Why this version

Why old versions are still here

The project convention is that an improvement ships as a selectable version and the old one is never deleted. That is not sentiment. Every one of these rewrites was justified by an A/B that a person had to be able to run, and several of the old paths still own mechanisms the new path never ported. A toggle that is only ever left at its default is a toggle that was never worth the cost; each section below therefore ends with a concrete reason to move it.

Almost every entry in the changelog carries "verified": false. That flag means the numeric bounds were machine-checked and the build ran clean, but the question of whether the thing reads right was deferred to a human looking at it. Treat the numbers as measured and the aesthetics as claimed.

What the badges mean here

A generation marked superseded is one where the only remaining reason to select it is diagnosis or demonstration. A generation marked selectable owns behaviour the newer one does not have. The distinction matters: several of the nine families below have no strictly dominant generation at all.

Why the current backend is good enough

It removed a cause instead of adding a compensator. The field is now a closed-form sum of at most 32 softened 2D pole kernels evaluated per particle: exact at every distance, zero solver cost, zero lag, and structurally incapable of truncating. Once the field was exact, all five gates became unnecessary at once, because every one of them had been tuned against the truncated field. The ring had an outer edge (Jacobi never reached there) and an inner edge (hf_suppress and sink_cancel closing at the pole and inside blobs); both edges disappear when the field is right. What survives as force shaping is one per-material acceleration cap, so the falloff you see is physics rather than tuning.

The collapse those gates were also fighting was independently re-diagnosed as a Langevin feedback runaway rather than mechanical compression — M read off the full field grows M — and that loop is now broken structurally rather than damped. The dead numerical seed is gone at the source. Holding became a spring-damper, so a stationary held magnet holds instead of braking.

It is also two orders of magnitude cheaper: seven cold-start 1024² solves at 52 to 96 iterations per frame (roughly 380 to 700 million Jacobi cell updates) became one warm-started 256² solve at 32 iterations (about 2.1 million), and that solve accumulates convergence across frames instead of throwing it away.

What is still wrong

• The biggest one is a regression the other docs do not mention. Selecting New or Fantasy sets suppressed_by_new_backend, which makes physics_active() return false, which sets u_use_real_magnetics = 0 in both p2g and g2p. That silently switches off five of the seven anti-compaction mitigations the status page still advertises: the demag pressure, the occupancy anti-collapse pressure, hf_suppress, the positional cursor gate and the per-cell M cap (CELL_M_CAP = 28), plus the legacy surface force and chain force. Under New and Fantasy, compaction is held by the unconditional Tait j_floor alone (0.85 ferro, 0.90 heavy, i.e. 10 to 15% maximum compression) and one accel cap. Ferro compacts more than physical under a strong source is therefore not fixed by the current backend; the stack that used to fight it is off and a blunt clamp replaced it. The project's own remedy is a slider.
• A world-up bias. With poles off and ambient zero but self-field on, field_dir falls back to a hardcoded (0,1) and chi_eff = dot(m, field_dir)/M_REF collapses to m.y/M_REF. A remanent body whose M points sideways gets near-zero force; one pointing downward is repelled from its own gradient.
• The demag solve lags. Warm-started but never converged (32 iterations, 256² over the whole MPM grid, far coarser than the legacy 1024²). Chain spacing and spike sharpness are resolution-limited, and fast material outruns its own field. Truncation was traded for lag, not eliminated.
• Silent pole overflow. kMaxPoles = 32 with an early return; permanent magnets cost two poles each, so past about 16 magnetic SDF objects the extras contribute literally nothing, with no warning.
• Soft iron does not self-consist. It is a single-pass induced estimate evaluated in pole-insertion order, so two soft-iron pieces near each other never mutually magnetize. The legacy ran three induction passes for exactly that.
• Fantasy's anti-flicker is a ratchet, not hysteresis. There is no per-material coercive field and no switching model, so the 0.15 / 0.60 thresholds are global constants applied identically to a hard magnet (saturation 22) and Para Mist (1.0). A latched M can rotate but effectively cannot shrink along its own axis under the self-field alone, so a body can hold a physically unjustified magnetization for a long time. Fantasy also runs self-field 1.6 and chain boost 2.2, which makes it the generation most likely to over-compact.
• The default is still Legacy (g_magnetic_backend = 0). Out of the box you get the ring, despite the consolidation proposal recommending New as canonical. Only the combo, or a scene preset carrying "magnetic_backend": 1, switches it.
• Doc drift, code wins. The backend combo tooltip still claims New does not yet do per-particle remanence, Rosensweig spikes, or the field overlay — all three shipped. The status page's Heavy Ferrofluid row still lists saturation 14 and alpha 3.0 while both shaders use 8.0 and 2.0 (its own caveat list, four lines above, has the correct numbers).
• A deliberate 2D compromise. The pole kernel falls off as 1/r for a monopole and 1/r² for a dipole pair, not the 3D 1/r³. Consistent for a 2D world, but no ring is partly bought with the magnet reaches the whole world.
• Vestigial, harmless. magnetic_rasterize.comp still carries the old max(alpha*hmag, 1e-4) form. It cannot reproduce the clumping — its only call site is guarded by hmag > 1e-5 and applies to SDF shell pixels, not thousands of particles — but it should go.

