Reported by a tester and reproduced here: finish a race, come back to the
lobby, start another, and the game dies a few seconds in.
It is a stack overflow, from CockpitShellProc calling itself. The cockpit
subclasses the game window to catch WM_SIZE and re-fit the canvas, and
kept SetWindowLongPtr's return as the proc to chain on to. But the game
window is not the cockpit's - it outlives it, and carries the console
screen from one race to the next - and nothing ever unsubclassed it. So
the second race subclassed an already-subclassed window, SetWindowLongPtr
handed back CockpitShellProc itself as the "original", and from the next
message onwards the proc chained to itself until the stack ran out.
Nothing in the log, because nothing in the game had gone wrong yet.
So the destructor puts the window's own proc back, and the install site
will not subclass the same window twice even if it could not.
While there: the destructor also left activeCockpit pointing at the
object it had just freed, so GetCockpit() handed CockpitShellProc a dead
cockpit to lay out. Harmless until someone resized or maximised the
window at the lobby between races, which is not a hard thing to do. Now
cleared.
This came in with the cockpit resize work in 6b43971, so every build
since has had it.
Verified under cdb: before, the crash is a c00000fd stack overflow with
CockpitShellProc / CallWindowProcA repeating the whole way down. After,
four consecutive races - launch, race, results, CONTINUE, lobby, launch
again - complete with no exception at all, and the process exits only
when asked to.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The volume keys wanted a partner, and the bass trim could not be one as it
stood: it scaled the sample data as it loaded, so by the time anyone pressed
a key the audio was already sitting in OpenAL buffers and nothing short of a
restart would move it.
So the trim is now a per-zone gain applied in the mix instead. Each buffer's
depth - how much of the low band it occupies - is still worked out once at
load from its playback rate, but the trim itself is read every frame, which
is what lets Home and End move it while sounds are playing. It is the better
form regardless: no rewriting of sample data, and no quantisation on top of
audio that has already been through one gain stage.
Home raises, End lowers, in steps of 0.05, and the setting is written to
bass.cfg beside the exe exactly as the volume writes volume.cfg. Together
with PageUp and PageDown that is the amplifier and the crossover the
cabinets had in hardware and a desktop does not.
Builds clean, runs, and neither knob fires unprompted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The cabinets had no volume control - they ran at unity and left level to an
external amplifier - so a player without that hardware had nowhere to turn
it down but environ.ini and a restart. PageUp and PageDown now step the
master volume by 0.05 while you play, from silent to double, and whatever
you leave it on is written to volume.cfg beside the exe and used from then
on. The environ.ini figure decides where a machine that has never been
touched starts out; the keys are the knob, and a knob stays where it was
left.
Page keys because they produce no typed character, so they cannot collide
with the character-keyed commands the engine already answers to, nothing
else in RP binds them, and they are on every keyboard including tenkeyless.
They are polled rather than read off the key-message path, which is worth
recording because the message path looked like the obvious home for them
and was tried first. RP's keyboard pump only takes WM_KEYUP, WM_SYSKEYUP
and WM_CHAR off the front of the queue, and the front end runs message
loops of its own, so key messages get raced for and lost: six deliberate,
well-spaced presses arrived as two. Fine for the abort chord, useless for
something you tap repeatedly to find a level. Reading key state directly
costs nothing and cannot be dropped. That losses figure is a pre-existing
property of the input path, not something this change introduced, and is
worth knowing before anything else gets bound there.
Builds clean, runs, and does not fire unprompted. The step function itself
is proven - it was driven end to end through the message path before the
switch, stepping the right way, clamping, and persisting. What I could not
test from here is the polling trigger, because Windows would not hand the
game foreground and injecting keys without it would have sprayed them
across whatever else was open.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The cabinets ran the game at unity and shaped volume and tone outside it,
in an external amplifier and a 3-way crossover. That is why there is no
master volume anywhere in the original code and none in AUDIO.INI - an
operator turned a knob on an amp. A desktop player has no amp and no
crossover, and the recovered soundbanks are a good deal livelier than what
4.12 shipped with, so the game has to offer the two controls the pod got
from hardware.
RP412AUDIOVOLUME, 0.0 to 4.0, is the amplifier: a listener gain, which the
port had never set at all. RP412AUDIOBASS, 0.0 to 1.0, is the crossover's
low band. Both default to leaving the mix exactly as the pod played it, so
neither changes anything for anyone who does not go looking.
The bass trim is not a filter, and the reason is worth writing down: the
OpenAL we ship is Creative's, not OpenAL Soft, and it implements only
AL_FILTER_LOWPASS. It rejects highpass and bandpass outright. A bandpass
would have been the tidy answer, carrying the authored brightness model on
GAINHF and the trim on GAINLF across the single direct filter a source
gets. It is not on offer.
So the trim scales sample data as it loads, which suits how this low end is
actually built: the weight lives in discrete deep layer zones whose per-zone
tuning bakes out to a very low playback rate - thirteen zones below 8kHz,
three to five octaves under their recorded pitch, against four fifths of the
set at 22kHz and up. Baked rate is a dependable proxy for band, so pulling
down the low-rate zones is a real low-band trim and not a blunt cut. It eases
in below 22kHz and reaches full depth at 5.5kHz.
Caught while building this, and the reason for the probe: EFX_Initialize
checks alGetError after configuring the scratch filter, so asking for a
filter type the driver refuses leaves an error pending and takes the entire
bridge down - reverb included. The bandpass attempt did precisely that and
would have silently killed the reverb and brightness work. Initialize now
survives losing the filter and says so.
Builds clean, runs with both knobs set and with neither.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Recovering the soundbanks took voice demand per sound from about one zone to
about two and a half, and the audio path allocated an OpenAL source for every
sound event and destroyed it again on release. Sources are a hard
per-context resource - this driver grants 256 - so that churn doubled at
exactly the moment it got more expensive. Sources are now generated once and
recycled through a free list: measured, three sources generated across
twelve thousand acquisitions.
The BT tree reached the same conclusion the expensive way, from field logs
full of failed acquisitions: raising the source budget is not the fix,
because the ceiling also acts as a governor and more voices mixing is real
CPU during exactly the busiest moments. Recycling is the fix, and it costs
nothing.
Two older bugs were sitting underneath, both reproduced against the driver
rather than assumed:
Releasing a set leaked it. alDeleteSources is atomic - one bad name in the
array and nothing at all is deleted. ReleaseSourceSet handed it the whole
fixed-size array and then parked the slots at -1, so any partial set, and
any double release, leaked every source it held. Sources are now handed back
one at a time and slots park at 0, which is never a valid name.
A source set began life uninitialised. The constructor set only the count,
and the acquire path decided whether a slot was already filled by asking
OpenAL about uninitialised stack garbage. Garbage that happened to match a
live name meant two sounds silently sharing one source. Pooling would have
made that more likely, not less, since it keeps small names in circulation.
