geo_push.hpp and replay.hpp cited memllib @ 0a541cc. upstream/main pins
e291192, where the same code had been deliberately redesigned — and because
InterfaceRL was not in the tree (fixed one commit ago), we carried the
superseded version for months. Three changes, all upstream's:
kGeometricPushScale 0.5 -> 1.0 (InterfaceRL.hpp:409)
kNegLRBase 0.5 -> 1.5 (InterfaceRL.hpp:410)
/(1+len) taper deleted (InterfaceRL.tpp:724)
Upstream's own comment on the taper: "a 'no' should clearly move the mapping
away even from a sound already far from the liked region (the taper used to
kill exactly that case)". The direction is already a unit vector, so the
taper only ever shrank the push for exactly the sounds a user is most likely
to be rejecting.
Cold start is folded into the same path. Upstream's useRandom is
`!havePositives || len <= 1e-4`: with nothing liked yet there is no centroid
to push away from, so every dim goes in a random direction. Ours instead
kept the older 0a541cc fallback — train AWAY from the heard action at a
NEGATIVE lr — which was inert whenever the heard action equalled the net's
own output, i.e. in the common case. One path now, and a "no" moves the
mapping before any likes exist (ml_bench E1: 0 -> 2.3e-3). The
GeometricColdStart action is still reported so callers keep their "like a
few sounds first" prompt; only the training changed.
Measured (ml_bench, one dislike at a point):
A4 0.0157 -> 0.0533 at-point displacement (3.4x), so end to end across
this and the RMSProp fix: 5.3e-5 -> 5.3e-2, ~1000x. The gap to the
legacy Diffuse design closes from ~4100x to ~4.2x.
D1 effective_lr 4.7e-4 -> 1.5e-3; 10 presses now reach 0.34, 100 reach
the full intended push.
A5 compounding 0.87 -> 0.96 (a second press at the same spot is no
longer noticeably weaker than the first).
A7 damage_ratio essentially unchanged (0.87-1.57) — the collateral
damage to protected positives scales with the push and is NOT
addressed here; it is the negative-feedback design question.
NOT adopted, deliberately: upstream's per-tick batch retraining over all
live negatives, and its fixed kDislikeLifetimeMs=2500 in place of our
proportional decay. Both need something the core does not have — a per-tick
call site and a millisecond clock inside nisps/ml — so they change
FeedbackControllerCore's interface rather than its constants. Recorded in
ALIGNMENT's deferred-debt entry alongside the existing one-press-one-step
divergence, and filed as its own task.
test_mlp_geo_dislike.cpp: the taper test now pins its ABSENCE (equal
displacement near and far), the cold-start test pins movement where it used
to pin inertness, and a new test covers the random-direction branch.
ALIGNMENT defect 6b resolved. Gates: build-cpp-tests 139 tests / ctest 4/4,
parity-check PASS (WASM rebuilt), lint-cpp clean, firmware slpworkshop
SUCCESS.
Upstream memlp (github.com/MusicallyEmbodiedML/memlp @ ea777502, the commit
upstream/main pins) applies gradients with RMSProp everywhere: Layer.h:239
ApplyAccumulatedGradients, the m_sq_grad_avg running average at Layer.h:601,
StaticMLP.h:268. nisps/ml/training.hpp shipped SGD only and filed the
difference as an optimiser-choice research question. It was not one.
RMSProp divides each step by the running gradient magnitude, so an upstream
lr is a NORMALISED step; under SGD the same number multiplies the raw
gradient. Every learning rate ported from upstream therefore landed in an
optimiser that reads it differently — most visibly feedback.hpp's
`geo_lr_ = 0.001f // upstream InterfaceRL.hpp:312`, an RMSProp LR pasted
into a single SGD step.
rmsprop_step() ports Layer.h:239 exactly: clip at +/-10, sq = min(0.9*sq +
0.1*g^2, 1e6), adj = min(lr/(sqrt(sq)+1e-6), 1.0), w -= adj*g. The
adjusted-LR clamp stays one-sided as upstream's std::min is, so the negative
lr used by train_targets' "train away from this target" path behaves as it
does upstream. The per-weight squared-gradient average is new persistent
state and lives in the storage policies (FixedStorage arrays /
DynamicStorage arena) so nisps/ stays allocation-free and the firmware's
zero-heap contract holds. It is optimiser state, not model state: excluded
from weight_count()/get_weights()/set_weights(), matching upstream, and
cleared by MLPCore::reset_optimizer_state() (upstream ResetOptimizerState).
draw_weights() deliberately does NOT clear it — upstream's DrawWeights
doesn't either.