Why Culture is good enough

It is the first generation where the pattern engine and the material are coupled in the direction that works. Colony asked spawning to be the growth (quantised, popcorn); expand+fill asked inflation to be the growth (never terminates); Automata asked the geometry to find holes the field never carved (8 firings in 21 seconds). Culture separates them: a properly resolved continuous automaton — 489 taps at radius 13, with occupancy confining and nucleating it but never overwriting it — genuinely produces membranes and holes; then a gradient force drags existing matter onto the ridges.

That is the whole argument. The porous structure is achieved by rearrangement, which is mass-conserving and therefore bounded by construction, and insertion becomes an optional budgeted extra rather than the mechanism. Sculpt mode cannot run away; grow mode's bound is a counter, not a tuned feedback loop.

The regrowth rule is deliberately asymmetric — living tissue heals to full unless actually hot, dead tissue converts back only where green is nearby — so green always wins at a boundary and a burn scar heals as a visible front. Every earlier heal/starve pair diluted into a grey equilibrium instead. The project confirmed the result in-app before building on it, and tagged a checkpoint at that state.

What is still wrong

• It cannot die of anything but heat. The starve rate is still passed in, and the shader marks it unused: living tissue is sticky, only burning kills — while the panel tooltip still tells you the body dies if you do not beat the starve rate. That tooltip is false.
• No kill switch. Culture has no cores and no energy budget, so it lost Colony's best mechanic: a permanently latched dead core. Any surviving green reclaims any adjacent white, so the only kill is burning 100% of the body.
• The changelog overstates the material transition. It says burnable tissue becomes a brittle husk material; the code never changes a material id — the husk is a vitality <= u_dead_floor branch (default 0.15) inside the tissue material, and the husk id is only ever placeable by hand.
• The revival field is not self-limiting, contrary to the changelog. Its source is the blurred living occupancy; the husk texture is read only by the stats pass. Ash gating happens per particle, not in the field, so the field is computed over empty air around any green tissue.
• Hidden coupling. One mask texture is used both as the field's confinement mask and as the revival field's green source, so moving Confine Radius silently changes regrowth reach. No tooltip says so.
• It only reads correctly at low gravity. The showcase scenes hard-code −1.5 g with the comment "so the porous structure shows". At −9.81 the membrane impulse (clamped to 1.8) loses to gravity and the tissue puddles.
• It is the most expensive backend: 489 taps × 65,536 cells per frame, 8 to 14 ms/frame against Automata's 6 to 8.
• Grow-then-hold is a lifetime counter, not physiology. The counter resets only when the module is disabled, so a body cut in half can recover vitality but never mass.
• Discoverability. The Creation menu still lists all eight living materials, but placing a Colony core in a default scene does nothing, because scene load forces the Version selector back to Culture and disables the other two modules. Only one tooltip mentions the Version dropdown at all.

Why v2 is good enough

The legacy failure was not a tuning problem, it was an ownership problem: two writers to one buffer, with the loser recomputed from scratch every frame. That is why the shader comment reads like a log of failed thresholds. v2 makes g2p the sole owner and leaves p2g's values as read-only intermediates for the current frame's stress. The second fix is equally structural: the neighbour rule had been gated behind a toggle scene loading reset to false, so the feature was unreachable in normal play; v2 forces the branch on and auto-enables the bio passes from spawn-time material detection, so it works without the user knowing a bio field exists. Aligning the temperature bands — the only part that looks like tuning — only became possible once there was a single owner to align.