Recycled sources are scrubbed before parking - stopped, buffer detached,
looping, gain, pitch, relative flag, position and velocity reset, and the
EFX filter and reverb send dropped. Without that last part a dry cockpit
sound inherits the wet send of whatever 3D source held the name before it.
Verified: a deliberately dirtied source comes back clean.
Builds clean. Runs with memory and handle count flat.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The OpenAL port kept the whole authored audio model and then threw most of
its output away. Every frame the engine computed a distance-attenuation
curve, a high-frequency rolloff, doppler cents, a reverb level and a
front/rear placement, and every one of those consumers had been commented
out when the two AWE32 cards were replaced. What reached the speakers was
OpenAL's own defaults instead: a straight-line fade to silence, no
filtering, doppler at the wrong constants with an inverted velocity, no
reverb, and every cockpit sound dead centre.
Restored, per AUDIO.INI, which is byte-identical to the file that shipped
in August 1995:
- the authored knee/rolloff distance curve, replacing AL_LINEAR_DISTANCE.
This also un-blinds the transient cull, the voice-steal weighting and
the mix ducking, which all key off it and were treating far sources as
full presence
- the CC7 squared volume law; writing the scale linearly ran everything
about 6 dB hot at mid-scale
- brightness and distance muffling, and the wet-exterior/dry-cockpit
reverb split, both through a new OpenAL EFX bridge
- doppler on the moving-source path only, as the original had it
- front/rear placement from the authored position enum
The larger find is that AL_PITCH was never called anywhere in the tree, so
the entire pitch chain was inert - not only doppler but pitch_mix_offset,
which our own sequences author 97 times. Doppler alone would have changed
nothing audible.
Note pitch is applied for parity with the BT engine but is identity here:
our content predates NoteAudioControlID, so every source runs at note 60.
Builds clean on VS2022 Release|Win32. Smoke-tested against vRIO on COM1 -
reaches gameplay and holds a steady frame loop. ALC_EXT_EFX is present on
the build machine with all nine entry points, so the filter and reverb work
is live rather than inert. Not yet listened to on the pod, which is the
real test: the volume law changes the level of everything.
docs/SOUND.md documents the original two-card quadraphonic design, where
the surviving original assets are, and what remains.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
BT411's f99003c, brought across. Its playtesters reported the cockpit
lighting going slow or stopping altogether while the 3D view stayed
smooth, and RP412 has the same structure exactly: the on-screen vRIO
buttons light themselves from PadRIO::GetLampState, but what FILLS that
store is lampManager->Update() in GaugeRenderer::ExecuteForeground - once
per full gauge cycle.
Which is the cycle the previous commit was about. Measured on a starved
frame budget it now completes 3.1 times a second, and completed 0.7
times a second before that; either way far too slow to carry a flashing
lamp. So sweep the lamps once per frame from the main render instead,
which runs regardless of how little frame is left over. It is cheap, and
AssertNewLampValue already drops anything unchanged, so this pushes no
extra traffic - it only stops changes arriving late.
Only when a PadRIO is active, i.e. cockpit-less play, and only while a
mission is actually running. With real serial hardware selected the pod
keeps its authentic bandwidth-paced cadence, untouched.
RP412LAMPSWEEP=0 restores the once-per-cycle behaviour.
BT411's other half, 02ce9f5, does not apply. That one is about Windows
throttling WM_TIMER and paint messages for background windows, which
made the glass panels' flash crawl whenever they did not have focus.
RP412 has no timer-driven repaint anywhere - the MFD windows are D3D
devices presented from SVGA16::Update, and the panel strips repaint from
there too - so there is no throttled message path to bypass. That path
was starved rather than throttled, and the previous commit is the fix.
Verified: no regression at either budget, 20.0 display sweeps/s at a
normal frame budget and 3.1/s starved, both unchanged by this commit;
mission runs clean. The lamp win itself is structural - the sweep is now
an unconditional per-frame call - and would want a busy multiplayer
mission to see directly.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two testers reported the map and the countdown clock freezing, one of
them only on larger, more complex maps, and one of them until a death.
Both details point at the same place.
The gauges and the cockpit displays are redrawn in whatever time is left
after the 3D view. The background loop is guaranteed a single pass per
frame and gets more only while time remains before the frame is due, and
one pass drew exactly one gauge. So a full sweep of ninety-odd gauges
needed ninety-odd passes - free when there is spare frame, but on a busy
map the 3D view eats all of it, the loop drops to its one guaranteed
pass, and a sweep takes ninety-odd FRAMES. Seconds. A death makes the
renderer skip every static object, the budget frees up, and the backlog
drains at once: the display appears to come back to life.
Worse, the copy phase that follows ended after a SINGLE display, so the
map - one of three - came round only every third sweep.
So: draw gauges to a 2ms slice rather than one per pass, which ties the
refresh rate to elapsed time instead of to how much spare frame there
happened to be; and copy every display before reporting the sweep done.
Measured on a deliberately starved frame budget, which reproduces the
reported symptom: 0.7 sweeps/s before, 3.1 after. At a normal budget
20/s, against 18-19 before - no cost to the healthy case. RP412GAUGESLICE
tunes the slice and 0 restores the old behaviour, which reproduces the
0.7 exactly. RP412GAUGEDIAG=1 logs the rate; watching the screen cannot
tell a display that has stopped refreshing from one whose picture simply
is not changing, which is what made this hard to see.
Also fixes the constructor calling Update() three lines before it
initialised mDisplayToUpdate, so the first pass indexed the D3D device
and surface arrays with whatever was on the stack.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Ported from BT411, which needed the same thing for its glass cockpit.
PadRIO reads XInput, which covers Xbox-class pads and nothing else. A
flight stick, a HOTAS throttle, a twist grip, rudder pedals or a wheel
arrive through DirectInput instead, and until now the game could not see
any of them - the only generic-joystick path left was the 1995 single-
device DIJoystick behind L4CONTROLS=DIJOYSTICK, which is untouched here.
L4JOY is the reader: up to four devices as normalized state blocks, hot-
plug re-enumeration on the same ~3 s cadence PadRIO uses to look for a
pad, and a device lost mid-race zeroed rather than left holding whatever
was pressed when it went. XInput-class devices are excluded by VID/PID
against the RawInput paths carrying the "IG_" marker - without that an
Xbox pad arrives through both APIs and every button counts twice.
bindings.txt gains four rows in the grammar it already had, using its own
vocabulary (deadzone/rate) rather than BT411's:
joydev <slot> [product-name substring]
joyaxis <src> axis <axis> [invert] [deadzone <d>] [rate <n>]
joybutton <n> button <addr> [toggle]
joyhat <n> <up|down|left|right> button <addr>
Slots resolve to a live device every poll, by name substring or ordinal,
so unplugging and replugging does not rewrite anyone's file.
