Two items from listening-gate round 2.
Mute: AcdreamToggleAudioMute (default Ctrl+M; bare M is selection) flips
OpenAlAudioEngine.Muted, implemented as the AL LISTENER gain - unused
since A2 moved all mixing to the CPU, so it is a free master switch that
silences already-playing voices instantly and restores them exactly,
without touching the retail mixing math, the -50 dB allocation cutoff, or
any persisted volume setting. Rebindable like every other action; console
line confirms each flip.
Inn chatter: the user hears talk-and-laughter ambience in retail inns and
not in acdream. Three installed-dat scans (pinned as conformance tests in
EnvCellSoundEmitterInventoryTests) prove the mechanism is NOT client
data: no interior static in the town landblock carries an ambient-slot
sound table, no Setup among all 5,935 in the portal dat references one,
and yet 23 sound tables carrying ONLY Ambient1..8 slots exist - pure
soundscape banks with nothing client-side pointing at them. They are
wire-bound: the server attaches one to an emitter object via
CreateObject's sound-table field and fires the slots over 0xF750 - ACE
implements exactly this (EmoteType.Sound heartbeat emotes ->
GameMessageSound broadcast). Our 0xF750 receiver (slice A3) is live and
now instrumented (ACDREAM_PROBE_SOUND_WIRE=1, via the new
AudioDiagnostics owner per Code Structure Rule 5, with per-event drop
reasons in AudioHookSink.PlayServerSound). A probed session against the
local ACE received ZERO 0xF750 events across a town walkabout: the
silence is server world-content (no emitters configured/firing), not a
client drop. The first scan's assertion originally encoded the
emitter-object hypothesis; the data refuted it, and the test now pins the
negative so the conclusion cannot silently rot.
Full Release suite green (the one failure during development was the
hypothesis-pinning assertion, corrected to pin the finding).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
acdream had no ambient system: StartAmbient minted a handle and played
nothing. Retail's is a weighted-accumulation + timer-queue engine, not
looping voices. On every objcell change (24 m) CellManager::ChangePosition
rebuilds per-sound weights over the 3x3 landblock ring x 64 land cells
each, decoding each cell's terrain word through the region file's
terrain -> scene -> AmbientSTBDesc chain; playback is a min-heap of
absolute deadlines drained from the frame tick, where each pop fires a
one-shot and re-arms.
A continuous bed (base_chance == 0) is non-positional, crossfaded by its
share of the TOTAL weight, and re-fired every min_rate seconds — that
rate is the author's intended loop period, and re-firing is how retail
fakes a sustained bed with no looping voice, re-rolling the variant and
the crossfade each time. An intermittent one keeps its authored volume,
plays at a random accumulated compass bearing at min + (max-min)*t^2,
and is dice-gated. Indoors is silent by design: CEnvCell's contributor is
a folded ret and EnvCell carries no sound data.
The Opus review caught four bugs before this landed, one fatal:
- Cell offsets were built in ABSOLUTE world coordinates and differenced
against the listener's STREAMED-frame position, so every one of 576
offsets came out ~32 km, every contribution was culled, and the whole
feature was silent with nothing logged. Offsets are now landblock-local
the way Position::get_offset builds them, and the streamed-frame
position is carried separately for playback, where it belongs.
- The cell's weight was added to the shared denominator once per
DESCRIPTOR instead of once per CELL, dividing every bed's crossfade by
the table's entry count — enough to push a typical authored volume
under the 0.03 audibility floor.
- The drain used where retail's UseTime is strictly
below, so a descriptor authored with a zero rate re-armed at the same
instant and spun the frame forever.
- Arming only enqueued; retail's UpdatePlayQueue PLAYS and then re-arms,
so a newly audible ambient was silent for a full period after the
crossing that made it audible.
Also: beds now go through retail's single 16-voice priority pool rather
than acdream's UI pool (retail has one pool; parking beds in the UI pool
let an A4 portal cue chop one mid-wave and discarded the authored
priority), and CalcDir's in-block test is XY-only, since CalcWeight
includes Z on purpose and CalcDir excludes it on purpose.
