Eight byte-identical field stacks from night 6, all one player, all in an
Owens: ParticleEngine::Destroy +0x11, access=0 target=0x0, from the plain
per-frame render path. Nothing in the stack touches weapons or the Owens.
Conn Man's Surface Pro 9 (Iris Xe, 128 MB shared) is simply the only GPU in
the fleet that ever actually LOSES the D3D9 device -- his two-trigger
missile+laser bursts are what provoke the timeout, not what crashes.
What crashed is our device-loss handling, which was wrong three ways at once,
in two inline copies (the scene Present and the wait-screen Present):
1. On D3DERR_DEVICELOST it called Reset() IMMEDIATELY. Reset on a
still-lost device ALWAYS fails, and V() only logs. There was no
TestCooperativeLevel gate at all.
2. It then ran ParticleEngine::Initialize against the lost device. The
creates fail there and NULL their out-params -- proven, not assumed:
the bench repro faults at target=0x0, not at a dangling address.
3. The next lost frame called ParticleEngine::Destroy again, which
Release()d those NULLs blind. Read of vtable at 0x0. Dead.
So: lost frame 1 tears down and leaves NULLs, lost frame 2 crashes. Two
frames, every time, deterministic -- which is exactly why all 8 field stacks
are byte-identical.
Reproduced before fixing. BT_DEVICELOST_TEST=<frame>,crashrepro runs the
field sequence on the bench; on the unfixed build it died at Destroy +0x11,
access=0 target=0x0, and symbolized to the same four frames as the field
logs. Same shape, same offsets-modulo-hook. That run also proved the
out-param-nulling assumption the whole diagnosis rested on.
The fix -- one shared DPLRenderer::BTResetLostDevice() replacing both inline
copies:
- Destroy() is idempotent and null-safe, and nulls after release.
- Reset() is gated on TestCooperativeLevel() != D3DERR_DEVICELOST; while
the driver still says lost, skip the frame and retry.
- The Reset HRESULT is checked; on failure, log and retry next frame
instead of driving on.
- On success, re-create via the new CreateDeviceObjects(), NOT
Initialize(): Initialize memsets the installed-effects table, so every
reset that DID succeed silently killed all particle effects for the rest
of the mission. The quieter sibling bug, fixed by the same split.
- Initialize checks its HRESULTs and defends MAXPARTICLES<=0; the draw
paths guard the NULL buffer, and ExecuteParticles keeps draining
particles while the engine is dormant so they cannot pile up.
Verified: the crashrepro shape now logs SURVIVED and play continues; three
forced full loss/reset cycles each log "[render] device reset OK"; a plain
run is assert-free.
Found while verifying, worth its own line: VIDEO\particles.png has NEVER
existed -- not in the tree, not in BTL4.RES, not anywhere in git history.
The texture load has failed on every machine since the engine was written,
and every billboard particle ever rendered was untextured quads via
SetTexture(0, NULL). RenderParticles deliberately does NOT gate on the
texture -- that would disable all particles everywhere; untextured IS the
shipped look. Filed separately; a real particle sheet is a content task.
The field verification that counts is Conn Man flying his exact crash
loadout on this build: instead of a dead process he should see at worst a
brief hitch and "[render] device reset OK" in his log. #35 stays open until
that happens.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Correcting my own framing in 7e57816. The dropout players actually reported was
fixed on 2026-07-23 by a999e5c, and nobody has complained since. I read the
leftover ACQUIRE FAILED lines in the night-5 logs as the same live bug and
raised the ceiling by default. They are not the same bug.
What is still in the logs is transient exhaustion, not a leak: the census sits
at 45-60 live, spikes during a firefight, and drains straight back (226 -> 42),
which is the priority steal loop working. A voice dropped in that window is
competing with ~256 already sounding, so it is very likely sub-perceptual --
which is consistent with the thing nobody is reporting.