Two things the pod's shape forced that BT411 solved differently:
Pedals - a signed composite axis that decomposes into the pod's two
pedals, positive right and negative left. The pod has a pedal each
side; a twist grip or rudder bar is one signed control, and pressing
one or the other but never both is exactly what it wants to say. It
is a channel name like any other, so a pad stick can drive the turn
too.
A joyaxis on Throttle with no rate is a real lever and OWNS the
channel - full travel maps onto the 0..1 the pod runs on, instead of
nudging the accumulator that a spring-centred pad stick has to use.
RP412JOYCONFIG=1 (joyconfig.bat) runs the capture wizard before the
console screen: it asks the player to move each control, and derives the
sign convention from the DIRECTION of the move. That is the point of it -
a stick that reads positive pushed right and one that reads negative are
equally common, and no amount of documentation gets a player to work out
which they own. It writes only its own section, between marker lines, so
hand-edited keyboard and pad rows survive re-running it.
The wizard also prints every axis at rest before it starts. A driver that
refuses the +-32767 range we ask for reports its own, and an axis then
sits hard over instead of near zero; seeing "X +1.00" on an untouched
stick is the difference between a five-minute fix and a bug report that
says it configured itself. Each capture reports the move it saw for the
same reason.
Verified on the Logitech Extreme 3D on this machine. Enumeration finds
it and excludes the Xbox pad, which still arrives separately through
XInput. Every row shape parses - 7 axes, 2 buttons, 4 hat directions -
and three deliberately malformed rows (a bad axis name, button 99, a
"sideways" hat) are each rejected by line number rather than silently
dropped. The wizard lists the device with its axes at rest reading
X +0.00 Y -0.01 RZ -0.04 SL0 +1.00, waits on the first prompt without
self-triggering, and with a hand on the stick captures X to steering,
Y to pitch, RZ to the pedals and SL0 to the throttle, inverting the ones
that read backwards.
Running the captures through to a written file needs a hand on the
stick, so that part is the machine's to confirm, not this build's.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
It is a draggable top-level window whenever L4PLASMA=SCREEN, which makes
it the last one still being placed fresh every launch. It joins
mfd_layout.cfg under "Plasma Display", the caption it already carries -
the same key-is-the-title rule the display panes follow.
Position only, like the panes: the glass is 128x32 at L4PLASMASCALE, so
its size is a setting rather than something to drag.
It registers and loads at the point it creates its window rather than
leaving that to SVGA16. The glass comes from the gauge renderer and the
panes from the video mode, and nothing guarantees which is built first;
loading in both places means whichever runs second simply re-applies a
placement the first already has. Its window procedure picks up the same
WM_EXITSIZEMOVE save the panes have, so a drag writes the file straight
away, and the destructor forgets the window before destroying it.
Verified by round trip in the exploded view: dragged to 640,880, the
file took "Plasma Display=640,880,528,167" alongside the panes and the
game window, and a fresh launch in load mode put it physically back at
640,880. The three load lines in the log - one window, then two, then
eight - are the ordering doing its job.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
,noframe takes the title bar away, and with it the only way out of the
game. The console screen now offers its own, bottom left: half width and
diagonally opposite LAUNCH GAME, because it is the one button on that
screen you cannot undo and it should not sit next to the one everybody
is aiming for.
It goes through the same door as closing the window - fe.closed, so
RPL4FrontEnd_Run returns False and the race loop breaks - rather than
opening a second shutdown path.
Two things had to move for it to make sense.
The saved placement now loads in RPL4.CPP, right after the main window
is shown, instead of only when SVGA16 builds the cockpit. That was not
until a mission started, so the console screen came up at the default
rect with its title bar still on and the window only jumped to the saved
placement once a race began - which, for a flag whose whole purpose is
to take the title bar off, meant it did nothing on the screen you land
on. RPL4.CPP now owns the main window's registration outright and
SVGA16's branches only reload; the reload after CockpitShellProc goes on
still matters, since its WM_SIZE is what re-fits the canvas.
That in turn made the exploded view's position-only registration
incoherent - the startup load had already applied the size - so the
game window is simply position and size everywhere now. The earlier
reasoning that its exploded size IS the -res render size does not hold:
the back buffer stretches to the window in either view, exactly as it
does for the cockpit.
WM_EXITSIZEMOVE moves from the cockpit subclass to RPL4.CPP's own
WndProc, which the subclass chains to anyway. In its old home it only
existed once a cockpit had been built, so dragging the window on the
console screen - the obvious moment to put it where you want it - saved
nothing. There is also a save on the way out of WinMain, for a session
that never started a race and so never ran SVGA16's teardown save.
Verified: on a bare-framed window the console screen comes up at the
saved 1280x760 with client == window rect, EXIT GAME ends the process
with code 0, and a screenshot shows it clear of the column content. A
console-only session dragged to 333,222 900x640 wrote that on the drag,
kept it through the exit, and came back to exactly it on relaunch -
without a race anywhere in the round trip. The noframe and cockpit
round trips still pass unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Append it to any line - "RPL4=240,120,1000,620,noframe" - and that
window comes up with no caption and no border. For the game window that
is a cockpit filling a monitor edge to edge at a rect you chose, which
-fit could only do by taking the whole screen; for an exploded pane it
is a display photographed without chrome.
Per line rather than global, so the shell can go bare while the panes
keep their captions, or the other way round.
The flag is an instruction rather than something measured off the
window, so Save carries it back out - otherwise the first finished drag
would rewrite the file and quietly drop it. Windows are always built
framed and Load only ever strips, so deleting the flag is all it takes
to get the frame back; there is no un-strip path to get wrong.
A bare window's rect IS its client rect, so the client area is what
survives: a window that had a size in the file keeps it as the client,
and a position-only pane keeps whatever client it had. That also makes
the round trip stable - once bare, what Save records is already the
client, so load-save-load does not creep.
WS_SYSMENU stays on. It draws nothing without a caption, but without it
DefWindowProc will not honour Alt+F4, and a window with no title bar and
no way to close it is a trap. Nothing else can be dragged either, hence
the note in the file header and environ.ini: place it first, add the
flag after.
Verified in both views. Cockpit: the same 240,120 1000x620 line with and
without the flag, CAPTION|THICKFRAME and a 984x581 client becoming POPUP
with a 1000x620 one, and a screenshot showing the displays hard against
all four edges where the framed shot had them inside a letterbox. Save
mode with the flag set, nudged with WM_EXITSIZEMOVE, rewrote the line
with ",noframe" intact. Exploded: Map bare at 777,333 with its 500x640
client preserved while the shell beside it kept its caption.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
RP412MFDLAYOUT already kept the exploded view's display panes where they
were dragged. The main window is the one people move most, and it was
still being placed fresh every launch, so it joins them.
MFDSplitView_LoadLayout/SaveLayout become RPWindowLayout_Load/Save, with
Register/Forget taking any HWND rather than the module reaching into a
pane registry. Same file, same format, one more line in it.
