Design notes — Part XII — v3.0, Phase 13: Suzu foundation and the §13.7 deferred physics (steps 78–82: the four-lane engine and the ×4 oversampler, the membranes and the plate, the stick-slip bow, the bore v2, the editable table and the Scala tuning — opened 2026-10-18 at step 78; folded 2026-10-19; formerly `DECISIONS_12.md`)

Opened 2026-10-18, the day Phases 11 and 12 were folded (Parts X and XI); merged into DECISIONS.md as Part XII at the phase’s fold. The spec is PROJECT_SPEC.md (v5) §13 with TO_PROJECT_SPEC.md §13.8–§13.12 as the phase’s draft (the proposals, corrected by each step). THE CORE IS FROZEN from this phase to step 92 (libsumi 1.12.0): a core change is a bug → regression test → fix. The instruments: tests/voxo_suzu_tests.cpp, the storm (midi-sink --dev --voxo-source suzu --voxo-storm <s> from the Release build), tests/voxo_fuzz.cpp, the sound profiles, the web gate tools/suzu_web_gate.mjs and the lab gate.

1. Step 78’s working rule, recorded BEFORE the refactor: the web gate’s contract under SIMD (2026-10-18). The gate (DECISIONS_7 #29) compares the lab’s wasm against a native reference built from the same sources with no fused multiply-add (voxo_nofma, -ffp-contract=off -O3), bit for bit for every voice kind whose path calls no math-library function whose last bit differs between the platforms’ libraries, within a declared tolerance for the kinds that do; and reports the wasm against the shipping desktop build in dB, as information. Restated for the SoA SIMD engine: the reference build is the SCALAR path — voxo_nofma compiles with VOXO_NO_SIMD (the one header’s scalar fallback) — and the wasm compiles with -msimd128; the gate stays BIT FOR BIT for the chain and the lattice, because the SIMD paths perform the same operations in the same order per node and per mode (the Laplacian’s two subtractions and their sum in the scalar’s order, the kick’s product then its add, the rotation’s two fused-or-not multiply-adds, the per-node damping before the drift), horizontal sums are taken in scalar order from the stored lanes, and the header’s mul_add is a plain multiply then add wherever VOXO_NO_FMA is defined (the reference and the wasm, which has no FMA) and a fused one elsewhere (the shipping native builds, whose scalar code the compiler already contracts). If the comparison turned out not bitwise on any kind, the gate would become a TOLERANCE gate for that kind with the tolerance declared here — it did not have to (#2). The shipping native builds keep their dB report. The strobe of step 75 is compared before and after on the shipping build’s own records (suzu_web_reference_desktop, captured before the first line moved) and on the strobe returns the suite dumps — bit-identical where the compiler’s contraction of the scalar matched the intrinsics’, the last bit elsewhere, reported either way. The lane widths: four on NEON (arm64: the Mac, the iPad, the Tab), four on SSE2 (the x86-64 baseline the Windows and Linux lanes build — AVX2’s eight lanes would need a flag the release lanes do not set or a runtime dispatch; not taken: the roadmap’s “AVX2 on x86-64” is read as “the x86-64 path”, SSE2 being what ships) and four on WASM SIMD128; one header, voxo/src/simd.h, one source.

2. Step 78 — the four-lane engine and the ×4 oversampler, as shipped (2026-10-18; Voxo 0.18.0, additive; evidence docs/evidence/step78/). The contract of #1 held without a tolerance: the suite’s twin comparison is 100 % bit-identical over 4800 sub-steps for the chain (80 nodes) and the lattice (16 modes, a fold shear, a muted mode), the web gate’s SIMD128 wasm against the scalar no-FMA reference stays bit for bit on the cell, the chain, the hybrid, Duffing, the rotor and the sax (the lattice, the flute and the trumpet within their libm tolerances as before), and BOTH reference records — the no-FMA scalar and the shipping FMA NEON — are bit-identical before and after the refactor, as are the twelve strobe return sets of step 75. The fusion matched because the intrinsics follow clang’s contraction rule (the first product of a·b + c·d, the product of a − b·c): simd.h documents it.