Measured with tests/cpp/ml_bench.cpp:
D1 one geometric dislike moves the mapping 1.6e-2, up from 5.3e-5 (~295x),
and repeated presses now CONVERGE on the intended 0.5 push (0.12 at 10,
0.56 at 100) instead of creeping linearly forever.
A4 geometric-vs-Diffuse gap narrows from ~4100x to ~14x in one press.
U4 the upstream-LR positive path actually trains now (range_util 0.71 at
100 ticks/gesture, was 0.016 — it was inert under SGD).
Not fixed by this, and now tracked as ALIGNMENT defect 6d: the dose
asymmetry. lurch_max is still ~1.08 against a [0,1] output range.
Golden vector stages 2 and 3 re-captured; stages 0 and 1 are pre-training
and did not move, which is the cross-check that only the update rule
changed. manifold/public/nisps.wasm rebuilt so parity-check compares like
with like — it FAILED at up to 5e-2 against the stale artifact and PASSES at
2.4e-7 against a fresh one. parity-check.sh only builds the WASM when it is
missing, never when it is stale; noted in MAP.md and filed separately.
ALIGNMENT defect 6 resolved (moved to Recently resolved); 6b's optimiser
cross-reference updated; new defect 6d for the positive-training dose.
Gates: build-cpp-tests 138 tests / ctest 4/4, parity-check PASS, lint-cpp
clean, manifold typecheck clean.
Four items from one workflow, committed together because their build and CI
wiring genuinely interleaves — nisps/CMakeLists.txt, run-all-tests.sh and
ci.yml each carry hunks from two of them, and the stage renumbering (1/5 ->
1/6) touches every line. Splitting would produce commits that do not build,
which is worse than a commit that does four things and says so.
S26 part 2 — the curve declaration now matches reality. params[].curve stays
the mode-wide DEFAULT; a voice_spaces entry may now be {name, curve_overrides}
declaring only the slots where THAT voice space deviates. The 6 modes with one
voice space are byte-identical. The values were derived MECHANICALLY by a new
codegen/curve-audit.ts that models the four idioms a p[N]*p[N] regex misses
(alias form, memlcelium's implicit-counter sq() lambda, loop-generated indices,
smooth_params_), inlines helpers, and RAISES rather than guessing when it
cannot reduce an expression. A drift gate cross-checks 1179 (voice space x
param) slots against engine source on every run and was proved to fail loudly
on three drift classes. Application stays in the engine: nisps/engines,
nisps/pipeline and nisps/core are untouched, generated output is pure insertion
(755 insertions, 0 deletions), and the rebuilt nisps.wasm was byte-identical.
S4 / 7.2 — firmware reads the active mode's driver config at mode start, and
mic/line is real. My brief assumed the engine owns this; the code disagreed and
the code was right. sound_analysis_midi's EngineT is NoOpEngine — the mic lives
on a separately-composed AnalysisEngine member — so engine-level wiring would
have compiled, passed every gate, and left the one mic mode on line input.
Hence a mode-level seam defaulting to engine().driver_config(). Separately,
DriverConfig's defaults (line_level 0, output_volume 1.0) had drifted from
memllib's actual 3/0.8 because nothing had ever read them; wiring them as-is
would have made every silent mode louder and its line input maximally
insensitive — a behaviour change disguised as plumbing. Now pinned by a test.
Also: GetSysClockSpeed() panic()s on unsupported sample rates and runs on the
first line of setup(), so sample_rate needed a fallback ahead of clock setup.
CI's firmware env list gains soundanalysismidi — it is the only mic variant and
nothing else compiles that path.
Plan 5e — telemetry is real. A loss_history C-API entry across the full 5-layer
chain lets the browser read the per-iteration loss the core already records.
The audit named one fabrication site; there were two — wasm-iml.ts's
synchronous train() published lossHistory: [loss] as well. A third, ctx.loss,
was not merely dead but actively synthetic (fallbacks of prev * 0.82 and a
literal 0.5, rendered by nothing) and is deleted. The firmware buffer stays
untouched, per the L25 call. EngineApi.lossHistory() reads spine state rather
than the MLP handle, because trainAsync() fits on the worker's mirror net and
the handle would give a subtly-wrong second answer.