What comes back depends on the window, so Register takes it as a flag:
display panes position only, as before. A pane's size follows its
content and its button banks, so an old size from a
different build must not distort it.
the game window position and size in the cockpit view. Nothing
derives that size - the cockpit fits itself to
whatever client area it is given - so a window sized
to suit a monitor should come back that way, and
half-restoring it would be the strange behaviour. In
the exploded view its size IS the render resolution
-res asked for, so there only the position returns.
Registered after the CockpitShellProc subclass is installed, on purpose:
the restore's WM_SIZE then runs LayoutCockpit again and the canvas
re-fits the restored client area. -fit does not register at all - it
owns the whole monitor, so there is no placement of the player's to
keep.
CockpitShellProc gained the WM_EXITSIZEMOVE hook the panes already had,
so dragging or resizing the shell writes the file immediately rather
than waiting for teardown.
Two hazards the panes were small enough to get away with and the game
window is not:
- Save reads rcNormalPosition rather than GetWindowRect. A minimised
window reports a nonsense rect and a maximised one reports the
screen; since the file is rewritten whole, either would have
replaced a good line with a useless one. rcNormalPosition is the
restored placement whatever state the window is in.
- Load drops any placement that intersects none of the monitors
currently plugged in. Restoring the game window onto a display that
is no longer there would leave nothing to drag back.
Verified by round trip in both views. Cockpit: dragged and resized to
240,120 1000x620, the file took it, a fresh launch in load mode came up
exactly there with a 984x581 client - and a screenshot confirms the
canvas re-fit it, displays at the corners and the map centred at the
bottom, nothing spilling. Exploded: the shell came back at 60,60 still
1280x720 from -res while Map came back at 777,333, which is the
size-flag split doing its job.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Ported from BT411's BT_GLASS_LAYOUT (29c502d).
The exploded view's panes are draggable desktop windows, but the
arrangement is recomputed on every launch, so dragging one somewhere
useful never survived the menu-race-menu loop.
RP412MFDLAYOUT persists it to mfd_layout.cfg beside bindings.txt:
off / 0 / unset computed arrangement only, no file (default)
load / restore restore saved positions at startup, never write
save / adjust restore, then rewrite on each finished drag
(WM_EXITSIZEMOVE) and on teardown
One "<title>=x,y,w,h" line per pane. Position is restored and the size
read and discarded: a pane's size follows its content and its button
banks, so letting an old size back in would misshape it after any
geometry change - and this port has changed that geometry twice already.
Load runs after the computed arrangement rather than instead of it, so a
pane the file does not mention simply keeps its computed spot. Only the
exploded view registers: the composited cockpit's panes are chrome-less
children with nothing to drag, so they have no position worth keeping.
RP412 needs no equivalent of BT411's "restored" flag. Its re-snap is
LayoutCockpit on WM_SIZE, which only runs in cockpit mode, so nothing
comes back later to overwrite a hand-placed window.
Verified by round trip: dragged Map to 777,333 in save mode, the file
took all six panes, and a fresh launch in load mode put it physically
back at 777,333. The harness also resized the window while moving it,
which incidentally proved the saved size really is ignored - the pane
came back correctly sized from a cfg that recorded 136x39.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The plates beside each spot came out blank. They ask for textures called
player1..player8, which are not files - the renderer draws each pilot's
callsign into a texture at run time - so the load failed and left them
untextured. Nothing bound the two together.
SortAndReloadNameBitmaps already builds those textures indexed by
finishing place, which is exactly how the plates are numbered, so the
plate beside each spot wants mNameTextures[place]. Binding them is the
whole fix, but it takes two steps rather than one.
The plates have to survive mesh consolidation first. Static geometry is
merged with D3DXConcatenateMeshes and its draw ops deduped by material -
and eight failed texture loads leave eight identical untextured ops, so
all eight plates collapse into one that could only ever carry a single
name. Each plate now gets a distinct 1x1 marker texture as it loads,
which keeps it a subset of its own. The marker is never seen.
Then the binding runs against the consolidated mesh, not the objects the
plates were loaded from - by podium time those have been merged away and
are no longer drawn, which is why re-pointing them changed nothing.
Verified on a race: 8 plates found in the consolidated mesh, 1 bound,
and the winner's plate reads their callsign under their vehicle. One
bound of eight is right for a single-pod race - the rest of the places
have nobody in them.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The reticle was still drawn over the podium. It goes out in the 2D pass,
after everything the presentation turns off, so it survived. The race is
over and nothing is being aimed at.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Your own vehicle is built insideEntity - a cockpit and no hull, because
you are sitting in it and never see it. That is fine for a race and wrong
for a podium: from the presentation camera your spot on the stand was the
one that was empty, and on a single-pod race that is the whole picture.
The renderer now gives the viewpoint entity an exterior before the shot.
Disconnected_Eye is the engine's own switch for this case, documented as
being there "so higher level renderers can fix the eye in one spot and
watch the viewpoint entity drive around", and it is what makes
NotifyOfNewInterestingEntity choose outsideEntity.
The exterior is added alongside what is already there rather than by
tearing the entity down and rebuilding it. Teardown-and-rebuild is the
path the interest manager uses constantly for scenery dropping out of
range, so it looked safe, but it is not safe for the viewpoint entity:
that one is never uninteresting, the eye renderable goes down with it,
and doing it mid-mission stops the scene rendering at all - the screen
went black from the moment the podium came up and never came back.
Adding the renderables directly does the same job with nothing removed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The podium arrived as a hard cut: the race was still on screen one frame
and the stand was there the next. The race has its own fade-to-black
already, and it was being suppressed to keep the fade from blacking out
the podium behind it - which threw away the transition along with the
problem.
Now the two are sequenced. StopMission lets the race fade out as it
always did and posts the podium to itself for when that fade has landed
on black; the handler stands the finishers up behind the black and ramps
back in. The fade-in is the end-of-mission fade run backwards - the same
multiply on the fog colour and both fog distances, from nothing up to
what the Winners Circle asked for.
Timings: 0.7s of fade-out and black, then a 0.45s fade in. Both come out
of the 11 second hold, leaving about ten seconds of podium.
RP412PODIUMFADEIN sets the ramp.
Verified by measuring frame brightness across the transition. The race
falls away and the screen reaches black, then the stand comes up - and
with the ramp stretched to 3s to make it resolvable at a half-second
sampling interval, it climbs 71.8, 77, 78.7, 79.7, 80.5 rather than
stepping, so it is a real fade and not a cut arriving late.
The camera also comes down and tilts up across the tiers, which is how a
podium wants to be shot. It is a balance in both directions: drop it
further or tilt harder and the sky takes the top half while the winner's
spot slides off the bottom of the frame; tilt down instead and the shot
turns into a floor plan.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The stand was composed to be looked at from a 4:3 pod monitor. On a 16:9
canvas its platform runs out at the sides and the shot fills up with sky
and void, so the podium is now pillarboxed: the scene renders into a
centred 4:3 viewport with the surround left black.
The projection has to use the cropped shape too, or the scene comes out
squashed into the narrower viewport instead of cropped by it.