Two behaviours are knowingly incomplete and registered rather than
guessed at slice end: TS-66 (sky-lit interiors should keep the outdoor
set) and TS-67 (contribution weight is computed in-plane). Retires TS-29.
The frame-loop hook is a typed IAmbientFramePhase, not a callback — the
first attempt used an Action<float> and the architecture guard
ExtractedUpdateOwners_DoNotRetainAnonymousCallbacks correctly rejected it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Retail's UI sound bank was absent, so three families of cue were silent:
the portal enter/exit stingers, the AdminEnvirons dungeon atmosphere
(chanting, drums, whispers, thunder — what players remember as dungeon
'music'), and every other interface slot.
The bank's DID is not a literal anywhere in retail: GetUISoundTable
@0x00563FB0 asks GetByEnum for enum slot 7, and DBCache::GetDIDFromEnum
@0x00413940 resolves it through two EnumIDMap hops off the portal dat
header's master map. UiSoundTableResolver walks that chain the way
RetailCursorResolver already walks it for cursors. Against the shipped
dats it resolves to 0x2000004B, and that table holds exactly the 32 UI_*
slots (UI_EnterPortal 0x6A .. UI_Thunder6 0x8A) — content that confirms
the walk independently of the decode. UiSoundTableResolutionTests pins
the walk, the DID, and the content, and skips when dats are absent.
Two corrections to the research along the way. The lane-5 note recorded
GetByEnum's arguments transposed: the 0x22 it called a fileType is the
CACHE type (CLOCache(cache, CSoundTable::Allocator, 0x22)) and the real
second-hop key is 0x10000003; walking it the other way finds nothing. And
its claim that the interface volume pref applies is wrong — GetAttenuation
with ambient=0 multiplies by the EFFECT knob, so retail's
interface_sound_volume stays the dead knob lane 1 byte-decoded it to be.
EnvironSoundCueMap is an explicit 21-case table read straight out of
Handle_Admin__Environs @0x0055DE20, not arithmetic: codes 0x65..0x72 sit
0x11 below their SoundType, but 0x73/0x74 have no case, so 0x75 lands on
UI_Squeal (0x84) where an offset gives 0x86, and the switch ends at 0x7B
with no 0x7C case. Verified case-by-case against the decomp rather than
from the lane note, whose tail table was ambiguous.
Cues are attached where retail plays them: the teleport-animation
boundary for the portal pair, and the AdminEnvirons handler for the
stingers. PlaySoundFromCenter's pan-0 / distance-0 shape is what
PlayUiWave already implements after A2.
Retires TS-54. Narrows AP-115 to its notice-presentation residual.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Retail is not a 3D audio engine. Every gameplay buffer is created with
m_3D = 0 and the DirectSound 3D listener the client sets up is dead code;
spatialization is two CPU scalars per voice, frozen at emission. This
slice ports that math and demotes OpenAL to a voice bank.
RetailSoundMixer (new, Core) carries the byte-decoded curve from
SoundManager::GetAttenuation @0x00550020: g = dist < 5 ? vol : 25*vol/d2,
clamped to 1 BEFORE the single master multiply, db = ceil(20*log10 g),
with a hard -50 dB floor at which retail does not start the voice at all
(audible radius ~94.2 m at unity). Pan is PlaySoundInternal @0x00550170's
(int)(-15*sin(delta-bearing)) in whole decibels, truncating toward zero,
forced to dead centre when (int)distance < 5, with no front/back and no
elevation cue. Every AL source is now source-relative with rolloff 0 and
the global distance model is None: AL's InverseDistanceClamped was
first-power (2/d), quieter than retail up close and far louder at range
with no cutoff whatsoever. That was the largest audible divergence in the
subsystem (AP-28, retired here).