What is still wrong

• "Same object" is a proximity heuristic. The changelog admits it: no per-particle object id exists. Two unrelated ash bodies touching will heal each other.
• The detection latch is sticky. It is cleared only by a full particle clear, so once any living material has existed in a session the Eulerian bio passes stay on for the rest of that scene even after the material is gone.
• It costs work the default configuration does not use. The detection list includes the Culture tissue materials, so placing Culture tissue flips the bio passes on — but nothing in the Culture backend reads that field. Culture has its own 256² revival field.
• Narrower scope than the name suggests. The v2 branch applies to the two ash-regrowth materials only. Morph Tissue and Cell Sheet are the neighbours that make ash regrow; they do not regrow through this path themselves.
• It is state recovery, not resurrection. No material-id transition and no particle spawn, so a fully destroyed region cannot come back.
• No measured regrow curve. Only a smoke-test recipe was recorded.

Why AnisoMPM is the right model, and why it is not the default

The single load-bearing change is the non-local term. Everything else — an energy-based drive, an elastic threshold, a monotone history, (1−d)² degradation — is correctness housekeeping that a careful person could have hand-tuned into the legacy branch. The Laplacian cannot be: it introduces a physical length l₀ into a model that previously had none, and that is what converts a threshold heuristic into an actual propagating crack with resolution-independent speed. It also makes the anisotropy real, because the structural tensor is built on the polar rotation rather than the deformation-gradient pushforward, which would contaminate direction with stretch.

It stays off by default because the shipped v1 had six real defects, all found by a re-run adversarial review after the first review was cut off mid-flight. The most instructive: the damage transfer was not weight-normalized — the scatter early-returned on d ≤ 0, so the denominator accumulated weight only from already-damaged particles, making the node value a mean over damaged particles that intact neighbours never diluted, and making the free-surface fallback fire for nodes full of intact material. All four review lenses converged on that one line independently. The gather's Laplacian was also a hand-rolled 5-point stencil centred on floor(cell_pos) while the scatter deposits on round(cell_pos), so for about half of all particles it was sampled a full node down-left, and the offset flipped discontinuously as a particle drifted across each half-cell line. A third defect is a good warning for this codebase generally: the per-particle damage aux was cleared only in init(), but the particle buffer is a bump allocator whose rewind does not re-run init, and the driving-force history is monotone — so a respawned particle could be born at maximum driving force and damage itself to failure under no load.

Storage deserves a note, because it is a reusable trick here: both p2g and g2p sit at the 16-SSBO driver cap, so the per-material fracture profile rides a samplerBuffer and the per-particle damage aux rides an imageBuffer — image uniforms live in a separate binding namespace. That escape hatch is now used by three subsystems on this page.

What is still wrong

• Enrollment is real work, not a data edit — and the first changelog entry said otherwise, which the correctness pass explicitly retracts. p2g has exactly two enrolled hooks, which is why exactly three materials work. Ceramic, splinter, tear-skin, laminated pastry, stoneware, composite and aniso each have their own constitutive branch with no hook, so setting the allow bit would enrol them in the scatter and gather — polluting the shared damage grid — while p2g never ran the damage step.
• The damage grid carries no material key. Damage can diffuse across a contact between two different enrolled materials under zero load.
• The free-surface test is a magnitude test. den < 1e-8 cannot express "negligible weight" given the 1e−6 fixed-point quantization.
• Φ is exactly invariant to thermal softening, because σ_c and the stress scale with the same moduli. That is arguably correct for a strain criterion, but the code comment claims otherwise, so one of the two is wrong.
• A claim in the record did not survive measurement, and it is left standing on purpose. The verifier argued Lap(d) was identically zero for an isolated damaged particle, i.e. that the regularization was inert. It did not reproduce: a seeded probe gives min −811 / mean −327 with the bug against −604 / −201 with the fix — nonzero and correctly signed in both, because the idealized isolation is destroyed within a frame or two. The real effect was that the length scale came out 1.5 to 2.8× too large, not that the term was dead. Fixed anyway, because it is not what the spec prescribes.
• Nine lower-severity findings were capped unverified when the review run hit a usage limit, and live A/B in the tearing-lab scenes — the actual calibration step — has not been done.
• Doc drift. The status page still says phase-field brittle has no Laplacian regularization and calls the AT2 PDE a future improvement of about 300 lines. That is essentially what shipped here, for the enrolled materials.

