18
0
mirror of https://github.com/MaSzyna-EU07/maszyna.git synced 2026-11-19 22:44:59 +01:00
Commit Graph

3567 Commits

Author SHA1 Message Date
maj00r
10c7527a64 AI: reduced signal speed ends at a marked end of the interlocking area (SignalAreaEnd) (#305)
* feat(ai): reduced signal speed ends at a marked end of the interlocking area

Ie-1 (from 17.01.2026, signals on colour light signals, par. 8): a reduced
speed shown by a signal (S10 "40, then Vmax" and the like) holds to the end
of the interlocking area the signal guards. Only on secondary main tracks
does it hold for the whole route, and on semaphore signals (Sr3) at an entry
or route signal 40 holds for the whole route as well.

The AI kept the speed of the passed signal (VelSignalLast) until it passed
the next signal, or until no signal, switch or W4 was left in the scan. On a
station the exit signal, the far switches and the W4 are always in the scan,
so after S10 at the entry signal onto the principal main track the AI ran 40
all the way to the exit signal.

A scenery can now mark where the interlocking area ends, with a putvalues
event read by the scan like a signal:

  event <name> putvalues 0 none x y z SignalAreaEnd 0 0 endevent

attached to event1/event2 of the track (the direction the train runs), past
the last switch of the area, on the principal main track only. In train mode
the reduced speed of the passed signal ends once the rear of the train is
past the marker, i.e. the whole train has left the switches. A signal at
stop passed (speed 0) is not lifted. The marker sends no command and is not
treated as a signal or a speed limit: it does not end the restricted run on
sight after S1 on automatic block, and it means nothing in shunting. The
driver hint for the end of the speed limit uses the marker when there is one
instead of the farthest switch in the scan.

Where the marker is left out (secondary main tracks, stations with semaphore
signals, every existing scenery) the speed holds as before.

Game data: nothing has to change in the SVN for this to work. Sceneries may
add the marker; a release without this change does not read SignalAreaEnd
(an unknown putvalues command is not scanned), so it ignores it.

* fix(ai): a marker behind the last passed signal does not end its speed

The marker acted on whatever speed the last passed signal had left when the
rear of the train passed it. A train longer than the way from the marker to
the next signal, or one which started standing over the marker (scenario
start at a platform, a direction change), had its head past the next signal
before its rear was past the marker of the previous area, and the marker
then lifted the 40 of the new signal while the head ran through its
switches. Passing a signal now drops the markers that lie before it in the
scan table: they end the area of an earlier signal.

In shunting the marker is no longer thrown away when the rear passes it, so
a train order given before the next signal still finds it.

A note to the previous commit: a release without SignalAreaEnd does not scan
the marker, it runs it as an ordinary putvalues event when a vehicle passes,
which logs it and sends the mover a command it does not know, to no effect.

* AI: a higher road speed (W27/W27a) takes effect once the rear is past the sign (#306)

* fix(ai): a higher road speed (W27) holds only once the rear is past the sign

The road speed of a RoadVelocity event (W27/W27a) took effect as soon as
the head of the train passed the sign, both a lower and a higher one, and
the entry left the scan table in the same pass. A lower speed from the head
is right: the AI brakes before the sign, it is in the scan. A higher one let
the AI speed up while the rest of the train was still on the stretch of the
lower speed.

Now a higher road speed (compared through min_speed, so -1 is no limit)
holds once the whole train is past the sign. Until then the entry stays in
the table and is left out of the rest of the table update, which would have
dropped it. What the update did to a passed passive event when it dropped
it (command, VelSignal, the end of the restricted run on sight) now happens
when the rear passes the sign instead of the head. One-way signs and 2w27
stay as they were: the scan reads only the event of the running direction.

The comments of the road speed and section speed branches named them W6
and W27; they are W27/W27a and W9.

* fix(ai): road speed of cars from the head, W27 does not end the run on sight

Two faults of the previous commit, found in review:

A car's passed events leave the scan table at the latest when its rear
passes them (TableCheck), before the table update could see the rear past
the sign, so an AI car never took a higher road speed and kept the lowest
one for good. A car takes the road speed from the head again, as before.