And raising the cap is not free. The 256 ceiling doubles as a governor: the
steal loop only steals when the incoming source outranks a running one, so a
higher cap means many more voices mixing simultaneously, and with EFX reverb and
the lowpass chains live that is real CPU -- spent precisely during heavy combat,
when the frame budget is already tightest. I verified the ceiling moves. I never
measured frame time under a real firefight, and a solo bench cannot produce one.
Shipping that to testers tonight would be gambling their framerate against a
complaint nobody is making.
So the knob stays and the default does not move. BT_AUDIO_SOURCES=<n> raises it
with no rebuild if dropouts are ever reported again -- measure frame time while
you do. Unset reports `granted mono=255 stereo=1`, exactly what testers have
been running.
Keeping the diagnostic read-back either way, since it earned its keep: asking
for 64 grants 240, because OpenAL Soft has a floor of its own, which is also why
the NULL default lands on 256. A request is not a promise.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Field report from 2026-07-23 was "audio cutting in and out toward the end of the
match". The night-5 logs say it precisely: 2017 ACQUIRE FAILED lines in a single
match, and every last one of them at live=256.
It is not a leak, which is where I started and where the code comments still
pointed. The source census sits at a healthy 45-60 live, spikes during
firefights, and drains right back down afterwards -- 226 back to 42 in one
window -- so the engine's priority steal loop is doing exactly what it should:
AudioSourceStop/SuspendMaintenance -> ReleaseChannels -> ReleaseSourceSet ->
alDeleteSources, with the live counter following it down. The 2026-07-23 fix
(a999e5c, deleting sources one at a time instead of the spec-atomic bulk call)
was real and is what makes that drain work. It just was not the ceiling.
The ceiling was ours. MakeAudioRenderer called alcCreateContext(device, NULL) --
no attribute list at all -- so OpenAL Soft applied its default budget of 256
mono sources. That number has nothing to do with the 1995 audio hardware; it is
just what you get for not asking. Meanwhile a busy match logs about 7200
explosions, each spawning three DPLIndependantEffect voices, so the pool pins at
the cap and every acquire during that window fails outright and drops its sound.
The peak we actually observed was 226 against a 256 cap, i.e. the bursts were
already scraping the ceiling.
So ask. ALC_MONO_SOURCES at 1024 by default, BT_AUDIO_SOURCES to override for
A/B testing, and read back what the driver GRANTED rather than assuming the
request was honoured. That last part earns its keep: asking for 64 grants 240,
because OpenAL Soft has a floor of its own -- which is also the reason the NULL
default landed on 256 in the first place.
requested 1024 -> granted 1024
requested 2048 -> granted 2048
requested 64 -> granted 240 (driver floor)
Costs mixing headroom, not hardware voices -- OpenAL Soft mixes in software and
only touches voices that are actually sounding.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
SAURON's "coolant flush sound glitched and perma" is one instance of a systemic
defect. SetupPatch (L4AUDLVL.cpp) decided looping from the SAMPLE flag alone:
AL_LOOPING = (info.loop != ForceStatic)
That armed every non-ForceStatic sample -- 224 of the 603 authored zones,
including unmistakable one-shots -- as an OpenAL source that never ends on its
own. Any such sound whose note-off never arrives plays forever. Measured 1946
LoopAtWill x Transient arming events in one short session.
The sample flag expresses PERMISSION; the SOURCE decides. Three facts [T1]:
* the enum's own comments (L4AUDLVL.h:14-19) -- LoopAtWill = "will play once
OR LOOP AS DESIRED", ForceStatic = "plays only once EVEN IF LOOPED" (a
veto), LoopAlways = "ramp up and then down";
* LoopAtWill is enum value 0, i.e. the DEFAULT an unauthored sample receives
(WTPresets.cpp:37) -- it cannot mean "loop forever";
* LoopAlways, the real always-loop, is used by ZERO shipped samples.
"As desired" is the source's authored AudioRenderType (Transient=one-shot vs
Sustained=held), streamed from AUDIO*.RES (AUDLVL.cpp:28) and already consulted
by the renderer (L4AUDRND.cpp:567, AUDREND.cpp:184). SetupPatch now takes it,
threaded from all three L4AudioSource call sites (Direct/Dynamic3D/Static3D).