   • What the chain cost. Sixteen chains of 80 nodes at 3.3 % of the callback on the Release build (the gate’s bar 6 %), beside sixteen cells with their shears at 3.5 %. THE BARS ARE THE RELEASE BUILD’S: the Debug build/ left the SIMD wrappers as calls (ten times the scalar, 150 % of the callback for the chains) until they were forced inline, and still measures with checks in; the suite asserts the headroom bars under NDEBUG only and prints the figure otherwise. build-release (Release, arm64, tests on) is where the number is read; build-universal has no tests and runs the storm. The structure-of-arrays across voices the roadmap named was not needed: within-chain lanes clear the bar.
   • The lane widths. Four everywhere (NEON, SSE2, WASM SIMD128), as #1 said; AVX2 not taken.
   • The oversampler’s scope. oversample (2 default, 4) is the CELL FAMILY’S — the cell and the lattice with their shears, kinds 0 and 1 — because that is where the fold is; the chains, the bores and the chaos voices keep the 2× section their grids are tuned to. The ×2 section is untouched (the two-tap average, bit for bit); ×4 adds two cascaded 2:1 half-band stages (suzu::HalfBand; 47 taps, Kaiser β 5.65, the stopband 64 dB down — the 39-tap design read 59 dB, a hair under the gate, so 47) and threads the section’s rate through the strike (ε, the decays, the bow’s rate, the release) and the block (the SVF, the retune, the compensation), the mode ceiling pinned to the OUTPUT Nyquist. The strobe’s sub-steps stay the 2× section’s: at ×4 a wrap lands in the first or second quarter with its fraction rescaled (at ×2 the mapping is the identity, bit for bit).
   • The alias gate as measured. The fold at full drive at C7: the floor in [12 k, 19.2 k] Hz is −112 dB under the fundamental at ×4 and −90 dB at ×2 (the red): both far under the roadmap’s −60 — the symplectic fold’s harmonics fall fast, so the aliases were never loud; C4 reads −160 / −129. The gate stands as the roadmap wrote it, with the red’s gap (≥ 10 dB) as the test of the decimator; the voices that need the ×4 section (the Toda ring, the billiard: Phase 14) are why it exists.
   • Found by the lab gate, a step-75 regression: the bus BREAKER (DECISIONS_10 #12) cuts a non-finite block inside the engine — the voices end, the block goes out as silence — so the worklet’s own non-finite check never fires any more and the lab’s two blow-up red controls (the strings’ CFL bypass, the winds’ naive junction) read “not caught” since 0.17.0; the pre-78 wasm rebuilt in a detached worktree failed them the same way. Fixed on the web surface: sw_breaker_trips (Voxo’s breaker_trips) and the worklet treating a new trip as the blow-up it is (the restart, the page told). The lab gate is green on all six pages; it is not in ctest and nobody ran it at 75–77 — a working rule from here: every Voxo change runs the web gate AND the lab gate.
   • The storm from the Release universal build on three session presets (the chain at 80 nodes, the trumpet, the fold at ×4) at 0 XRuns, render max 0.19 ms; the fuzz a minute clean, the strobe fuzz 30 s clean; the prewarp helper within 1e-7 of eps_for. Carried: the oversampler on the other kinds (their grids would need their own tuning at the section’s rate), AVX2’s eight lanes with a runtime dispatch, the structure-of-arrays across voices if a future voice needs it. The tablets run the patches on NEON through the same header: the Android emulator (arm64, the Tab unplugged) ran the Bowed string, the Trumpet and the Flute under the spike at 0 device XRuns; the iPad’s run is built and staged but the device was locked at the install — the author’s to unlock, the commands in the evidence.