With the sign now kept in the table until the rear passes it, the generic
handling of a passed passive event ran only then, and it ends the run on
sight after S1 on automatic block. A train which stopped at S1 with its rear
not yet past a W27 lying shortly before the signal would, once on the move,
end the run on sight and run into the occupied block at road speed. A road
speed sign is no signal: it no longer ends the run on sight at all, which
also ends the same thing happening, before this branch, to a train whose
head passed a W27 while running on sight.

* fix(ai): a passed road speed sign is handled when the head passes it, as before

The previous commit left a road speed sign out of the table update until
the rear passed it, so the generic handling of a passed passive event (the
SetVelocity command, VelSignal, the end of the run on sight) ran up to one
train length later than before. By then the order may have changed: a
train standing over a W27 at its last station, with Shunt as its next
order, would be switched back to Obey_train by that command once its rear
passed the sign on the next forward move.

Now the sign is handled as before in the first pass after the head passes
it. In place of dropping it from the table it is then marked as held
(spRoadVelHeld) and only waits for the rear, left out of the rest of the
update. So everything but the moment a higher road speed takes effect is as
it was, and the change of the previous commit that kept a W27 from ending
the run on sight is undone: it is not needed any more.

A car is told by CategoryFlag == 2, the same test TableCheck drops its
passed events by, so a unimog (rail and road) holds the sign for its rear
like a train.

* fix(ai): a road speed released by the rear doesn't exceed a lower sign after it the head has passed

A train longer than the distance between two road speed signs, as W27a
120 and then W27a 60, passes the 60 with its head before its rear is
past the 120. Releasing the 120 then set the road speed to 120, and only
the held 60 handled later in the same table pass brought it back down,
so the result depended on the order of the table and on the loop
reaching that entry (it stops at the first point that orders a stop).

When the rear passes a sign, the road speed is now the lowest of that
sign and of the signs after it that the head has passed and the rear
not yet. Without such a sign it is set as before.

* fix(ai): a passed road speed sign not yet marked as held caps the released one too

The sign after the released one may have been passed by the head in
the very tick the rear passes the released one, before the table update
marks it as held. It is now taken into the lowest speed as well (an
enabled road speed entry behind the head), so the result doesn't depend
on the table update reaching it in that tick.

---------

Co-authored-by: maj00r <maj00r@host>

---------

Co-authored-by: maj00r <maj00r@host>
Co-authored-by: jerrrrycho <jurek.pakula@gmail.com>
2026-11-11 10:44:42 +01:00
jerrrrycho
5bac4e1cf1 Fix the four issues failing the Sonar quality gate on master (#323)
* fix(scripting): forward the object in push_object of the Lua screens (cpp:S5417)

push_object takes a forwarding reference and called std::move on it, which
SonarCloud reports as a blocker bug. It only worked because every caller
passes an rvalue: with an lvalue Type_ is deduced as a reference, and
new Type_(std::move(Object)) doesn't even compile.

The object is now forwarded, and the userdata is sized and built for
std::remove_cvref_t<Type_>, the type collect<> and check_object<> are used
with: an rvalue is moved in, an lvalue would be copied. Every caller passes
an rvalue of lua_image, bitmap_font or glyph, so they move the object in
exactly as before.

* fix(simulation): turbidity from the overcast without branches of the same value (cpp:S3923)

The chain of conditionals setting Global.fTurbidity had pairs of branches
with the same result (<= 0.10 and <= 0.20 both 4, and so on, the last two
both 9), which SonarCloud reports as a bug. The thresholds that changed
nothing are gone and the mapping is a small function: up to 0.20 gives 4,
up to 0.40 5, up to 0.60 6, up to 0.70 7, up to 0.80 8, anything else 9.
The comparisons are the same ones, in the same order, so every overcast
value gives the turbidity it gave before.

* test: script screen tests keep their files in the build folder, not in the temporary one (cpp:S5443)

The folders of the script screen host tests were made in the system's
temporary directory, and the sandbox test put its "outside.txt" straight
into that directory. SonarCloud reports both uses as vulnerabilities:
the directory is writable by everyone.