New rule: LoopAlways->1, ForceStatic->0, LoopAtWill->the source's render type.
MEASURED before changing behaviour (BT_LOOP_AUDIT=1, scratchpad/loopaudit.py):
LoopAtWill x Transient -> loop=0 (was 1) 1946 events
LoopAtWill x Sustained -> loop=1 unchanged 65 events
ForceStatic x Transient -> loop=0 unchanged 292 events
The engine loops (EngineAccel07_z0..z2, EngineMotor01, EnginePower01) are all
LoopAtWill/Sustained and PRESERVED -- no regression there. All 24 reclassified
samples are genuine one-shots: laser charge/fire/explosion/loaded, missile
loading, engine shift, coolant pressure inc/dec.
A/B PROOF (scratchpad/loopab.py, same session both arms, only BT_LOOP_LEGACY
differs), asking the engine's own BT_AUDIO_DUMP what is still playing with
loop=1 long after every release:
[legacy] EnginePower01, CoolantPresInc03_z2, CoolantPresDcr03_z2
[fixed] EnginePower01
The two stuck coolant sources are the reported symptom. They were eventually
reclaimed when a later trigger reused the source, which is why the bug was
intermittent -- the "perma" case is when nothing else retriggers it. The fix
removes the mechanism.
BT_LOOP_LEGACY=1 restores the old rule for a field A/B without a rebuild. The
layers to listen to are the beam sustains: LaserA/CSustain* are authored
LoopAtWill on a TRANSIENT source, so they now end with the sample instead of
looping until note-off. The render type is T1 authored data; the interpretation
that LoopAtWill defers to it is a strong reading of the enum + defaults rather
than a disassembled statement, and is flagged as such in the KB.
Plausibly bears on #32 (audio cutting in/out late in a match): a stuck looping
source holds its pool slot for the rest of the round.
Rigs: scratchpad/flushsnd.py (flush start/stop pairing), flushsnd2.py (stuck-
source hunt), loopaudit.py (the classification matrix), loopab.py (the A/B).
KB: context/wintesla-port.md audio section, context/decomp-reference.md env
table (BT_LOOP_AUDIT / BT_LOOP_LEGACY / BT_AUDIO_DUMP / BT_AUD_TAIL),
docs/AUDIO_FIDELITY.md F38. checkctx.py CLEAN.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Field report: audio cutting in and out toward the end of the match.
Three structural defects in the port's source-pool lifecycle:
1. SourceSet.sources[] was uninitialized heap garbage until first
acquisition -- AL names are small ints, so a recycled-heap slot could
alias a LIVE source owned by another sound; alIsSource() skipped
generation, two sounds shared one source, and whichever released
first deleted the other's mid-play. SourceSet ctor now zero-inits.
2. ReleaseSourceSet's bulk alDeleteSources(count, sources) is ATOMIC on
invalid names (AL spec) -- one empty/-1 slot and NOTHING deleted;
after the first pool exhaustion every partial release leaked and the
pool never recovered. Now per-slot guarded delete, slots parked at 0.
3. Failed acquisitions stranded partial sets forever (dropped transients
are never released by anything). ReleaseChannels() handback at both
failure sites (StartRequest + dormant-resume) + a dtor backstop.
Verified: 2-node MP smoke (mission + lobby-loop relaunch clean, census
live/acquireFails 0/0) + 100s solo combat smoke (52 audio triggers,
zero AL errors). Awaiting live playtest confirmation at scale.