RetailVoicePool (new, Core) ports the allocator at 0x0054FEC0: ring scan
for a free or finished slot, then evict the first slot whose DAT priority
is strictly lower, else drop. Eviction compared GAIN before, so a loud
unimportant sound could silence a quiet important one. It lives in Core
because the engine's play path talks to native AL handles and could not
be tested; the pool now has 12 conformance tests.
The listener keeps using the camera position, which the decode shows is
retail-faithful (SmartBox::set_viewer @0x00452D36 hands the same collided
camera Position to SoundManager) — only the heading extraction changes,
since retail reads one compass bearing and never a forward/up basis. An
earlier draft of the plan called this a defect; corrected in the plan so
it is not fixed backwards.
Opus review found and this commit fixes: a linear pan-to-azimuth mapping
that saturated to full separation at 30 degrees (OpenAL Soft's own
speaker angle) where retail gives 15 dB — now inverts the constant-power
pan law, so full deflection reaches 0.776 of the arc and both channels
stay live; the stale FUN_00550ad0 / gain-eviction class header, which
contradicted the register row this commit writes; missing discriminating
tests for clamp order and pan truncation; dead PlayingGain state whose
comment invented a retail symbol; and a third in-tree copy of
Position::heading, now delegating to MoveToMath.PositionHeading.
MasterVolume folds into the mixer's one multiply instead of AL listener
gain, so the cutoff, radius and dB quantisation move with the slider.
Register: AP-28 retired; AP-173 (pan law), AP-174 (volume taxonomy),
TS-64 (two unimplemented sound prefs), TS-65 (volume-squared quirk,
applied on the ambient path only) filed. Research note corrected twice
where its summary contradicted its own decode (30 m dB, floor vs trunc).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The SoundTable probability field is a Bernoulli play/skip gate applied at
the play site (SoundManager::PlayProbability @0x005500E0), not a selection
weight — and variant selection (SoundManager::GetSound @0x00550680) is a
uniform index over (n-1) that ignores probability entirely. SoundCookbook
did the opposite: a cumulative-distribution walk weighted BY probability,
short-circuiting single-entry lists before rolling at all.
A dat census says 4,183 of 4,184 entries are single-entry and 686 of those
carry probability < 1.0, so the gate was categorically absent: Speak1 idle
chatter authored at 0.05 fired every trigger (~20x too often), wound/attack/
swoosh variants never dropped, and six 0.0001 entries always played.
Split into retail's two steps (PickVariant + PlayProbability, composed by
Select) over a new ISoundRandom modelling both retail roll ranges: the
variant roll clamped below 1.0 (0x00797D48) and the gate's 1/32767 grid,
which is why 0.0001 resolves to ~1.2e-4. PickVariant reproduces retail's
(n-1) off-by-one verbatim per the port-faithfully rule — the last variant
of a multi-entry sound is unreachable, costing exactly one wave
(0x0A00051E) in the shipped dats.
Also removes invented mechanism this review disproved: the dead Core
SoundEntry/ISoundCache scaffold (PitchMin/PitchMax, Loop, Is3D — retail
never calls SetFrequency, never sets the loop flag, and creates every
gameplay buffer 2D), the engine's pitch plumbing, the int 0..7 priority
cast (the dat field is a float in [0,1]; 4,100 entries collapsed to 0),
and the clamp-at-the-field on volume (an unbounded gain retail clamps only
after the distance divide).
Tests rewritten as conformance against the disassembled values, replacing
a self-referential suite that pinned the wrong model.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
SoundId was not a subset of retail's table, the way its comment claimed.
It was an invention: 23 acdream-local names on acdream-local values, and
the values were wrong in the way that matters. FootstepDefault = 0x02 is
retail's Random. SwingSword = 0x10 is retail's Death2. Death = 0x60 is
retail's Explode. Anyone who reached for one of those names to compare
against a wire or dat value would have got a different sound.
Nothing referenced any of them by name -- grep for `SoundId.` across src
and tests returns nothing -- so this was a trap rather than a live defect,
the same shape the enum campaign found in DamageType. All 22 invented names
are deleted and retail's 205 replace them.