Why v3 is good enough

Because it is the first version that made the three changes together. Raising E alone blows up an explicit scheme — that was the audit's own caveat — so the substep scaling is not a performance afterthought, it is part of the fix. Raising E without moving the SVD floor would have left the clamp as the resting state, so the material would still have been frozen in a shrunken configuration, just a stiffer one. And doing it per family rather than globally is what lets bread stay squishy while a column holds: the soft-body list is shared with the existing stress-ceiling branch, so the two agree by construction rather than by coincidence.

The honest framing is that v3 is a calibration, not a new constitutive model. It does not change any law; it puts the catalog's numbers into the right decade for the world size and then pays the CFL bill that implies.

What is still wrong

• The multipliers are per family, not per material. Twelve is one number covering every structural solid; the right answer is a re-authored catalog in physical units, which nobody has done.
• It costs up to 4× the substeps. That is the known trade, and it is charged to every MPM material in the scene, not only the stiffened ones.
• Real-Time Pacing is a separate, still-off toggle. The legacy loop runs exactly one fixed 1/120 s step per rendered frame, so at 60 Hz vsync the entire simulation runs at half real-time speed and sim speed depends on frame rate — a large part of the same floaty feel that v3 addresses from the other side. If a scene still feels weak with v3 on, check the pacing toggle before touching the multipliers.
• ×12 changes look, not just strength. Any scene tuned against v1 or v2 will read differently: stiffer bounces, less settle, different pile angles.
• The technique guide's glass section is now partly stale. "Why glass still collapses too easily" describes a v1/v2 world; glass is one of the materials the user specifically confirmed as fixed by v3.

Why Sedov is the better model

Two structural wins. First, single ownership of energy: under Sedov the legacy per-frame emitter is skipped entirely, so exactly one thing owns strength, heat, smoke and divergence, and a follow-on pass extended that discipline to vents (the trapped pressure inside a vented cluster is deducted so the legacy cavity-vent branch cannot release the same energy twice). Second, correct time semantics: all front-crossing deposits are dt-free impulses consumed by exactly one projection per step, which is what makes the behaviour independent of substep count.

The follow-on pass also made the drawn ring honest: it now follows the physical front radius instead of a fabricated integration, and a retired front hands its record back to the legacy integrator so it fades from its last radius with no pop. Casing particles are pre-seeded into the guard buffer at rupture, so the front no longer double-kicks the shell that the legacy burst pushes already carry.

What is still wrong

• It is not the default, because live-fire calibration was never done. Every number in it — ξ, c₀, overpressure, impulse, thickness — is a live knob with an unvalidated starting value.
• Vent fronts and backdraft flash fronts are both default-off even under Sedov, so out of the box a dug-wall rupture still uses the legacy path.
• Mirror and reflection fronts were deliberately deferred: the image-source origin lands inside the reflecting wall, so it would need a change to the front packing to paint anything. Reflections you see are the aftermath through the projection, not reflected fronts.
• Two real defects were found by review after shipping, both fixed, both worth knowing as failure modes of this design: the ring froze at radius 0 in single-substep pacing because the radius query shared its zero sentinel between "absent" and "alive at t = 0"; and backdraft fronts replayed frozen bins after the backdraft feature was turned off.
• The cosmetic emitters still fire at rupture (shockwave, scorch, crack ring) independently of the model, so there remain two sources of visual truth.

Why v2 is good enough

It replaces a comparison with an average, which is the entire content of the fix. Once the cell holds ∑w·mC and ∑w·mC·T instead of a maximum, three separate symptoms stop being symptoms: contributions add, so a crowd is hotter than an individual; the mean can be below a participant, so hot things cool; and water having twice the thermal mass makes quenching fall out of the same arithmetic instead of needing its own quench channel. Impact heating is correctly left as a numerator-only deposit, because kinetic-to-thermal really is generated energy and should not be conserved against the mixture.

It cost two dedicated passes rather than an edit to the existing shaders, because both g2p and the SPH force shader sit exactly at the driver's SSBO-block limit and adding two buffers fails to link at all. That is the same wall the fracture and coupling work hit, and it is worth internalising: in this codebase, new shared state usually means a new small pass.

Measured cost: no measurable frame-time change over 15 s forced on, about 7 ms/frame.