Plan 5f — engine throughput is measurable. One source compiled twice (CMake
natively, emcc for WASM) so the targets compare directly and no WASM export is
added. Sequencers are driven into a working state, and every row prints its own
working-state evidence so a number produced by an idle engine is visible rather
than plausible. Reports, never asserts: a wall-clock threshold on shared
hardware is meaningless or flaky, same call as the firmware size job.
ALIGNMENT: the telemetry defect is deleted (built, not deferred); the
performance defect is rewritten to what is actually left — these are HOST
numbers, and nothing measures the RP2350 at 150 MHz, which is the target the
mission's constraint is about. Q4 (memllib ownership) and Q5 (legacy feedback
modes) are closed.
Corrections to my own earlier claims, both found by agents contradicting the
brief: manifold/ONBOARDING.md was NOT "now accurate" — its primitives list
still named five deleted primitives and cited a seededGradient() that does not
exist. And the parity harness misses the sequencer engines because it runs 128
frames while their sequencers evaluate every 400-500 samples, NOT because
all-params-0.5 fails to trigger them (it does trigger: 0.5 maps to ratio 2,
firing three times per bar). The fix is a longer window, not different params.
Gates: run-all-tests.sh ALL GREEN — 4/4 ctest, parity PASS, lint clean, curve
drift 1179 slots ok, 39 e2e (was 33). Firmware: 5 envs built including the mic
variant.
The operator's call: "there should be one default learning rate and one
default max iterations and they should both be configurable at runtime."
There were SIX copies, not the four the audit described, and they did not
agree:
nisps/ml/mlp.hpp no-arg train() hardcoding 1.f / 1000u / 0.001f —
and firmware's ONLY training path calls exactly
this, so firmware had no runtime knob at all
wasm-iml.ts train() and trainAsync() TS default params (x2)
engine-api.ts learningRate ?? 1.0, with no maxIterations knob
vcv/src/iml.hpp 200 / 0.1 / 0.00001 — silently divergent
external_synth_midi.hpp its own kDefaultLearningRate/kDefaultMaxIterations
schemas/modes/*.json x9, identical, read by nobody at runtime
Now: schemas/ml_defaults.json is the single declaration (validated against a
sibling meta-schema, matching the midi_device.schema.json convention), codegen
emits it to C++ and TS in the same run, and MLPCore carries a TrainConfig whose
default member initialisers read the generated constant.
set_train_config()/nisps_ml_set_train_config() make it runtime-overridable on
every target; the explicit-argument train() overload is untouched. min_error
joins the tuple — it was duplicated identically and belongs with the other two.
The per-mode ml block loses default_learning_rate/default_max_iterations.
default_spread stays (genuinely wired on both targets) and input_channels stays
(codegen-time validated, real information for sound_analysis_midi).
VCV BEHAVIOUR CHANGE, deliberate: MEMLNaut.cpp constructs IML positionally and
relies on those defaults, so the module moves to 1000/1.0/0.001 — 5x the max
iterations, 10x the learning rate, and a 100x looser early-stop threshold. The
old values were never justified anywhere; they arrived with fbc68eb alongside
an unrelated module rewrite and no tuning rationale. Firmware and WASM have
shipped 1.0/1000 all along. It is now runtime-settable if this turns out worse.
The generated header lands in nisps/ml/generated/, not nisps/modes/generated/
where the rest of codegen output lives: training hyperparameters are an ML
fact, and nisps/ml sits below nisps/modes, so emitting them there would make
mlp.hpp include upward. The agent that built this flagged the directory-crossing
rather than hiding it; this is the fix. CI's generated-freshness gate learns the
new directory.
Gates: run-all-tests.sh ALL GREEN — 4/4 ctest, parity PASS (max delta 2.38e-7),
lint clean, manifold typecheck + 17 unit + 33 e2e (which exercise train() and
trainAsync() through a real browser).
feedback_/fbzm1_/fb_smooth_alpha_ were read every process() call but nothing
ever wrote feedback_, so the path was silently inert. Git history shows this
was a *muted* feature, not dead code: `feedback = v0 * feedbackGain` with
feedbackGain 0.1f (MEMLCeliumAudioApp.hpp @ d095688), then "0; //0.1f" —
zeroed with the value kept in the comment — then the write was dropped
entirely when the engine was ported to nisps/engines/ at the 2026-04-29
rewrite (8d0d47b). Operator decision: port it back, live at 0.1f.
The tap is voice 0's own enveloped output, post-envelope and pre voice-1/mix/
shape — NOT the final output. This differs from PAFSynth's feedback shape;
where the sibling and the reference disagreed, the reference won.