Three oracles agree exactly, on every name and every value: retail
acclient.h:4569 enum SoundType, ACE's Sound, and DatReaderWriter's Sound.
The third matters most. AudioHookSink already resolves SoundTable lookups
through DatReaderWriter.Enums.Sound, so that is the enum acdream actually
reads at runtime; our catalog now agrees with the values already flowing
through the dat path, and a conformance test pins the two so they cannot
drift apart.
On the "206 sounds" figure: retail's block holds 207 entries, being 205
sounds followed by NUM_SOUND_TYPES = 0xCD and FORCE_SoundType_32_BIT. The
first is a count and the second a width pin. Counting the former is where
206 came from. Neither is a member here, matching how the campaign treated
NUM_ATTACK_HEIGHTS and Num_HoldKeys -- a count is not a value the wire can
carry.
Behaviour is unchanged and could not be otherwise: the enum had no
consumers. IAudioEngine's three SoundId overloads are no-op stubs and the
live path takes wave ids and DatReaderWriter values.
The user's separate report that sound is "not working that good" is a
triggering, selection and attenuation question rather than a catalog one,
and is filed as its own Bucket B row in the post-Vulkan intake.
Also in this commit, by user decision: AC2D is retired as a reference. Its
clone and directory are gone and it must not be re-cloned. Everything we
took from it still stands and is written down -- the FSplitNESW terrain
split constants, the 0xF61C movement packet layout, the finding that a
client need not compute terrain Z itself -- so CLAUDE.md's reference list,
its hierarchy table, and the architecture doc's protocol row now point at
docs/research/2026-04-12-movement-deep-dive.md rather than erasing the
history. The reference count drops from six to five.
Core tests 3907 passed / 2 skipped.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Complete Slice D4 by adding aggregate lifecycle occupancy and traffic facts, validating physical source reports, and including decoded audio under the typed startup budget. Exercise every domain under forced pressure and retain the real cache/fence convergence gates.
Two audit-verified caches grew monotonically for process lifetime, which is
fatal for the 30-bot long-uptime headless-fleet goal:
- DatSoundCache._waves (Core) memoized every decoded PCM WaveData forever
in a bare ConcurrentDictionary — no LRU, no byte budget.
- OpenAlAudioEngine._bufferByWaveId (App) retained a native OpenAL buffer
copy of the same PCM per wave id until engine disposal — a second,
independent unbounded cache.
DatSoundCache now bounds payload residency with a 32 MiB byte-budget LRU
(decoded PCM waves run ~100-500 KB each, so this holds a comfortable
working set). Missing/unsupported-format waves are memoized separately in
an unbounded-but-cheap negative-result set (bounded by the finite Wave dat
id space) so they can never compete with or get evicted alongside payload
entries. Concurrent first-touch decodes of the same wave id are deduped
via a shared Lazy<T> so racing callers don't pay for WaveDecoder.Decode
twice. AcDream.Core cannot reference AcDream.Content (code-structure rule
2), so the LRU is a local reimplementation mirroring
BoundedDatObjectCache/DecodedTextureCache's shape rather than a shared
dependency.
OpenAlAudioEngine._bufferByWaveId now bounds native buffer residency with
a 48 MiB byte-budget LRU (AlBufferBudgetTracker), evicting least-recently-
used buffers once oversized. alDeleteBuffers fails on a buffer still
attached to a source, so eviction queries live AL per-source state
(GetSourceInteger.Buffer) rather than tracking a second, easily-stale
copy — several call sites (Play3DWave, PlayUiWave) set a source's buffer
directly. The buffer EnsureBuffer just created is explicitly protected
from its own eviction pass, since the caller hasn't attached it to a
source yet at that point. Evicted waves simply replay through
DatSoundCache -> EnsureBuffer on next use, identical to a first play.
