The panel was anchored to the right edge of the view, which is nowhere near
whatever the player just clicked. It now stands beside the selection: right of
it where it fits, otherwise left, otherwise the roomier side pushed inside the
view -- the one case where it covers part of what it describes.
Nothing told the panel where the selection was. The selection events carry ids
only, and the mode that separates a fresh selection from an expanded one is
consumed inside SelectionController before they are built, so the view now
publishes the selection's screen bounds itself, immediately before selecting and
only when the selection is starting. Freezing that rectangle is what holds the
panel still: it does not chase a scrolling view, a ship flying off, or a
selection being added to. Only the panel's own size still moves it, and even
then it keeps its side and the edge facing the selection.
The rectangles come from what the renderer was already computing for the
selection outlines, now shared rather than duplicated.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JcReq7hVk4KUPhTDKWAG7K
The panel decides nothing about what the controls are: it asks the same
resolver the key handling and the mouse dispatch ask, and renders each badge
from the binding that resolver matches. A chip cannot claim a key that does
nothing, and a label cannot describe a click that does something else, because
neither is written here.
Display text is the one part that is not shared -- lib/core says what is
available and what triggers it, ControlActionText.cpp says what it is called.
That split is why the table can stay free of UI strings and still be the single
source of the pairing.
It refreshes on a timer rather than by subscribing. Two things that change a
row -- a belt drag starting, the ghost crossing a transfer target -- happen on
mouse movement and publish nothing, and they must still show while the game is
paused, so there is no event and no tick to hang it on. Resolving is comparing
two vectors of enums, and the rebuild is skipped unless they differ.
Left edge, bottom-aligned in the same band the selection panel is confined to,
so it clears the build bar's strip and never meets the panel on the opposite
edge. Clicking the heading collapses it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YHcUerKAZKWNvSKJxYKbnG
The controls panel needs to say what each key does right now, and a panel that
keeps its own list of that is a list that goes stale. So the list moves into
lib/core/ControlAction.h: which actions exist, what each is bound to, and when
each does something. InputMapper stops deciding that and switches on the
resolved action instead, so the panel and the key handling cannot disagree
about what Q means -- there is only one place that says.
The table declares; it never performs. It holds no simulation access, fires no
events, and names nothing: display strings live in the ui target, which formats
the bindings this hands it, so a badge is rendered from the real binding rather
than typed beside it. What an action *does* stays exactly where it was.
ControlContext is the snapshot the rules read, which is what keeps this
testable without a world. Two of its facts come from BlueprintLibrary, which is
built after the world view and so arrives by setter; one comes from the new
hovered-transfer flag on BuildModeController, resolved once on mouse-move
through the same classifier the click and the ghost colour already use.
Behaviour is unchanged, deliberately. Ctrl still separates the chords and every
other modifier is still ignored, so Shift+A pans as before; matching modifiers
exactly would have silently swallowed those presses. Build hotkeys and F3/F4
stay outside the table -- the first are advertised on the build buttons and
already derive their badges from the handler's own table, the second are
development controls the panel must never offer.
The tests are the point of putting this in lib: every row's bindings must
resolve back to that row's action in that same context, which fails the moment
a shown row and its handler part ways.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YHcUerKAZKWNvSKJxYKbnG
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>