RP412PODIUMASPECT overrides the ratio, 0 turns it off.
The camera also came in closer, from 33 units rather than 45, and now
aims slightly below the group rather than above it. That tilt is what
buys back the sky above the grandstand - aiming above the group tips the
camera up instead and walks the winner's spot off the bottom of the
frame, which is the one position that has to be in shot.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The six secondary displays sit over the viewscreen like the pod's bezels
and have nothing to say once the race is over. Worse, the radar sits dead
centre along the bottom edge - directly on top of the winner's spot, so
the one position that matters was the one you could not see.
They are hidden when the podium comes up, which uncovers the canvas the
3D is already being drawn on. No matching show: the mission is over by
then, and the next race builds a fresh cockpit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The pod hall stood the finishers on a numbered platform when the race
ended. All of it shipped in this repo and none of it ever ran here: the
stand geometry, the eight ranked dropzones win1-win8 in every one of the
11 maps, and the sequence that places the racers on them. The sequence
lived on RPL4PlaybackApplication - the mission-review build - behind a
spool file, so the app the pods actually race has never called it.
RPL4Application now has its own StopMission handler that ranks the
finishers, drops each onto their spot, freezes them, re-sorts the name
plates into finishing order and frames a camera on the stand. It fires
once: StopMission arrives twice, from the console at the buzzer and again
from the player when the ending fade expires, and only the first is the
end of the race.
Ranking works in football as well as a race. CalcFootballRanking ranks
only the RunnerPlayers group, which would have placed the runners and
stopped - but nothing calls it. What runs is Player::CalcRanking, every
frame, over every scoring player by score.
Three pieces of the original had been stubbed out in the D3D9 port and
are restored:
SetViewAngle was an empty function, so the 45 degrees the sequence asks
for did nothing. It now rebuilds the projection the way DPLReadINIPage
does and pushes it, and sets viewRatio, which nothing had written since
the DPL body was commented out.
winnersCircleFogStyle was an empty case. The stand sits far off the
track in open ground where the track's own fog leaves it dark; this is
the blue-violet the original used, with the fog pushed back to 100/1050
and the clip plane pulled to 1100.
The end-of-mission fade had to be told to stand down. It multiplies the
fog colour and both fog distances toward zero every frame - correct when
a race just ends, fatal to anything shown afterwards. That fade is what
made the podium a black screen, and it took a while to find because
every frame was being built and presented correctly the whole time.
The presentation camera overrides D3DTS_VIEW between the eye renderable
writing it and ExecuteImplementation reading it back for the draw calls,
so no CameraShip is needed. It builds with LookAt LH, not RH: the
projection is LH, and RH aims the camera the opposite way - ask to look
down at the stand and you get the sky behind you. The engine's own eye
renderable is right to use RH, because its forward and up come out of the
entity matrix already in that convention.
The mission is held open 11 seconds rather than 3. That fade timer is the
only thing keeping the simulation and the renderer alive once the race is
over, and it has no upper bound on the ending path.
Switches, all off-by-default behaviour aside: RP412PODIUM=0 skips it,
RP412PODIUMCAM=0 keeps the cockpit view, RP412PODIUMSTANDOFF/HEIGHT/AIM
frame the shot, RP412MISSIONSECONDS overrides the menu game length (the
shortest it offers is 3:00, a long wait when what you are testing is the
buzzer), and RP412RENDERDIAG=1 reports what a frame is made of.
Verified end to end on Wiseguy's Wake: the stand, its tiers, the blue 2
and 3, the red 4 through 8 and all eight name bays, held steady for the
full 11 seconds and then handing off to the results screen.
Known gaps: the name plates are blank, because the player1-8 textures are
runtime name bitmaps that do not resolve as files in this port, and your
own vehicle has no exterior model - you see the others, not yourself.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
L4VB16.h carries GlassSize and CockpitGlassSizes, which L4VB16.cpp uses
in both the constructor and LayoutCockpit - 8dc6605 left it behind (the
add named the path in the wrong case), so that commit does not build on
its own.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Button banks
The exploded diagnostic view was still display-only - it predates the
button work - so it now builds the same banks as the cockpit, with
the pod arrangement laid out from the panes measured sizes rather
than a hardcoded 640/480 grid (the banks make each window bigger than
its glass, and the bottom row hung off the work area otherwise).
The map side columns were spread height/6 from the top, but the maps
own legend grid is not sixths: measured off the bitmap it starts 13
rows down with six 102-tall cells on a 105 pitch. Every button sat
high of its label, worst at the bottom. Each buttons top and bottom
now come off that grid separately and are subtracted - scaling a
height directly would let rounding drift them back out of step on a
resized cockpit.
Depth 100 to 240: against the 480 glass the two banks meet in the
middle bar the strips, so practically the whole display is a press
target. This mattered most in the cockpit, where the panes are small
enough that the halfway clamp governs - at 100 the MFDs had a 110px
dead band straight through the middle of the glass.
-fit (also spelled -windowed-fullscreen)
Borderless over the whole monitor, with the render size chosen to
match. The cockpit presents the 3D into a viewscreen that fills its
canvas, so the right -res is that canvas at the scale the cockpit
will settle on; computing it with identical arithmetic makes the
stretch a copy. On the 3440x1440 panel that is 133% and -res 2553
1436, against 125% for the windowed path that pays for the taskbar.
The pick runs after the whole command line, so an explicit -res wins
from either side of -fit. Capped at 3840x2160. Cockpit mode only -
mode 0 has to stay playable on real pod hardware and mode 2 is a dev
view - so those get the resolution and keep their windows.
Display layout, in environ.ini
L4MFDSCALE sizes all five MFDs, L4MFDSCALE_UL and friends override
any one of them, L4RADARSCALE the radar, and L4RADARPOS puts the
radar bottom centre, in either bottom corner, or halfway up either
side. Scaling is applied in canvas units before the canvas is fitted
to the window, so a number means the same thing on every monitor.
Sizing each display separately let the clamps become exact rather
than one conservative rule for all five: what limits a display is its
actual neighbour. Which neighbour that is depends on the radar, so
the clamps follow it - on the bottom edge it clears the one MFD above
its column, but centred on a side it has one above AND below and
grows from the middle both ways, so it must clear the taller twice
over. Clamping shrinks uniformly; these are photographs of real
instruments and a one-axis clamp would squash them.
Verified on the ultrawide: all five radar positions, per-display and
group scaling with the clamps biting, -fit with and without an explicit
-res, and the button geometry measured back off the screen.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The canvas was capped at 100%, so a bigger monitor got a 1920x1080
cockpit in the corner of the screen and maximising did nothing - the
layout only ran once at startup.
Now the fit is one uniform scale with no ceiling, recomputed whenever
the window changes size (maximise, restore, drag), and the canvas is
centred in whatever client area it gets. Uniform scale is what locks
the aspect: a wider-than-16:9 desktop letterboxes with even black bars
instead of stretching the cockpit. Panes gained Resize() so the glass
and its button banks re-scale with the canvas; the pixel buffers keep
their native source resolution.