The folders are now made in the build folder of the unit tests, passed to
them as EU07_UNIT_TESTS_SCRATCH, the way EU07_UNIT_TESTS_DATA is. The
scripts are one level down in the folder of a test, so the file the sandbox
test puts outside the folder of the scripts is still inside the folder of
the test and goes with it. Nothing depends on the working directory, and
nothing is left behind.
2026-11-11 06:04:31 +01:00
maj00r
b251b6fe42 feat(blackbox): a black box of the session: the run, the cab, the signals and the stations to a file (#322)
* feat(blackbox): a recorder of the run, the cab, the signals and the stations to a file

With the physics logs on (and the new eu07.ini setting blackbox, on by
default), the simulator writes a black box of the session to
physicslog/*.blackbox: the scenery map (tracks, signals, W4 stop points
with their heading, W6 horn points), the timetables of the user's train,
and at 10 Hz the cab of the vehicle the user sits in (speed, controllers,
brakes, pressures, lights with the pattern the run asks for, the movement
kind), the AI's reference for the same place (target speed, next limit,
braking distance), the consist, the cameras and the vehicles around.
Changes go in as they happen: switches, signal aspects, commands,
signals passed with their aspect, stations stopped at, passed or missed,
horn requests, the autopilot switched on or off, couplings and vehicles
removed.

The driving aid's next speed limit moved from the UI panel to
TController::next_speed_limit(), so the panel and the recorder read the
same. The format is described in blackbox_doc.md; the writer and the
reader of the records have unit tests.

* fix(blackbox): the aspect a signal had when passed, the lamps the ends have, vehicles taken off

- A passed signal is written with the aspect the train had of it, kept in
  the speed table, not the one its memory cell shows when the pass is
  noticed: by then the train itself on the track behind has often set
  it to stop, and most passes were written as at stop.
- The cab record gains the lamps the vehicles at the head and the tail
  have at their outer ends (136 bytes), so a reader can check the light
  pattern against them as the driver does (no upper headlight, the
  auxiliary lamps, the rear-end signals only).
- A vehicle taken off the scenery and still pointed at by a coupler or
  the camera gets no new number; the camera lets go of it.
- The timetable written before a station goes under the vehicle the
  user is in now, not the one of the last look.
- The file buffer is given once the file is open: MSVC passed over one
  given before.
- The doc tells the AI fields before the consist is started and when
  entering a vehicle is a driver change.

* fix(blackbox): the stop of an automatic block signal passed is written as a stop

The speed table keeps the stop of an automatic block signal (SetVelocity 0 x) as the speed allowed after it, with spStopOnSBL; a pass of it was written as that speed with no flag. It is written as 0 with the flag of an automatic block stop again.

* feat(blackbox): the doors, the brakes of the whole train, a diesel's engine and the rest of the cab

The cab record grows from 136 to 190 bytes, so that a reader can check
everything the driver did:

- the doors: the permits on each side, the departure signal, the steps,
  the door lock and whether the doors are locked;
- the brakes: the equalizing reservoir of the driver's valve, the main
  reservoir pipe, the highest brake cylinder pressure of the train, the
  hand brake, the releaser, the setting (G/P/R/Mg) and the mode
  (PS/PN/EP/MED);
- what drives the vehicle (electric, induction, diesel, diesel-electric,
  steam, none), its power, a diesel's oil pressure, water, oil and engine
  temperatures, overheating, fuel and oil pumps, clutch and torque
  converter, the motor fans;
- the cruise control and its speed, the integrated driving handle of a
  unit, the radio channel, the horn's tones.

New device bits use the upper half of the 32-bit field; the new fields
come at the end of the record, so older readers still read it.

* fix(blackbox): the driver's valve field is what its gauge shows, the radio channel has no none

Handle->GetCP() is the equalizing reservoir of most valves but the time reservoir of the FV4a, empty but after an overcharge: the field is named and documented as what the valve shows. With no driver the radio channel reads as the first one, not 0.

---------

Co-authored-by: maj00r <maj00r@host>
2026-11-11 04:30:52 +01:00
maj00r
18c56359ee fix(ai): the stop point on a W4 kept the margin of a passenger stop from the platform's end (#321)
The W4 stop point was shifted with fMinProximityDist (8 m), while the
train comes to a stand anywhere up to the margin of a passenger stop
(10-15 m by the number of vehicles) short of it. A train stopping by
the middle of a platform could so stand with its tail a few metres past
the platform's end (L053: Swoszowice 204.5 m of 201, Tumlin 244.3 m of
242). The point is now shifted with the same margin the stop is judged
by, so the tail stays inside (197.5 m and 240.4 m on the same run).