3. Step 79 — the membrane, the bus plate and the gong, as shipped (2026-10-19; Voxo 0.19.0, additive; evidence docs/evidence/step79/; the spec’s §13.9 rewritten as shipped). voice_kind 9 is a 2D symplectic leapfrog grid (suzu::Membrane<CAP>, four lanes a row behind simd.h, bit for bit on the web gate), and the bus plate is one Membrane<64> on voxo_t at the output rate. The decisions:

   • The boundary is simply supported (u = 0 and Δu = 0 on the ring): the biharmonic is the Laplacian twice through one zero ring, the mode table is closed-form, and the suite pins the (1,1), (2,1), (1,2) to it within a cent for the membrane and the plate (gate delta d). The grid is a 16 : 13 rectangle so the degenerate pairs split.
   • Pitch through c², the grid cut per note under the bar at E_max; the proposal’s “0.000 cent by scaling the grid” dropped (gate delta a). drum_kappa is the biharmonic SHARE of the (1,1)‘s stiffness, so the note stays on the (1,1) at any share and the CFL number is ε₁₁ times a factor of the grid.
   • The cap is exact on the plain energy (the momentum scaled to the remainder after every strike) and E_max = 2H_cap/(c² + κ²λ₁₁) follows by arithmetic — the roadmap’s 2H₀/c² tightened by the biharmonic term (gate delta b). drum_alpha is the swell the player hears (c² × (1 + α) at the cap), α_raw derived from it.
   • THE GONG’S BAR IS 1.80, NOT 1.95 — measured. A pure membrane at swell 1 displaced to exactly E_max at λ_2D 1.95 with the top mode seeded diverged: the (1,1)‘s stretching modulates the top mode’s stiffness by the whole swell at twice its own rate, and that close to the edge the modulation is not adiabatic (the linear gate is necessary, not sufficient). 1.90 held for swells to 2, the κ share lowers the depth, the gong patch’s real strikes held at every note; the bar for a swelling sheet is 1.80 (MEMBRANE_CFL_GATE_GONG), the linear sheet keeps 1.95. The reds as the roadmap wrote them: 1.2× the cap’s stretching on a sheet tuned to 1.95 diverges within 100 steps; the ABI’s forced 1.05 is refused, and bypassed trips the breaker within a second.
   • The damping on the momentum at twice the rate, so decay_s is the amplitude’s T60 as everywhere else (read back 1.19 s for 1.2); decay_bright the lattice’s law per (λ/λ₁₁ − 1). The release multiplies the held contraction (the momentum’s factor squared).
   • The plate at 1×, clamped not refused. The bus’s rate is the output’s; the plate’s share and then its pitch are clamped under the gate with a log line (a bus element never leaves the patch standing) — a 64² plate at 48 k reaches about 360 Hz and is near-membrane (the share 0.003 at 1.48; gate delta c: 32² on a VOXO_MOBILE build). The voices inject their own samples at their points (a DRIVE, bounded) scaled by (n/64)², measured so 32² and 64² read alike; the mix at 0.045 puts the plate 10 dB under the dry harp chord. The point is the shell’s: voxo_set_strike_point before the note-on — the iPad and the Tab hand the layout cell’s centre from their probes (the roadmap’s sumi_layout_cells does not exist), the desktop’s keyboard and a wire take the default map (the note along x, a golden-ratio scatter along y).
   • The drift gate on the leapfrog’s exact invariant H̃ (the staggered quadratic form), not the plain energy whose staggering wobble is 0.27 dB: −0.00016 dB over ten minutes.
   • Found beside it: the bench’s --preset applied a session preset to the engine alone (dev_apply_preset), so a storm “on a preset” played whatever the Sound settings held — step 78’s three storm lines were the settings’ patch, not the chain, the trumpet and the fold. Now the preset’s Suzu block reaches the settings for the run and the file is left unsaved; the three storms of this step ran the drum, the gong and the soundboard (0 XRuns, render max 0.79 ms). And the Android DEBUG APK’s engine is −O0: its spike read 60 ms of render and hundreds of XRuns; the release APK signed with the debug key reads 10 ms and 0 device XRuns on the drum and the gong, one on the soundboard (the emulator’s audio path; the Tab is unplugged). The iPad (M4) runs the three at 0 XRuns, render max 0.58 / 0.65 / 0.35 ms of its 2.667 ms buffer — a CSV copied while the device was locked was a stale run’s: a devicectl copy answers even when the launch was refused. Cost: sixteen 16 × 13 membranes 16.5 % of the Release callback, the 64² plate 16.3 %, 32² 4.5 %. The patch table holds fourteen: Drum, Gong, Harp on a soundboard, pinned to their JSON files by gate 28. Carried: the Tab’s own run, the author’s choice of the drum and gong references the profiles are read against (the profiles are the patches’ own), the lab’s membrane page (step 92), a strike profile from the layout cell (step 80), the AVX2 lanes.