Not schema-exposed: the reference hardcoded it, and MEMLCelium has no voice
spaces to vary it by. If feedback depth becomes a research axis it wants to be
a runtime parameter — noted, not built.
Verification, and a gap worth recording: parity-check.sh PASSES (max delta
2.38e-7) but that is NOT evidence for this change — its harness only exercises
PAFSynth and ChannelStrip, and the impulse baseline runs all-params-0.5, under
which voice 0's sequencer never triggers, so v0_env stays 0 and the new write
is 0 either way. That baseline is blind to this path, not a check on it. So it
was cross-checked directly instead: sequencer params forced to trigger early,
4800 samples native vs WASM through the nisps_engine_* C ABI —
native vs WASM max abs diff 1.19e-7 (tolerance 1e-5)
all 4800 samples nonzero, finite, max amplitude ~0.99 (tanh-bounded)
same run with gain forced to 0 diverges by up to 1.49 — not a no-op
No fixture regenerated. nisps.wasm ships here because it is a tracked artifact
under the Phase 0 freshness gate.
Follow-up: the parity harness covers 2 of 8 engines. "Firmware and WASM share
the same engines" is asserted repo-wide but tested narrowly.
Phase 3 (S1, S5, S6, S25, L11, L37, ST11, ST12). Behaviour-preserving by
construction: the diff on the generated directories is PURELY ADDITIVE (217
insertions, 0 deletions), so no emitted constant changed value. This moves where
truth lives; it does not change what truth says.
- S5: nine mode headers each carried a mechanically identical 12-field
positional ParamSchema aggregate. codegen now emits one
`inline constexpr ParamSchema k<Mode>Schema` per mode, and the struct itself
moved into generated/schema_types.hpp. Each param_schema() is a one-line
return.
- S6 + S25: every mode hand-typed its net shape a second time as MLP template
args, duplicating the schema's own dims. codegen emits a `<Mode>MLP` alias
built from the already-emitted constants (not re-literalled), and all nine
modes use it. NMaxExamples still defaults from kDefaultMaxExamples (Phase 2).
- S1: model.ts hand-imported all nine schemas by name and hand-paired each with
its overlay — so the SET of modes was hand-maintained and could silently drift
from codegen. codegen now emits ALL_MODE_SCHEMAS; SCHEMA_MODE_OVERLAYS is
purely display truth (label/glyph/css/order), which stays hand-curated.
- L37: deleted the hand-written modeEngineId switch, which duplicated
schema.engine_id and silently defaulted unknown modes to 'thru'. Routes on
MFMode.engineId with exactly one documented exception: sound_analysis_midi
declares engine_id 'thru' because its ModeBase audio slot really is
NoOpEngine, while it separately drives the real AnalysisEngine.
- L11: ExternalSynthMIDIMode's shape was literal in two places; now named once
in an ext_synth_defaults namespace with an ExtSynthMIDIMLP alias. Full folding
into the JSON pipeline is NOT done and the reason is recorded in-file: it is a
template family over an externally-supplied Device and variable NOut, with no
single (device, NOut) schema to author.
- ST12: extracted codegen/lib.ts for the helpers both generators duplicated, and
corrected the comment that claimed they had to be separate. Proof the
extraction was behaviour-free: regenerating the MIDI-device outputs produces a
byte-identical tree.
- ST11: deleted codegen/templates/ — dead "reference" files no generator reads,
already drifted from the real emitters.
Gates: run-all-tests.sh ALL GREEN; codegen idempotent (re-running both
generators yields no further diff), which is what CI's dirty-diff gate checks.
Phase 2, S35. Two real defects from one root cause, both confirmed by trace
rather than taken from the audit:
1. Divergence. WasmIML built its TS Dataset mirror with a cap of 100 while
every addExample() ALSO pushed into the C++ FIFO ring, capped at 128. Since
train() reads the C++ ring and trainAsync() reads the TS mirror, past 100
examples the two trained on different datasets — silently.
2. Latent OOB read. nisps_ml_train sizes its sample-weight span by the C++
side's example_count() (up to 128), but wasm-iml.ts allocates that heap
buffer from the TS dataset's size (<=100). Once the ring exceeds the mirror,
the span reads past the end of the caller's allocation.
Fix: name the capacity ONCE as nisps::ml::kDefaultMaxExamples = 128, used by
FixedStorage's default template arg, DynamicStorage's default ctor arg, and the
MLP<> alias (which is the only real FixedStorage instantiation path and carried
its own independent 128 literal — the last copy of this dual truth). Expose it
through nisps_ml_describe and have the TS side read it instead of hardcoding.