Co-authored-by: maj00r <maj00r@host>
2026-11-11 04:30:13 +01:00
5764e942a7 build(linux): load the shared libraries from the game folder and ship them in the build (#320) 2026-11-11 04:29:36 +01:00
226155d430 feat(physics): physical coupler model for the player's train and those it closes in on, plastic hit on solid buffers 2026-11-11 03:15:58 +01:00
aba8687678 feat(ui): repair the broken coupler next to the camera from the driving aid 2026-11-11 02:52:00 +01:00
8e97277bcb feat(debug): coupler section in the debug panel, with the parameters of the physical model 2026-11-11 02:28:46 +01:00
c85aab4f3f feat(physics): physical coupler model with slack, hysteresis and breaking 2026-11-11 01:43:46 +01:00
39e855d3a5 Merge pull request #319 from MaSzyna-EU07/docs/agents-md
docs: AGENTS.md as the one set of rules for every coding agent
2026-11-10 23:40:52 +00:00
b42d2f7b66 docs: AGENTS.md as the one set of rules for every coding agent build-docs-agents-md 2026-11-11 00:40:27 +01:00
b3e6e02779 ci(windows): build lua51.dll for the unit tests 2026-11-11 00:34:59 +01:00
maj00r
5c8cc6d9a5 feat(ai): PointVelocity, a speed limit for the head of the train at a point beyond the cell (#304)
* feat(ai): PointVelocity, a speed limit for the head of the train at a point beyond the cell

A memory cell (or putvalues) with "PointVelocity v d" tells the train that the head has to be at v km/h
at the point d metres beyond the cell (v -1: no limit). The cell is read by the speed table scan like a
signal, live until the head passes it; what it says then is kept for the point, later changes of the
cell don't concern that train. The limit ends when the head passes the point, it doesn't hold the rest
of the train and doesn't extend or break a speed limit of the line, like W27 of length 0 moved d metres
on. The AI brakes so the head reaches the point at v; for a human driver it only shows as the next
speed limit of the driving aid, the security systems don't see it.

It is made for the warning disc of a level crossing (TOP, Osp1/Osp2), which has to give 20 km/h at the
crossing about 500 m beyond the disc, as the disc showed when the train passed it.

The cell with PointVelocity as its first command makes its getvalues scanned (passive) as SetVelocity
does; PutCommand ignores it when such an event runs. Old commands keep working as they did,
SetProximityVelocity included.

* fix(ai): a point limit isn't a signal, and a PointVelocity 0 can be released

From the review of the PointVelocity commit:
- a cell rewritten from SetVelocity or ShuntVelocity to PointVelocity drops the flags of a signal, so it
  can't count as a semaphor or report a signal passed at danger;
- PointVelocity 0 is read on after the head passes the cell, as a stop of a signal is, otherwise the
  train would stand short of the point for good;
- a point limit doesn't become eSignNext, so a W4 between the disc and the crossing keeps its release.

* fix(ai): a PointVelocity limit outlives a rebuilt speed table, and holds when shunting

The state of the cell, latched when the head passes it, is now kept by the
controller apart from the speed table, and the distance to the point is
measured by the movement of the consist in the world. Stopping, backing up
and going on between the cell and the point, coupling or a takeover no longer
lose the limit or read the cell anew, unless the head backs up past the cell,
when the cell is read again.

Near the point the desired speed is capped to the limit, so a train that
coasts within its tolerance (as a shunting one does) reaches the point at it.

* fix(ai): review of the latched point limits

- a point the head has passed already isn't latched, and a latched limit at
  or behind the head isn't applied (it asked for a full brake for a cycle);
- a latched limit is dropped once the head backs up behind its cell, so it
  can't follow the consist onto another track; the table reads the cell again;
- a cell latched already, found under the consist by a rebuilt table, is
  ignored whatever it shows now (a 0 for the next train no longer stops it);
- PointVelocity 0 stops at the usual distance, not from the approach;
- the heading follows the track under the measured vehicle, not the chord of
  the consist, which a long consist on a tight curve holds far off the track;
- vehicles coupled or uncoupled at the end towards the point move the head;
- a jump of the consist (moved elsewhere) drops the latched limits.

* fix(ai): a cab change and curves don't move the latched point limits

- the end of the consist towards the point is taken into the distance only
  when the consist changes (its end vehicles or its length), not when another
  vehicle is measured from (a cab change at a reversal) nor when a curve
  shortens its projection;
- the way of travel follows the direction of the driver's vehicle against
  its own couplers, not the chord of the consist, which on a balloon loop
  points against it.