What is still wrong

• The SDF-wall and air-injection heat paths are still one-way, explicitly deferred. A hot wall still heats a particle without cooling; that half of the audit finding is open.
• The relaxation rate is a single global constant (0.45/s) that stands in for every real contact conductance. Two materials with wildly different conductivity exchange at the same rate.
• Conservation is aggregate, not exact. "Roughly equal rates keep the exchange near-conserving" is the actual claim; nothing enforces a balance per pair, and the fixed-point accumulators quantize.
• The deposit is once per frame, the relax is its own pass, so the effective exchange rate has a mild dependence on how the frame is subdivided.
• The sound-field SDF diffusion is untouched and remains a separate thermal path with its own semantics.

Why v2 is good enough

Because the floor was not a tuning parameter, it was a lower bound on the state variable, and no other subsystem could beat it. Nothing about a heat-conduction term, a water quench, an air-cooling term or an oxygen gate matters if the last thing that runs sets the temperature back to 505 K. The replacement — a smoothstep(330, 380) sustain multiplier on the self-heating — keeps the property the floor was there to provide (a burning particle resists ordinary grid cooling) while restoring the property it destroyed (enough cooling wins).

The other four fixes are each small and each removes a distinct reason the front smeared. The one worth calling out is jitter: a deterministic threshold on a regular particle lattice is a synchroniser, and any spatial randomness at all breaks the lockstep. Water Quench is the complementary half — the legacy SPH-to-MPM heat exchange was strictly hot-to-cold via atomicMax, so water had no channel through which to cool anything, which is the same defect the contact-heat family fixed properly a fortnight later.

What is still wrong

• Water Quench is now partly redundant. Contact Heat v2 makes cold water quench natively through the mixture mean. Two mechanisms now cool with water, both default-on, and nothing reconciles them.
• The sustain thresholds are global constants (330 to 380 K for burning material, 320 to 370 for embers), not per-material. A material that should smoulder at low temperature and one that should not are gated identically.
• Burn-through speed is a look knob, not a fuel property. The 2.2× default multiplies every material's burn rate; per-material burn duration remains hand-authored.
• Ignition jitter is a hash of the particle index, so it is spatially uncorrelated. Real flame fronts are rough at a length scale; this is rough at the particle scale.
• It interacts with the later fire oxygen chemistry and structural fire damage, neither of which was re-tuned against v2's thinner front.

Why v2 is the right model

Restitution above 1 is not a tuning choice, it is a sign error in disguise: any resting contact with k > 2 is a pump. Clamping k to 1 + e with e ≤ 1 makes the response passive by construction, which is a property rather than a setting — you cannot tune your way back into the bug. Making recovery a rate in 1/s rather than a fraction of dx per frame is the same kind of change: it removes dx and frame rate from a physical quantity's units.

The back-reaction is the part that makes it a coupling rather than a boundary condition. Recording the impulse at the one site where the contact force is actually known, rather than recomputing it in the scatter pass, is a deliberate anti-drift decision worth copying.

The verification is the most instructive part of this entry. Forcing it on with a temporary GPU readback caught a bug that a clean build and a no-crash run would both have missed: the consume-and-zero sat after grid_op's empty-cell early-out, but SPH deposits with B-spline weights and therefore reaches nodes carrying no MPM mass. Those cells were never cleared and the fixed-point accumulator climbed monotonically, 0.6 to 126 over 25 s — int32 wraparound in roughly seven minutes, and a phantom kick of minutes-worth of stale impulse for anything that later drifted in. Moving the read and clear above the early-out made it flat: about 140 cells at about 0.18 peak impulse, steady over 30 s.

What is still wrong

• It is not the default, because the live A/B is the calibration step and it has not been run: does a water jet now visibly move a sand pile, and does fluid stop climbing out of pockets?
• One frame of latency on the back-reaction, by construction. Acceptable in an explicit scheme, but it means action and reaction are not simultaneous.
• The applied velocity change is clamped to one cell per step so a near-empty node cannot produce an unbounded kick. That is a safety clamp, i.e. exactly the kind of compensator this page is generally suspicious of.
• It is still a penalty contact, not a unified pressure solve. The relationship map's own wording remains accurate: soft coupling for contact and thermal exchange. Two-way momentum makes it honest, not unified.
• SDF geometry still receives no momentum back from either solver. That one-way arrow is a separate, still-open item.