Scaling up is free quality on the MFDs - their glass is a downscale of
a native 640x480 channel until about 200%.
Verified on the 3440x1392 ultrawide: opens at 125% (2400x1350) instead
of 100%, maximises to 126% centred with even bars, a 1884x661 window
fits 61% letterboxed left/right, 1084x961 fits 56% letterboxed top/
bottom, and a 2584x1461 client scales up to 134%.
The 3D still renders at -res and D3D stretches it to the canvas, so
raise -res to match a large screen for 1:1 pixels; start-windowed.bat
and the README say so.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The amber Secondary/Screen columns get the same treatment as the red
MFD banks, turned on its side: each button reaches 100px in behind the
map with a 10px indicator clearing the edge, so the lamp reads as a
slim column and the radar picture is the click target. The shared
buttonDepth/indicatorStrip constants now drive both banks, and the
columns are clamped so they can never meet behind a narrow map.
The map pane narrows from 404 to 344 for a 324-wide glass, handing the
difference back to the viewscreen.
Verified live: clicking 60px inside the radar from either edge presses
the column button behind it, with feedback in that side's strip (720
and 1350 pixels).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The red banks were thin strips sitting outside the glass, so the click
target was only as tall as the strip. They are now 100px buttons that
extend BEHIND the MFD picture with just a 10px indicator clearing the
edge: the lamp still reads as a slim strip along the bezel, but the
region of the display above or below it is what you press. Paint order
flipped to match - buttons first, glass over them.
The banks are clamped so they can never meet in the middle on a
scaled-down cockpit, and the panes lost 40px of height each (260 vs
300 for a 240-tall glass), which hands that space back to the
viewscreen.
Verified live: clicking 50px inside the picture presses the button
behind it and the press shows in the indicator strip (770 pixels
changed, all of them within the strip).
Amber map columns are untouched for now.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
They were mapped to the config buttons (0x37/0x36), which the 9 and 0
keys already reach from the Upper Right MFD bank. Freeing them gives
players two pad buttons to assign; bindings.txt carries the mapping as
commented example lines so the way to re-enable it is in the file.
Docs and the dist README follow.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two pieces of MFD work that had been sitting in the tree (already in
every build and dist since).
The compact cockpit glass shows the full 640x480 gauge canvas at about
half size, and COLORONCOLOR did that by dropping every other row and
column - which shredded the 1-bit vector strokes and small text.
HALFTONE area-averages instead (with the brush origin set, as MSDN
requires) so the downscaled MFDs stay legible.
L4MFDSPLIT=2 adds an exploded diagnostic view: every display in its
own full-size desktop window at native resolution, decoded exactly as
the pod VDB split them from the single gauge canvas, with no cockpit
compositing and no downscale. It makes an individual MFD readable and
screenshottable at full resolution for comparison against the
emulator per-channel reference windows. L4MFDSPLIT=1 keeps the
composited glass cockpit and stays the default.
Also: .gitignore now covers the packaged RedPlanet-*.zip releases,
which live on the Gitea release page rather than in the tree.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The map/radar glass grows from 240x320 to 360x480 on the 1920x1080
canvas (user request - the radar reads best big). The flanking
Secondary/Screen columns, bottom-center placement, and canvas scaling
all follow from the measured size automatically.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The polish-backlog item, implemented from vRIO KeyboardLampMirror:
game-commanded RIO lamp states paint per-key RGB keyboards through
Windows Dynamic Lighting (WinRT LampArray). Keys bound to lamp
addresses in the active bindings profile glow with the panel palette
(red banks, yellow Secondary/Screen columns), flash modes use the
exact L4MFDVIEW formula so keyboard and on-screen buttons blink in
step, unbound keys are blacked out so the board reads as the button
field, and zone-lit keyboards fall back to a board-wide mirror of the
strongest lamp. Advantage over vRIO: Dynamic Lighting grants LEDs to
the FOREGROUND app - which is the game - so no Windows settings
dance.
Isolation: L4KEYLIGHT.cpp compiles /std:c++17 + DEFAULT packing +
conformance (per-file vcxproj settings; the engine /Zp1 would break
the WinRT ABI) with a scalars-only interface, and all WinRT work runs
on a private worker thread (watcher, claiming, 100ms paint loop).
On by default with a bindings map present; RP412KEYLIGHT=0 opts out;
missing Dynamic Lighting logs once and stays dormant.
Verified live on the dev laptop: claimed its 24-zone keyboard
(board-wide mirror) during a race; race cycling with per-race
start/stop of the mirror thread stays green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Per the user: the game never reads the pilot keypad, so the numpad
becomes the flight cluster - 8/2/4/6 stick, 7/9 pedals, 0 trigger -
with Shift/Ctrl as throttle up/down and Alt as reverse thrust. That
frees the entire letter board: W/A/S/D/Q/E return to their printed
MFD bank positions and B goes back to being the gap key, so the
default profile now carries the complete unmodified vRIO bank layout.
The keypad addresses stay bindable (arcade key events) but ship
unbound.
Alt as a held flight control meant every release popped the window
menu and stole focus - WndProc now eats SC_KEYMENU. Shift/Ctrl/Alt
key-name aliases added to the parser alongside the .NET names.
Profile parses clean (59 key buttons, 8 key axes); single-player
cycle and key-bomb tests green (Alt+Q abort unaffected).
Machines with an existing bindings.txt keep their old map - delete
the file to take the new defaults.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PadRIO now loads bindings.txt (vRIO profile grammar: key/pad/padaxis
lines with toggle, deflect/rate, invert/deadzone options) written
self-documenting with the full default layout on first run. The
default is vRIO board-complete map - number and letter rows are the
MFD banks as printed on the panel, F-keys the secondary/screen
columns, numpad the pilot keypad (0x50-0x5F delivered as arcade RIO
KeyEvents, a new PadRIO capability), Space/arrows the joystick
column - with the desktop driving keys carved out: WASD stick, Q/E
pedals, PgUp/PgDn throttle, B reverse (vRIO gap key; R returns to
its bank). Pad bindings unchanged in spirit, plus Panic on LB and
config on Start/Back; axis signs are encoded in the profile now, so
L4PADFLIP flips on top of it.
Default profile parses with zero rejected lines (68 key buttons, 8
key axes, 12 pad buttons, 5 pad axes); single-player cycle and the
key-bomb tests stay green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Full input-surface audit of the PC keyboard channel. The arcade abort
was the typed ampersand character (0x26), but the Win32 port feeds
WM_KEYUP VIRTUAL-KEY codes into the same channel - and VK_UP is also
0x26. The hat look-up key aborted the mission: that was every mystery
abort across the multiplayer test rounds. Likewise the E key (0x45,
the right pedal!) dumped the event queue on every release.