4. Step 80 — the stick-slip bow and the strike position from the cell, as shipped (2026-10-19; Voxo 0.20.0, additive; evidence docs/evidence/step80/; the spec’s §13.10 rewritten as shipped). bow_mode 1 is suzu::StickSlip: the hyperbolic friction curve solved implicitly against the point’s mobility, bounded by the sticking limit, on three bodies — the lattice’s bowing point, the Verlet chain’s node and the hybrid string’s loop (two taps against Z = 1: the textbook waveguide bow). The decisions and the findings:

   • The hysteresis is part of the model. Without the stick/slip branch (McIntyre–Woodhouse) Newton from the free velocity picks the slip root every step and every body whistles; with it the chain sticks 70 % of the period at the default force on the note, the waveguide 48 % rising to 68 % at twice the force. The branch is the voice’s state (bow_stuck).
   • The bow acts between the kick and the drift. Applied after the drift a strong bow pinned the chain’s node into a static deflection (the string’s kick cancelled the bow’s velocity before the node moved: the tone read as a tone a sixth low with a buzz). VerletString::kick / drift split the step; the fused step is bit for bit.
   • The clean Helmholtz lives on the delay line. The chain’s lattice dispersion (its partials as sin of the mode number) smears the corner: its stuck share climbs gradually and its pitch wanders a dozen cents with the force. The hybrid’s loop is dispersionless; the bow there is two taps on the ring — the wave about to pass the bow and, inverted, the wave from the bridge side — and the loop gains exactly f·v a sub-step. The roadmap’s “the string’s or the mode’s impedance” is read as both the node (G = 1) and the waveguide’s 2Z (G = ½). The patch Bowed string (stick-slip) is the waveguide; Bowed chain the chain.
   • The ledger’s gain is read at the synchronized velocity. The leapfrog invariant’s change under a kick at the half-step is Δp times the mean of the half-step velocities around it; read at the post-kick velocity the chain’s gain summed to −5.3 against 4.1 of work with the energy flat — impossible. With every loss zeroed the balance is now exact per quarter second on both bodies (gain_probe.cpp).
   • The red is the raw Newton, not the explicit force. A converged implicit root never crosses the bow, and the explicit force’s excess is friction heat; neither crosses the roadmap’s line E_held ≤ ∫F·v_bow, which sits a hundred times above the energy held. The tight line is per step — what the string takes, ∫f·v, against the bow’s work, ∫f·v_bow — and the three Newton steps taken raw on the folding flank cross it: 0.156 against 1.46 on the chain, 7.0e4 against 6330 on the waveguide. Both lines are the gate.
   • The controls: the mouth is the speed, the press and the velocity the force (×2 at a full press, ×½ at a soft touch) — §13.3’s “press as a sustained drive” is this. The scales measured: KV 2.3 / 1 / 0.45 and KF 6 / 1 / 15 (the chain / the lattice / the waveguide), v₀ a tenth of the speed, μ_s 0.8, μ_d 0.3. bow_position is the bowing point on the strings and the profile on the lattice.
   • The web gate drops the lattice under the bow: the branch is a discontinuity, and the lattice’s per-sample sinf flips one at a sub-step, moving the snapshot’s decimated trace past any tolerance while the audio holds within −80 dB; the chain and the waveguide under the bow are bit for bit. The first twelve sections’ audio and traces are bit-identical to the step-79 reference (the inspection records carry the new k slots).
   • The strike position from the cell (voxo_set_strike_pluck; the tablets’ touch-begins map the distance from the cell’s centre, the centre the middle, the rim the bridge) replaces pluck on the chain, the plucked lattice and the membrane; the bar keeps its fitted 1/m. Gates: the suite’s 39–41 green (the stuck share rising with the force on both strings, the ledger’s two lines, the red on the chain, the lattice on the note, the servo bit for bit under the new knobs); the web and lab gates green; the storms on both patches at 0 XRuns (render max 0.26 ms); the fuzz clean; ctest 14/14; the chart bow_force and the two profiles. The emulator’s release APK ran both patches at 0 device XRuns (the spike has no press: the bow lifts; the load’s evidence); the iPad took the build and locked before the launches. Carried: bowing under the finger on both tablets is the author’s; “a struck bar’s” strike profile (the VCSL fits would move); the lattice’s pitch pull under the bow; the raucous pin on the chain above the playable range; the bowed lattice in the web gate.