* fix(ai): a coupling with a cab change in one order still moves the head of a latched point limit

The end of the consist measured from the previous vehicle is moved to the place of the new one, so a coupling or an uncoupling followed at once by a change of direction to the other cab (the shunting orders) is taken into the distance to the point.

---------

Co-authored-by: maj00r <maj00r@host>
2026-11-10 18:44:33 +01:00
ee6f9a0aa2 Merge branch 'master' of https://github.com/MaSzyna-EU07/maszyna into feat/wasm-screens 2026-11-10 16:23:35 +01:00
e433b43fef Script screens: led_post and resize of the Lua screens on the gpu 2026-11-10 16:22:35 +01:00
c4924f756d Lua script screens beside the WebAssembly ones 2026-11-10 15:29:25 +01:00
a73585dca0 Dont copy entire python env when already present 2026-11-10 01:05:58 +01:00
18c333b089 wasm display support 2026-11-10 00:57:37 +01:00
c9792672c6 fix(render): textures of a far vehicle no longer come back black with gfx.resource.move (#308) 2026-11-10 00:41:14 +01:00
maj00r
685258d787 Merge pull request #274 from MaSzyna-EU07/feat/elf-master-controller-emergency-brake
Elf: emergency brake position in the master controller (eu07/maszyna#413)
2026-11-09 18:59:12 +01:00
jerrrrycho
12299bb1fa Merge remote-tracking branch 'origin/master' into feat/elf-master-controller-emergency-brake 2026-11-09 16:33:31 +01:00
1760c90934 Merge pull request #271 from MaSzyna-EU07/feat/diesel-throttle-position-marker
Throttle position marker for diesel locomotives (#74)
2026-11-09 15:23:17 +00:00
jerrrrycho
6f1055e09d fix: a duplicated cab control keeps its scale table
MirrorCopy() of a duplicate: yes control (#285) now copies the scaletable: of the
duplicate's definition too, so the copy is animated on the same non-uniform scale.
build-feat-diesel-throttle-position-marker
2026-11-09 16:13:43 +01:00
jerrrrycho
478727850d Merge remote-tracking branch 'origin/master' into feat/diesel-throttle-position-marker
# Conflicts:
#	vehicle/Gauge.h
#	vehicle/Train.cpp
2026-11-09 16:07:48 +01:00
jerrrrycho
cf3080c1f3 Merge remote-tracking branch 'origin/master' into feat/elf-master-controller-emergency-brake
# Conflicts:
#	McZapkie/Mover.cpp
2026-11-09 15:30:38 +01:00
d5e524ee60 Merge pull request #231 from dos1/aiui
driveruipanels: Allow to toggle AI driver from the driving aid
2026-11-09 13:00:26 +00:00
f908ae307b Merge pull request #265 from pwarszawski/feat/ai-braking-gradients
AI on up gradients: hold a train coming to a stand, start on the train brake
2026-11-09 12:59:34 +00:00
d6b9a94079 Merge pull request #267 from MaSzyna-EU07/fix/light-programmer-coupled-side-headlights
Light programmer: switch off the lights at the coupled end, not the opposite one (eu07/maszyna#395)
2026-11-09 12:58:57 +00:00
7c103b5d2d Merge pull request #268 from MaSzyna-EU07/fix/ws-hold-coil-lever-bistable
Bistable hold coil lever of the line breaker (eu07/maszyna#364)
2026-11-09 12:57:49 +00:00
f61a4ab2e7 Merge pull request #269 from MaSzyna-EU07/fix/diesel-smoke-after-ignition
Diesel exhaust smoke only once the engine fires (eu07/maszyna#523)
2026-11-09 12:57:02 +00:00
7edf23b8ff Merge pull request #270 from MaSzyna-EU07/fix/fpslimit-frame-pacing
fpslimit: pace frames in integer clock ticks, show the real frame rate (eu07/maszyna#416)
2026-11-09 12:56:40 +00:00
fa00fff594 Merge pull request #273 from MaSzyna-EU07/feat/6dg-power-demand-5pct-steps
6Dg: power demand in 5% steps (eu07/maszyna#412)
2026-11-09 12:55:39 +00:00
ff88868d32 Merge pull request #275 from MaSzyna-EU07/fix/sm42-shunt-mode-toggle
SM42: shunt mode at the first move of the switch, CA off in the shunt mode