Dataset's constructor default is removed entirely: a default was what invited
this bug class, and the sole call site now always supplies the describe() value.
ABI NOTE: this extends nisps_ml_describe from a 6-int to a 7-int descriptor.
nisps_ml_describe always writes 7 ints regardless of the caller's buffer, so
every call site had to grow in the same change or it would overflow the WASM
heap by 4 bytes per call. All five sites updated: three in wasm-iml.ts (init
defaults, init per-instance, reshape re-describe — the finding said there were
two), one in wasm-worker.ts, one in tests/cpp/parity_wasm.mjs. The parity
harness's expected-dims check now also pins the new max_examples slot.
Regression test: tests/cpp/test_mlp_storage_defaults.cpp — pins the two storage
policies to one constant, and drives MLPCore<DynamicStorage> exactly as
bindings.cpp does past the old TS cap, asserting it saturates at 128 and not at
100. Fail-before/pass-after confirmed by temporarily setting the constant to
100: 2 failures, named. Reverted: green.
Audit correction: the cited dataset.ts:81 is the FIFO eviction check; the
hardcoded default was at dataset.ts:45.
Gates: run-all-tests.sh ALL GREEN, parity PASS.
Phase 1 group 4 (S33, S34, L54).
- S33: removed 12 dead entries across the full 5-layer registration chain
(bindings.cpp KEEPALIVE -> EXPORTED_FUNCTIONS -> the NispsModule declaration
table -> the WasmIML wrapper -> the EngineApi facade): nisps_ml_reset,
example_count, move_weights, feedback_learning_paused, feedback_drag,
jolt_lr_scale, jolt_tick_lr_ramp, pipeline_state_size/save_state/load_state,
feedback_placing and feedback_state. Each was grepped against manifold/src,
manifold/tests, the e2e specs, manifold/tests/wasm-load.ts and
tests/cpp/parity_wasm.mjs — the parity gate builds its own API via cwrap and
is a real consumer, so it counts.
KEPT deliberately: EXPORTED_RUNTIME's heap views + ccall/cwrap (the parity
harness and wasm-load.ts depend on them), and nisps_ml_feedback_static_output,
whose C export IS called directly by parity_wasm.mjs even though no TS
wrapper reaches it. Also dropped parity_wasm.mjs's moveWeights cwrap, which
was declared but never invoked.
- S34: deleted the publishWeights_ channel — EngineSink.setWeights,
Spine.setWeights/weights()/liveWeights and every call site. It ran a C->heap
copy plus a fresh Float32Array allocation at up to 200 Hz into a field
nothing read. getWeights survives for persistence and the debug probe.
- L54: worklet loader — deleted the unused imports object, the 'c' branch,
exMap and the duplicate second loop, and replaced the silent `() => 0` stub
with one that throws, so a missing import fails loudly instead of returning
plausible zeros into the audio path.
manifold/public/nisps.{js,wasm} rebuilt with the trimmed export list (emcc
3.1.69, the CI-pinned version) and committed — the freshness gate added in
Phase 0 requires it, and the webhook ships this artifact to production.
Gates: run-all-tests.sh ALL GREEN, parity 1273 floats within 1e-5.
Operator-approved ABI change (P2 stop-point). The WASM MLP is now
MLPCore<DynamicStorage>:
- nisps_ml_create(input, output, hidden[3], n, seed) honours its args;
non-positive/null fall back to the historical 32→[10,14,18]→126, so
every pre-P2 caller (manifold, worker, parity harness) stays
bit-identical. Invalid/oversized dims (>4096) → null.
- NEW nisps_ml_reshape(ml, in, out, hidden, n, spread): fresh net at the
new dims, warm-started via nisps/ml/warm_start.hpp (overlapping region
copied; rest keeps spread init); feedback controller re-created (state
resets — reset-on-reshape modal is the front-end contract). Failure
leaves the old net untouched.
- nisps_ml_describe(ml, out): takes the handle; null reports defaults.
- FeedbackController got the same storage split: algorithms in
FeedbackControllerCore<FbStorage>; FixedFeedbackStorage keeps firmware/
tests source-identical via the old alias; DynamicFeedbackStorage (one
arena) sizes to the runtime net. Firmware .text unchanged (122692).
- MLHandle: per-instance scratch vectors; dropped the dead 2MB
batch_out_scratch.