Disarmed: plain 0x26 and E are swallowed at the L4 layer; the abort
answers only to the deliberate Alt+Q chord (translated to the legacy
engine code, so APP.cpp is untouched); the debug toggles (wireframe
Alt+W, predator vision Alt+V, frame dump Alt+F, perf stats, event
queue on Alt+E now) arm only with RP412DEVKEYS=1. The cheat-string
manager has no PC-keyboard strings and the trace-log keys are
compiled out of release - both inert.
Verified live: a volley of VK_UP releases mid-race leaves the mission
running; Alt+Q aborts on demand. Docs and the dist README updated.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round six raced all three machines (staging fix confirmed) and then
exposed what happens when a pod leaves mid-mission - which arcade pods
never did.
The B crash dump named it exactly: VTV::TakeDamageMessageHandler
resolved message->inflictingEntity to NULL (the entity belonged to the
departed owner) and dereferenced it - Verify is compiled out in
release. Collision damage from an entity that no longer exists is now
ignored.
And the race B and C were left in was a zombie: the owner (console)
had aborted, so the mission clock would count up forever and the
death/respawn flow hung with nobody to arbitrate. Lobby-member races
now set gConsoleLossEndsMission: losing the console mid-mission posts
StopMission locally, the pod tears down, and lands back in the lobby
room. Arcade -net pods keep the re-listen-and-wait behavior.
Loopback hosted race still green.
For the drivers: the ampersand key is the arcade mission-abort - that
was every crash-on-keypress so far; and a sleeping Bluetooth pad wakes
on the Xbox button and hot-connects within 3 seconds (PadRIO
re-probes).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round four reached one step from the race: eggs delivered, both
members ACKed with complete meshes, one member entered LoadingMission.
Two failures remained, both now fixed.
One: Steam reports incoming callers under locally-allocated ALIAS
FakeIPs, not their global ones - the owner accepted both mesh legs
but its identity check compared the alias against the egg address and
never counted the connections (no Connected to GameMachineHost on the
owner). The peer table now carries each member SteamID (lobby go
roster gained a field) and Accept resolves the caller identity back
to the global FakeIP the egg promised.
Two: mission load stalls the game thread for 10-30s with nothing
pumping, and Steam default 10s connected-timeout sheared every
connection mid-load (end reason 4001, rx ages 11.5-20.5s - right at
load duration). Connected timeout is now 90s; TCP never timed out an
idle arcade link and races pump every frame once running.
Self-test still green (ping, survives listener close).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round-three diagnosis from the three-machine logs: the console
channel connected first try on SDR both ways - then died seconds
later with end reason 5010 (PeerSentNoConnection). Cause: the arcade
engine closes its console LISTENER the moment the console connects
(correct under TCP, where accepted sockets outlive the listener), but
Steam''s CloseListenSocket closes all accepted connections
ungracefully. The members silently killed their console connection at
accept; the owner''s next packet got no-connection back, no eggs were
ever fed, and all three pods sat waiting.
The transport now bridges the semantics: an engine close only marks
the listener (new callers are rejected, pending queue dropped); the
Steam socket is destroyed in Cleanup at mission teardown. Re-listening
on the same engine port reopens the marked listener.
The loopback self-test now covers exactly this: accept, close the
listener, then push data both ways over the accepted connection -
PASSED (ping 1, survives listener close 1).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Root cause of the three-machine failure found and fixed. The
connection-status callback was registered as a per-connection config
value pointer - which the steam_api flavor of the library never
dispatches (only the standalone lib does). So members never saw the
incoming connection request, never accepted, and the owner timed out
after 120s of state-5 retries. Valve''s own SpaceWar example registers
SteamNetConnectionStatusChangedCallback_t with STEAM_CALLBACK; the
transport now does the same through a CCallback listener constructed
after SteamAPI_Init.
Proven with the new RP412STEAMSELFTEST=1 loopback: one machine
listens on its fake console port and dials its own FakeIP - the log
now shows incoming/accept/connected on both ends and a verified data
round trip, connect succeeded on attempt 1.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The three-machine test failed at exactly one step: the owner''s
ConnectByIPAddress to a member''s fake console port retried silently
for 120s while the member WAS listening on that port. The transport
threw away Steam''s reason for killing each attempt.
Now every connection state change logs the connection description,
and drops log the end reason plus Steam''s debug string - the next
run will say exactly which subsystem refused (cert, FakeIP directory,
routing, accept). Connect also polls GetConnectionInfo directly
instead of trusting only the callback flags, logs each attempt''s
final state, no longer holds a pointer into the connection table
across mutations, and paces retries at 1s.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
RPL4LOBBY implements the multiplayer front door on ISteamMatchmaking.
The setup menu grows HOST STEAM RACE / JOIN STEAM RACE buttons when
the Steam wire is live; hosting creates a tagged public lobby, joining
finds one. Every member publishes FakeIP + fake ports + persona +
loadout as member data; the room screen lists members (host marked)
and gives the owner a launch button.
Launching writes a nonced go-roster into lobby data. Each pod
registers every peer with the Steam transport (two-port peer table:
engine console/game ports map to Steam fake ports on connect) and
enters the race: the owner through the hosted-race path - it builds
the multi-pilot egg from real personas and loadouts and its console
marshals everyone - and members as network pods that boot straight
into WaitingForEgg for the owner to feed over the wire.
The lobby outlives races: members loop back through WinMain into the
room (no local console needed - MissionCompleted is waived for member
races), and the owner returns to the room after its results screen.
Leaving the lobby clears the hosted-race priming.
Verified on this box: menu buttons appear under RP412STEAM=1, hosting
creates a lobby on the Steam backend, the room runs and leaves back to
the menu; single-player cycling and the LAN hosted race both still
pass. Full three-account mesh test is next, on real hardware.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The lobby-owner-as-console architecture, in-engine. RPL4CONSOLE gains
RPL4LocalConsole_InstallNetworkRace: the owner pod switches to network
mode and meshes like any pod (egg fed locally via FeedLocalEgg, which
now opens the ConsoleOnly state gate), while the console tick also
marshals REMOTE pods over NetTransport speaking the exact arcade
protocol - egg chunks with 5s-retry-until-ACK, 1Hz state polling,
RunMission once every pod stages at WaitingForLaunch, StopMission at
expiry (remotes first, local pod holds until their EndMission scores
land), score intake labeled with [pilots]-order names on the results
screen.
The front end builds the multi-pilot egg: RP412HOSTPODS lists member
console channels (lobby stand-in; the Steam lobby feeds the same
path), RP412HOSTPORT/RP412HOSTADDR set the owner side.
Winsock Connect now redials with a fresh socket per attempt (a refused
TCP socket is dead; the old loop reused it) bounded at 120s - needed
whenever a peer boots after the caller, which is the normal Steam
lobby launch order.
