/explosion
|
||||
case, passing 5/5 standalone). Those intermittent failures are pre-existing
|
||||
prototype/AI gaps and are not new regressions.
|
||||
revisions, so they are pre-existing nondeterminism rather than a regression.
|
||||
The Occupy policy failure (2/20 at `4ed6beb`, 3/20 here) is now explained by
|
||||
an obsolete immediate-plan assertion: seed 8 legally activates Ecaz special
|
||||
Karama before sealing. All four profiles finish that captured state in two
|
||||
actions; the deterministic regression now checks finite legal progression.
|
||||
The Harass/explosion failure is now reproduced at deterministic seed 62:
|
||||
the fixture's donor exchange removed its already-selected Harass card.
|
||||
Staging now uses a conserved deck replacement while hand capacity remains.
|
||||
The repaired physical scenario returns two forces before the explosion,
|
||||
discards Harass once and reaches Collection. Neither finding changes rules
|
||||
or bot behavior; the earlier attribution to prototype/AI gaps was incorrect.
|
||||
|
||||
**Gated captures resumed under the new prototypes.** The 14 captures that
|
||||
remained gated after the 7 October rulings were resumed in place against their
|
||||
|
||||
@@ -1,5 +1,31 @@
|
||||
# Dune implementation status
|
||||
|
||||
## 9 October 2026 — two Occupy failures isolated to test contracts and staging
|
||||
|
||||
Deterministic seed 8 reproduces the recorded missing-immediate-plan failure.
|
||||
The bot actually offers a legal Ecaz special Karama action; Easy, Medium,
|
||||
Hard and Brutal each then seal the selected lead's plan on their next action.
|
||||
The actual captured engine smoke consumes Karama once and submits the plan.
|
||||
No bot or rule behavior changed.
|
||||
|
||||
The regression now pins entropy and follows actual first-policy actions to
|
||||
a sealed plan for both leads and active/canceled Occupy instead of requiring
|
||||
the first action list to contain a plan. That case passes.
|
||||
|
||||
Deterministic seed 62 also reproduces the Harass ally/explosion failure:
|
||||
when Lasgun must come from another hand, fixture staging gives that donor
|
||||
the actor's sole held Harass, then tries to seal the now-unavailable defense.
|
||||
The engine correctly rejects it. The staging helper now funds a donor
|
||||
replacement from the original deck when the recipient has room, retaining
|
||||
its earlier selected physical cards. Full hands keep the existing exchange.
|
||||
The real repaired scenario returns two forces before the explosion,
|
||||
discards Harass once and reaches Collection. Its deterministic regression
|
||||
passes. Neither repair changes production engine or bot behavior.
|
||||
|
||||
Frozen verification: `check:quick` passes and all 270 cases in the nine
|
||||
affected Occupy, Stronghold and native-tech files pass. No full offline
|
||||
suite, HTTP campaign, database audit or deployment was run.
|
||||
|
||||
## 9 October 2026 — provisional Kulon/Ornithopter coexistence
|
||||
|
||||
Active Kulon no longer blocks either mode of the Richese movement card.
|
||||
|
||||
Reference in new issue
Block a user