5. Step 81 — the acoustic bore v2, as shipped (2026-10-19; Voxo 0.21.0, additive; the core frozen). The spec draft’s §13.11 is rewritten as shipped; the proposal’s four items each took a correction on the probe.

   • The toneholes are branch ports, and the pickup had to move. A hole is an inertance kicked by its node’s pressure at the half step (as u), radiating through a one-pole resistance at twice the bore’s corner, its flow in the node’s continuity; closed, a wall. The first build read the bore at the far end as before, and the holed flute’s fundamental sat 40 dB under its fifth partial — the oscillation inside was right (the pitch landed), the far end sits past the lattice where the fundamental is evanescent. The output is now what leaves the bore: the end’s velocity plus the open holes’ flows. Found on the way: a flat hole resistance ate the fundamental (the hole’s R was above its reactance at the pitch — the one-pole form of the ends’ fixes it); the hole’s size must be physical (1/L_h = S/8 at the node: a flute’s hole is eight cells’ inertance; at S/1 the scheme grew at 0.99 of the bound and the holes cut nothing cleanly); the hole mode takes room under the CFL bound (Gershgorin, λ ≤ λ_bound/√(1 + h/4): at 0.999 of the bound the third register’s bores grew); a cone’s holes count the apex (the sax’s corrections saturated at +300 cents until they did); a hole sits at a fractional cell, split over two nodes (the top register’s semitone is a cell and a half); the vent is a real register key (90 cells’ inertance at an eighth of the length — a quarter-hole at a third detuned the second register by 230 cents), its detune calibrated per register.
   • The calibration is the woodwind maker’s, per register, and the sax’s is blown. Each fingering pulsed on the register’s bore, the first peak read, the hole moved by the miss, twice; on the sax the passive peak and the reed’s note differ by up to a quarter-tone on the hole bores (not the plain bore’s pull), so the second pass BLOWS each fingering (0.4 s: the upper register takes a third of a second to settle; a 0.25 s blow left the second register 50 cents flat). The cost at patch load: 228 ms (the flute), 770 ms (the sax), cached per patch slot by the holes’ parameters. The holes serve two registers (D4..C♯6); the third is past the lattice’s cutoff — a hole’s impedance is the bore’s and it no longer cuts (measured: −60…−100 cents however the holes were moved) — and takes the v1 cut, as does any note the keys disagree with.
   • The quasi-periodic A3 was the bore’s slow mode, not the reed’s inertia. The proposal’s L_r·dq/dt ships as proposed (dflt 0) but barely touches the ±16 Hz sidebands (L_r 8 cleans breath 70 only, 30 degrades). The one-pole radiation port leaves the bore’s slow modes undamped; a resistance on the bell’s slow flow (a one-pole at 40 Hz, z_l 0.4, positive-real) takes them: −62…−63 dB at breath 50/70/100 (the red: +13 dB). Its reactance is a negative end correction the cut counts (uncounted: the flute +40 cents sharp). AND THE LEAK IS THE SAX’S ALONE: on the flute’s two open ends and the trumpet’s bell it cost the bass (22 dB at the flute’s C2, 7 at the trumpet’s) — the leak at 65 Hz is a hundred times the radiation loss there. The flute’s profile is back within 7.8 dB over C2..C7. And a leaky bell takes room under the CFL bound (0.96): the sax at G7 — thirteen cells, λ 0.989 — blew up under z_l 0.4, a full-scale spike at one note of the profile that no gate reads (the sound profile does, again).