Verified on loopback: member pod in -net, owner hosting from its menu;
mesh completed both sides, 30s race, remote score collected over the
wire (host 3), local stop after the drain, results screen shows both
pilots by name in one process that returns to the menu.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Steamworks SDK 1.64 vendored at extern/steamworks_sdk_164 (headers +
win32 redistributables only; .gitignore trims the rest). Both projects
build with RP412_STEAM; activation stays behind the RP412STEAM=1
environment switch, so plain desktop runs never touch Steam.
L4STEAMTRANSPORT.cpp implements NetTransport on ISteamNetworkingSockets
with FakeIP: SteamNetTransport_Install brings up SteamAPI, relay
network access, and a two-port FakeIP identity (fake port 0 = console
channel, 1 = game mesh), then swaps the process wire; any failure logs
the reason and the game carries on over TCP. Addressing keeps the
engine untouched: all pods share the -net port convention, eggs carry
fakeip:engineport, and the transport alone translates engine ports to
Steam fake ports via the lobby-fed peer table (RegisterPeer). Connect
mirrors the TCP retry-while-refused loop; Receive normalizes message
lanes back into the stream semantics CheckBuffers expects.
Runtime verified on this box: RP412STEAM=1 under AppID 480 came up as
169.254.59.52 (fake ports 32256/32257); without Steam credentials it
falls back to TCP cleanly; default boot logs no Steam lines at all.
steam_api.dll ships in the dist.
Next: the lobby layer (ISteamMatchmaking member data -> RegisterPeer +
egg build + RPL4CONSOLE marshal), which needs a second account to test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Steam-multiplayer prerequisite from the design doc: L4NET keeps
hosts, message queues, and the deterministic mesh ordering, while
connect/listen/accept/close, send/receive, startup/cleanup, the local
interface list, and ip[:port] parsing move behind NetTransport
(L4NETTRANSPORT.h). WinsockNetTransport carries the existing TCP
behavior over verbatim - including the connect retry-while-refused
loop the egg-ACK ordering relies on - and is the process default;
NetTransport_Set installs a replacement before the network manager
comes up.
L4STEAMTRANSPORT.h documents the ISteamNetworkingSockets mapping
per method (FakeIP keeps [pilots] entries as IPv4 strings) behind
RP412_STEAM until the Steamworks SDK lands.
Also fixed in passing: the ExclusiveBroadcast path sent
sizeof(network_packet) - four bytes of pointer - instead of the
message size, which would have sheared the stream framing had it
ever fired.
Verified: single-player race cycle (menu, race, results, menu, race)
green; -net 8000 boots, listens for a console through the transport,
and idles stable.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
WinMain now wraps the engine block in a loop: when a front-end-launched
mission ends under the local console, the setup screen comes back in the
same process instead of exiting (the arcade relaunch-per-mission model).
Replaces the CreateProcess self-respawn - required for Steam, where the
lobby and sockets must survive across races.
Second-cycle re-init crash fixed: d3d_OBJECT kept a static texture cache
keyed by filename, so race 2 got IDirect3DTexture9 pointers created on
race 1 destroyed device and died at first draw (DrawMesh AV). The cache
is now flushed in ~DPLRenderer before the device is released, and
ParticleEngine::Initialize drops particles left over from the previous
mission. Verified: three consecutive 30s races in one PID, each stopped
on time by the console with final scores collected.
Also: L4CONSOLELEN env override for test-length races, and the console
exposes MissionCompleted() for the loop.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Domain correction from playtest: hand-fed eggs are a developer shortcut
- a mission only ends on a console command, so the clock hits 00:00 and
counts up forever. Even single-player games need a console marshal.
RPL4CONSOLE is that console. Like the real one it lives on its own
thread: it owns the mission clock and raises the stop request at the
selected length; the app-manager per-frame hook (new gPerFrameHook seam
in APPMGR, called while the application global is live - the loop
condition NULLs it on exit, which ate the first attempt) executes the
engine-safe part, dispatching the same StopMissionMessage TeslaConsole
sent. Final scores flow in through a new RP-layer sink
(gConsoleScoreSink in RPCNSL): RPPlayer feeds it the same score it
sends a real console at mission end.
It also inherits the launcher role: the application tears down after a
stop (arcade pods were relaunched per mission by TeslaLauncher), so
WinMain respawns the process when the console ended the mission,
landing back on the race-setup screen. L4NetworkManager grows
FeedLocalEgg (the single-user egg-inject path, callable mid-session)
for the future in-process loop.
Verified end to end: menu -> 3:00 race -> stop dispatched exactly on
time -> final score collected (host 1 = 4113) -> process respawned with
the front end up. -egg runs stay unmarshaled (the dev shortcut).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Starting without -egg, -net, or -mr now boots a race-setup screen
(RP_L4/RPL4FE.cpp) instead of aborting: track / vehicle / color / badge
/ time-of-day / weather / race length plus pilot name, populated from
TeslaConsole''s RPConfig.xml catalog (Death Race scenario). LAUNCH
builds the egg exactly as the console did - the RPMission.ToEggString
port, including the pilot name pre-rendered to 1bpp plasma bitmaps
(128x32 + 64x16) via GDI with the console''s auto-shrink font logic and
the verbatim ordinal graphics - writes frontend.egg, and injects it
into the standard egg-load path (new L4Application::
SetEggNotationFileName).
The menu is a GDI child of the main window (pod green-on-black, double
buffered, mouse driven, EDIT control for the name) running a modal loop
before engine init; closing the window exits cleanly. Found and fixed
along the way: the empty egg CString holds a NULL representation
(operator! is the safe emptiness test), and the modal loop needed a
queue nudge for launch clicks delivered via SendMessage.
Verified end to end: boot -> menu -> LAUNCH -> generated egg (7.5KB) ->
racing in the 1080p cockpit with score and mission clock running.
start-windowed.bat now boots into the front end.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 4-8px black padding framing every MFD and the map read as a jarring
border over the 3D. Glass and button strips now run to the pane edges
(last button in each strip/column absorbs the rounding remainder);
only the thin separations between buttons within a strip remain.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
SS_BLACKRECT paints the system window-frame color - gray on modern
Windows - so the viewscreen child is now a plain STATIC subclassed to
erase true black. The MFD/map panes flickered because Paint() cleared
the on-screen surface before redrawing at the fill cadence; the pane
now composes off screen and lands in a single BitBlt.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Playtest: the half-height red buttons did not read well; back to
display_height/8 (18..40px). The contiguous amber map columns stay.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Playtest direction on the canvas layout: all five MFDs at the compact
320x240 glass size - upper pair pushed to the top corners, score glass
top-center, lower pair in the bottom corners, map bottom-center - and
the viewscreen now fills the entire 1920x1080 canvas. Launched with
-res 1920 1080 the 3D renders native 1:1 (the 2007 D3D9 path takes the
1080p backbuffer and 16:9 aspect without complaint). start-windowed.bat
updated accordingly.
Verified live: full-screen native 3D with the cockpit floating over its
edges, mission running, preset lamp lit on the map column.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CydandClaude Opus 5