Simulation (and ArenaSimulation in the balancing tool) now owns the factory's
world data; BuildingSystem holds a reference to it. This is what lets the systems
that operate on the data be handed the same state — phase 3's construction and
deconstruction systems, and later the ecs/system/ classes that today take a
BuildingSystem& only to query it.
reset() clears the state alongside m_admin and m_beltSystem, matching how the
subsystems were already rebuilt from scratch.
Falls short of the tick-argument form I sketched: BuildingSystem still reaches
the data through a member reference rather than a parameter. Making it truly
stateless means the const query surface has to find the data some other way, and
that surface is large — findBuilding alone has 64 call sites, with findSite,
getAllBuildings, getAllSites, isTileOccupied and the rest behind it. Doing that
needs a queries facade behind Simulation::getBuildings() so the callers do not
all move, which is its own decision rather than a side effect of this one.
The constructor gains a parameter, so the four owners and the three test fixtures
that build a BuildingSystem directly are updated; the 33 files that only use one
are untouched.
Verified with a golden-checksum capture before and after — all four sample ticks
identical.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YHcUerKAZKWNvSKJxYKbnG
The scripted session only placed buildings, so every unlock container stayed at
its initial value for the whole run and the checksum never saw them change. It
now destroys the enemy stations twice and takes the offered schematic choice,
so awarded groups, per-schematic levels and the derived recipe/item sets are
exercised too.
Adds a test that pins down that unlock state actually reaches the checksum: two
sessions in lockstep, one takes the choice, checksums must diverge. The
scripted-session tests cannot show this themselves — they compare runs against
each other, so they pass whether or not UnlockState is in the fold. Verified by
temporarily removing the fold: the new test fails, the old two do not.
The first choice is asserted rather than assumed, so a config change that stops
offering a group fails loudly instead of silently dropping the coverage.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YHcUerKAZKWNvSKJxYKbnG
rotateInPlace re-implemented the belt-tile re-registration switch inline
instead of calling reregisterBeltTile, and its splitter branch omitted the
setSplitterFilters call the canonical version has. Since an operational
splitter keeps its filters only in BeltSystem, removeTile discarded them and
rotating a configured splitter silently reset it to "accept all".
Replace the duplicated switch with a call to reregisterBeltTile, capturing
the filters beforehand via getSplitterInfo — the same idiom deconstruct
already uses. This removes the second copy of the switch that allowed the
two to drift apart in the first place.
Add a regression test; the existing [rotate-in-place] cases covered belt
tiles only, which is why this went unnoticed.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GH8ZMRY3vhxxXcaUxBqxkk
Add deterministic record/playback for a run.
Recording captures `(seed, config hash, ordered tick-tagged commands)` and re-simulates on playback — no state snapshots. `DotaFactory.exe --replay <file>` re-plays a recorded run view-only with manual speed/pause.
Reviewed-on: #4
Co-authored-by: Malte Langkabel <malte.langkabel@gmail.com>
Co-committed-by: Malte Langkabel <malte.langkabel@gmail.com>