   • The shock is dissipative by a flux limiter, and the local-CFL story was not the ringing’s cause. The local Courant number on both sweeps (on the momentum sweep alone the steepening was too weak), saturated at 20 P_REF with the base λ leaving room; the Lapidus flux limiter’s loss logged in the ledger (k[10]). Measured at C4/127: the blare +10 dB over the v1 bore in the 6–20 kHz band. THE RED moved under the probe: with the leak at the trumpet’s bell the limiter off had rung 23 dB above the limited in the output’s high band; without the leak (where it ends) the ringing stays in the bore — its profile 1.46 rough against 1.01 limited, the bell keeping it out of the output — and at twice the steepening the unlimited shock is refused by the valve gate at patch load (the junction makes energy with the losses zeroed, 1.1× the mouth’s work: the 58c gate caught it) while the limited one is admitted and holds. The gate reads both. The trumpet’s λ sits at 0.28 (its cells capped), so the ringing is the dispersion at a shock, not a Courant crossing; the saturation stays as the safeguard. The roughness is a relative measure (a linear bore reads 0.70; read against the limited bore’s at the same note). Under the limiter more steepening dulls (shock 2: −32 dB, 4: −37): the blare peaks near κ, the default.
   • The shock is the trumpet’s; the sax keeps the v1 bore. Built on both valve winds, the shock cost the sax 80 % of its render for no claim the spec makes of a cone. And the trumpet’s storm threw XRuns at thirteen voices (render 2.0 ms of 2.67): the limiter’s passes vectorise but a running sum for the loss serialised them (790 µs a block against the v1 bore’s 380); the terms are written beside the updates and summed after (615 µs, the loss identical). And every wind strike ran the bore’s power iteration twice (inside the build on the fresh 64-cell bore, again on the cut one), 40–60 µs each: a block carrying fifteen strikes cost the trumpet 1.5 ms. The profile’s μ_max is taken once per patch (WindCal.mu64), the flute’s once per engine. The storms pass.
   • The two-mode lips ship as a knob (lip_shear, dflt 0): the registers hold at share 0.3; at 0 the one-mode lips render bit for bit under the ratio. The author’s ear decides the trumpet’s default.
   • The web gate: the sax joins the libm class (−80 dB) — its calibration’s log2f/exp2f on the pull; the leak moved the pull to values whose last bit differs (measured −102 dB). Gates: the suite’s 42–45 green with their reds (the sax’s sidebands, the shock’s ledger and its Gibbs red, the lips, the holes on both woodwinds: the chart’s twelve fingerings, the registers, the disagreeing keys); the web and lab gates green on the final wasm; the fuzz clean (the strobe 0 blocks differ); ctest; the storms on the three winds from the Release universal build (the trumpet 0.97 ms, 0 XRuns); the three profiles; the emulator’s release APK on the three winds at 0 device XRuns; the iPad’s three runs at 0 XRuns (render max 0.42 / 0.41 / 0.70 ms); the Tab at its 4 ms low-latency burst throws one to four XRuns on the winds — and did at step 80 too (the trumpet’s render max 4.0 → 4.3 ms there): twelve wind voices at 4 ms were at the Tab’s edge before the step. Carried: the hole mode’s third register; the sax’s hole intonation to ±10 cents; the lab’s shock-limiter and leak reds on the winds page; the tablets’ key panels driving the holes (the Boehm layout’s keys reach the engine already; the mode is a patch knob); the Tab’s wind headroom at 192 frames (a mobile cap on the voice count, or the spike’s burst, is the author’s call).