KB: context/wintesla-port.md audio-dropout section.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
PARALLAX (#16): the pick/fire ray was anchored at mech.y+5.0 (a port
improvisation) while the sight line ran from the eyepoint (y=6.23) --
two parallel rays whose constant offset grew into the reported low-miss
as range closed (measured ry +0.072 @50u -> +1.54 @2.7u). The decomp's
sight and pick share the eye origin (HudSimulation @4b7830 chain). Fix:
the viewpoint mech's cockpit eye owns the aim-camera publish in BOTH
views, origin = its own eye translation; leveling + deliberate elevation
untouched; chase view now converges to cockpit ballistics (V cannot
change where shots land). After: pick pinned to the crosshair (ry <=
5e-6) from 50u to point-blank; 26 laser + 8 missile center-mass hits at
3/4-screen. Awaiting the reporters' approach-test.
VERDICT INSTRUMENTATION (#4 closed authentic, #5 verdict posted):
BT_RANGE_LOG per-frame pick tracing + BTGroundRayHitExact analytic
cross-check (0/8000 arena fall-throughs; cavern 6/8400 single-frame
grazes -- the 'crazy sliding' is the authentic world-pick + 500 m/s
slide over depth discontinuities); BT_AUD_TAIL StopNote/fade timing +
BT_FIRE_PULSE single-shot driver (the energy 'buzz' is the AUTHORED
charging loop: crescendo through recharge, 1.309s authored release,
measured within one frame). CAVERN.EGG: solo cavern test egg.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- decomp-reference.md: the binary ATTRIBUTE-TABLE section -- 16-byte
{id,name,off+1,0} row format (alignment warning), the walker technique,
and the three recovered tables (Mech ids 21-56 incl the corrected
EyepointRotation@0x360 / RearFiring@0x410 / DistanceToMissile@0x400
labels; HeatSink 3-12; Torso 3-15; the weapon RearFiring 'b' marker).
- combat-damage.md: the REAR-FIRE + look-view system (roster survey: the
Blackhawk's ERMLaser_2/3 are the game's only rear weapons; every missile
rack forward), the missile mount-frame launch truth, and the BANKED
full Missile flight-model decode (three performances, proximity fuse).
- wintesla-port.md: post-Phase-4 closures (instability model live, F14
static filter baked, the footstep warm-up bug + its trace-cap lesson);
deferred list trimmed to F21/HRTF.
- AUDIO_FIDELITY.md: status block -- F14 FIXED, warm-up bug noted.
- open-questions.md: "which mechs used rear fire" answered from data;
remaining look-view key bindings noted.
checkctx: CLEAN.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Correct the audio digest in wintesla-port.md + project-overview.md: the "no
sound ever" root cause was that BOTH engine/lib backend DLLs were no-op stubs
(libsndfile ordinal-only sf_open->NULL; OpenAL32 imports only KERNEL32). Now
real OpenAL Soft + LoadWavPCM; soundbank cracked (241 samples). Only remaining
gap = game triggering (AudioEntities).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
User reported never hearing any audio. Root-caused: the KB "Audio: DONE" was wrong.
The OpenAL playback chain IS fully implemented (device/buffers/sources; PlayNote
really calls alSourcePlay), but three gaps kept it silent:
1. GATED OFF (the "no sound" root cause): BTL4Application::MakeAudioRenderer returned
NULL unless the pod's AWE_FRONT/AWE_REAR AWE32-card env vars were set -- authentic
1995 pod behavior, dead on modern hardware, so the renderer was NEVER created.
FIXED: default audio ON (BT_NO_AUDIO=1 restores silence; AWE vars still force-on).
Verified: the OpenAL device now opens with no env vars ([audio] device OPENED).
2. Soundbank STUB: allPresets[2][100] (btstubs.cpp) is zero-init -> PRESET_isImplemented
false -> 0 buffers load. Sample data exists (AUDIO1/2.RES: EnginePower/LaserAFire/...)
but the event->sample map is gone (not in the decomp). STILL OPEN.
3. No triggering: the reconstructed game never creates AudioEntities on events. STILL OPEN.
So the device opens but nothing loads/plays yet -- enabling real sound needs the
soundbank reconstructed + the game triggers wired (a proper audio wave). Added a
BT_AUDIO_LOG trace harness across L4AUDRND/L4AUDRES/L4AUDLVL. KB corrected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