Verified before implementing: AudioHookSink is the only GetWave caller
(single per-frame render-thread path per AnimationHookRouter's own
threading doc), and PcmBytes is read only at DatSoundCache.Admit (byte
accounting) and EnsureBuffer's first-upload branch — confirmed dead after
AL upload on the steady-state replay path, so bounding either cache
independently is correctness-safe; a cold replay after both evict simply
falls back to a full re-decode + re-upload, identical to a first play.
AlBufferBudgetTracker's eviction/budget decision is extracted as pure
logic (no AL dependency) specifically so it's unit-testable: the existing
OpenAlResourceLifetimeTests fake exposes a null AL, which short-circuits
every native buffer call before it runs, so the engine's actual AL wiring
isn't testable headless.
Tests: 9 new DatSoundCacheTests (Core.Tests) covering eviction order, byte
accounting, negative-result memoization, oversize-single-entry handling,
and concurrent-access smoke tests; 10 new AlBufferBudgetTrackerTests
(App.Tests) covering the pure LRU/budget/protection logic. Full suite:
3214/2 skip (Core.Tests), 3471/3 skip (App.Tests) plus one pre-existing,
unrelated failure (LandblockBuildOriginTests.FarLoad_..., reproduces
identically with these changes stashed out — landblock streaming, not
audio).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 76c880d35bcf30b50e3f7b4cb8635dc9ab9e3ec7)
Full audio pipeline from MotionHook → OpenAL 3D playback. Faithful to
retail's 16-voice pool, inverse-square falloff, and SoundTable
probabilistic variant selection.
Core layer (AcDream.Core/Audio):
- WaveDecoder parses the WAVEFORMATEX in Wave dat headers. PCM
(wFormatTag=1) decodes directly; MP3 (0x55) and ADPCM (0x02) return
null + log (ACM compressed decoders need Windows winmm; cross-platform
path deferred). Cites r05 §2.1-2.3 + ACE Wave.cs.
- SoundCookbook.Roll implements the probability-weighted entry pick that
gives retail footsteps their variation. Cumulative-distribution walk;
silence tail when probabilities sum to <1.
- DatSoundCache: ConcurrentDictionary-backed lazy load of Wave /
SoundTable dats, decoded PCM memoized.
App layer (AcDream.App/Audio):
- OpenAlAudioEngine (Silk.NET.OpenAL): 16-source 3D pool with
round-robin first-free, then evict-quieter-slot algorithm matching
retail chunk_00550000.c FUN_00550ad0 exactly. Separate 4-source UI
pool (source-relative). AL buffer cache keyed by Wave id.
InverseDistanceClamped distance model. Fail-open when AL driver
missing or ACDREAM_NO_AUDIO=1 — client continues without audio.
- AudioHookSink routes SoundHook / SoundTableHook / SoundTweakedHook
from the Phase E.1 animation-hook router into OpenAL. All three
hook types fire on both player AND NPCs/monsters (the sequencer
dispatches per-entity and the sink uses entity worldPos for 3D pan).
- DictionaryEntitySoundTable holds per-entity SoundTable mapping,
populated from Setup.DefaultSoundTable at hydration time. Server-
sent overrides would take precedence here when wired.
GameWindow integration:
- OpenAL init in OnLoad after dat collection, suppressible via
ACDREAM_NO_AUDIO=1.
- SetListener called each OnRender frame with camera position + view
basis vectors (fwd = -Z, up = +Y of inverse view).
- AudioEngine disposed in OnClosing before dats.
Tests: 6 WaveDecoder (PCM / MP3-null / ADPCM-null / stereo / truncated
/ peek) + 6 SoundCookbook (empty / single / 50-30-20 distribution
within 5%, silence tail, table lookup, missing table key). Verified
against r05 §2 + ACViewer export-path.
Build green, 497 tests pass (up from 485).
Ref: r05 §2 (Wave format), §5.3 (16-voice pool + eviction).
Ref: FUN_00550ad0 (chunk_00550000.c:527) eviction algorithm.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Erik