6. Step 82 — the editable modal table, the Scala tuning and the unison pair, as shipped (2026-10-19; Voxo 0.22.0, additive; the preset schema 2; the core frozen). The spec draft’s §13.12 is rewritten as shipped.

   • The table is modal_preset 5 with three arrays beside the knobs (ratio, t60, weight; a zero defers to the formula), read by the one table builder every call site already used; voxo_suzu_table hands any preset’s resolved table out so a shell edits from the formula’s values. The compensation becomes a flag (modal_comp): a unison pair has no separate targets for the solve — the load gate’s spectral radius ran the solve on the raw ratios, diverged (16.6 against 4) and refused the patch until the raw bound was its own; and under the raw lattice THE SWIRL ADDS NOTHING to the coupling, for the same reason the compensation exists (0.5·κ would detune every mode).
   • The split is the chain’s, not the pair’s. The proposal’s figure assumed √(1 + 2κ); the first pair’s two normal modes sit at +κ(3 − √5)/4 and +κ(3 + √5)/4 of the frequency, one mode having one neighbour and the other two — measured +3.25 / +22.50 cents at κ 0.01 against +3.30 / +22.37. The patch’s κ 0.006 beats 11.6 cents apart.
   • The schema is 2. The suzu block’s arrays and modal_comp ride as JSON arrays; a schema-1 reader keeps what it knows. The preset test’s three schema-1 assertions moved. sumi_preset_suzu_to_voxo is the one file-to-params mapping with every step’s sentinel, in the presets library (types only); the gate pinning the patch table to its files now reads the arrays element for element and skips the serializer’s own keys (the trace routes, oversample) — the unison file is the first one written by the serializer itself.
   • The tuning is one site. voice_retune was already the only place a note became a frequency, for Suzu and the sampler alike; the Scala table sits under it, double-buffered, the bend an equal-tempered offset on the tuned note. The parser takes the files’ text (the tablets read through their pickers) or their paths; it refuses with a reason. 19-EDO reads 0.0001 cent off the formula over the keyboard, the cell 0.0003 at the oscillator; Pythagoras’s fifth 1.5000000 under a .kbm.
   • The tablets expose the table and the file as files: a Suzu patch .json through the document picker (the table inside it) applied over the table’s patch, and a Scala .scl, both copied beside the instruments and persisted; a sixteen-row editor stays the desktop’s. Gates: the suite’s 46–48 green (the table resolved / inspected / at the oscillator with its decays; the Scala table, oscillator, mapping and refusal; the unison’s two lines and its deaf swirl); preset_tests 40/40 (the round trip whole); the patch table’s seventeen pinned to their files; the web and lab gates green; ctest; the storms; the unison’s profile; both tablets under the spike on the unison. Carried: a table editor on the tablets; the tuning inside the session file; the lab’s table; §14.6’s microtonal mapping (the core’s, as the roadmap says).

Phase close — 2026-10-19

Steps 78–82 shipped (#1–#6), each libvoxo additive (0.18.0 → 0.22.0) and the core frozen throughout; TO_PROJECT_SPEC.md §13.8–§13.12 stand as shipped (their fold into PROJECT_SPEC.md is the release’s). The evidence folders of the five steps left the tree at this fold (git show <commit>:docs/evidence/step78…82 keeps them). The tag v3.0.0-alpha.3 is the author’s to cut. Phase 14 opens DECISIONS_13.md (Part XIII) at step 83.