- TS: types.ts decls (+_nisps_ml_reshape), wasm-iml re-describes the
created instance, worker carries a shape-contract note for P2.3.
Verified: ctest 4/4 incl. new warm-start grow/shrink test; reshape ABI
smoke (dims honoured, overlap survives, invalid rejected, outputs
bounded); parity PASS unchanged (2.4e-7); lint clean; manifold 9 unit +
20 e2e green; firmware .text 122692 (+0.30% vs pre-P2 baseline).
Algorithms (forward, backprop/SGD, init, move_weights, diagnostics) now live
once in MLPCore<Storage> (nisps/ml/mlp.hpp). Storage models:
- FixedStorage (storage.hpp): template-sized std::array, zero heap. The
classic MLP<NIn,H1,H2,H3,NOut,...> is an alias preserving kInput/kHidden*/
kOutput/kNumLayers/weight_count() constexpr — firmware + bindings + modes
compile unchanged.
- DynamicStorage (dynamic_storage.hpp): runtime dims, ONE arena allocation
at construction, nothing per-call. #error under NISPS_TARGET_EMBEDDED
(new macro in core/perf.hpp); sole lint-cpp.sh heap-allowlist entry, plus
a lint check that fails if the #error guard disappears.
Verification:
- new ctest test_mlp_storage_parity: fixed↔dynamic BIT-identical across
init/draw/inference/train(FIFO)/move_weights(pin mask)/eval_loss/
layer_stats/set_weights/infer_batch/reset; invalid+moved-from inert
- golden ML vectors (pre-refactor constants) pass → bit-stable refactor
- native↔WASM parity PASS, max delta unchanged (2.4e-7)
- chokepoint B compile: PAFSynth .text 122324→122692 (+0.30%, ±1% budget);
RAM +416B (eval scratch)
- fix: firmware-common.sh used bare 'python' (absent here) → ${PYTHON:-python3}
Part of one-core-engine-refactor P2. nisps_ml_create ABI untouched (P2.2 is
an operator stop-point).
Wire the modular input layer into the Console and reshape the browser
engine so input axes are genuine independent dimensions.
Inputs (manifold/src/inputs/):
- gamepad-source: emit press+release edges with standard-mapping labels
(enables hold-and-move); single/double-stick already present.
- midi-input-source: single-device selection + batch "MIDI Learn"
(every CC swept while armed becomes an axis); notes stay discrete.
- input-layer: compose() forwards each axis 1:1 (no mean-blend);
add onReducedInput so the manifold tracks gamepad/MIDI position.
- types: InputAction.phase, InputMode.
Console (manifold/src/console/):
- ConsoleApp: bind gamepad buttons to verdicts (RB up / LB down /
X randomise / Y nudge / B undo / A-hold reposition); mirror composed
position onto the manifold.
- Drawers: rebuilt Inputs drawer (source picker, gamepad legend, MIDI
device picker + batch-learn flow, learned-control meters).
Engine (nisps/wasm, manifold/src/engine):
- DefaultMLP widened MLP<2,..> -> MLP<32,..> (32 = MAX_AXES); each
active axis gets a dedicated slot, unused slots held at 0 (inert).
Rebuilt nisps.wasm (playground + manifold).
- spine/engine-api: setInputs writes the full N-D vector (was dropping
arr[2+]); primary pair keeps the 2-D pipeline; process() re-ticks the
whole vector via spine.reprocess().
Tests:
- parity_check/parity_wasm: ParityMLP -> 32 inputs, widen example bufs.
- CMakeLists: build parity binary with -ffp-contract=off so native
matches FMA-free WASM (training amplified the gap past 1e-5).
Inputs dock is still an exclusive picker; mixing toggles, reshape modal,
and the >2-D slider view (inputs-spec.md) are groundwork-laid but not
yet wired. See docs/redesign/midi-gamepad-inputs-worklog.md.
Adds the Manifold app (manifold/) — the convertible-mode React front-end on
the real NISPS ML+audio engine, served live at meml.lnfinitemonkeys.org/next/.
Console chrome trimmed per UI pass:
- drop mode label + subtitle from the top-left overlay (keep MEMLNaut wordmark)
- remove the composite split preset/ratio readout (top-centre)
- remove the OUTPUT corner tag above the bars
- remove the A/B compare toggle from the verdict cluster
- remove the follow button + input/noise readout (bottom-left)
- remove AltitudeNav focus switcher (bottom-right)