2026-11-09 12:54:21 +00:00
889aa2d5db Merge pull request #278 from MaSzyna-EU07/feat/shunt-lock-resistive
Field shunting locked at zero and on resistive positions (eu07/maszyna#402)
2026-11-09 12:53:57 +00:00
4a9ab1447e Merge pull request #279 from MaSzyna-EU07/fix/multiple-unit-electric-diesel
No multiple-unit control between electric and diesel locomotives (eu07/maszyna#520)
2026-11-09 12:53:30 +00:00
0c95dbfa5a Merge pull request #280 from MaSzyna-EU07/feat/radio-fuse
Radio fuse: a cut-off in the radio's power supply (eu07/maszyna#438)
2026-11-09 12:53:14 +00:00
ebc9eb8740 Merge pull request #281 from MaSzyna-EU07/feat/shunt-pantograph-config
AI: suggested pantograph setup for shunting in the .fiz (eu07/maszyna#461)
2026-11-09 12:52:58 +00:00
3710b2e5d6 Merge pull request #284 from MaSzyna-EU07/feat/light-programmer-cab-light
Light programmer: cab light flag per position (eu07/maszyna#529)
2026-11-09 12:51:37 +00:00
fedbf458b5 Merge pull request #285 from MaSzyna-EU07/feat/duplicate-cab-control-submodels
Cab controls with more than one submodel (eu07/maszyna#527)
2026-11-09 12:50:54 +00:00
28337b5f28 Merge pull request #286 from MaSzyna-EU07/fix/dimmer-position-bounds
fix(lights): keep the headlight dimmer position within the DimmerList
2026-11-09 12:50:35 +00:00
c1089a6b16 Merge pull request #292 from MaSzyna-EU07/fix/en57-ep2-brake-charging
EN57 brakes: charging when hauled with the brake hose alone, Knorr EP brake only with control power
2026-11-09 12:48:13 +00:00
cb1dd60630 Merge pull request #293 from MaSzyna-EU07/fix/door-inner-faces-lowpoly-lighting
Door wings: let the low poly interior have its own, moving inside of the doors (eu07/maszyna#512)
2026-11-09 12:47:58 +00:00
8ec5d91307 Merge pull request #294 from MaSzyna-EU07/feat/second-loco-engine-start
Separate engine start switches for the vehicles in multiple
2026-11-09 12:47:44 +00:00
af44ed72af Merge pull request #298 from MaSzyna-EU07/feat/editor-tutorial-gifs
feat(editor): the clips of the tutorial lessons
2026-11-09 12:47:22 +00:00
71733a343a Merge pull request #299 from MaSzyna-EU07/fix/ui-tooltips-stay-in-window
fix(ui): tooltips and popups stay in the main window, and a long tooltip wraps
2026-11-09 12:46:56 +00:00
edea0dc0f5 Merge pull request #301 from pwarszawski/fix/python-screens-pillow
Python screens: ship Pillow with the Windows build, and the screen module for Pillow 10+
2026-11-09 12:46:42 +00:00
65468a2eae Merge pull request #302 from miloszunio/fix/gizmo-yaw-rotation-master
fix(editor): model yaw rotation with the gizmo jumps past 90 degrees
2026-11-09 12:46:24 +00:00
c0524552f1 Merge pull request #303 from MaSzyna-EU07/fix/legacy-renderer-instance-scale
fix(render): scale of model instances in the legacy and simple renderers
2026-11-09 11:39:41 +00:00
963d65192a fix(render): scale of model instances in the legacy and simple renderers build-fix-legacy-renderer-instance-scale 2026-11-09 11:23:46 +01:00
Claude
58e4bbff49 fix(editor): a model turned with the gizmo about the vertical axis no longer jumps back and forth
The angles of the whole gizmo matrix were decomposed every frame and the model was turned by their difference from its own.
The decomposition gives the angle about the vertical axis only within -90..90 degrees: past 90 it comes out as 180 - angle
with the other two axes turned by 180, so a drag one way turned the model back and flipped it. The turn is now taken from
the delta matrix of this frame (small, so its angles decompose with no ambiguity) and wrapped to -180..180.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DDcMn1Q7wdYkbxDUWscNHZ
2026-11-09 07:56:51 +01:00