FLUID_BULK_IMPLICIT.md

Implicit coarse-liquid transport — opt-in checkpoint

Full-goal dependency: flowing Eulerian ownership needs transport that can empty a closing cell, not just pressure coverage. The current explicit donor cap leaves at least 5% behind. This work replaces that transport update in a separate opt-in path; it does not remove particles or declare Narrow Band FLIP complete.

Integration map

• FluidBulk keeps source admission, the resident/pending ledger, canonical FP64 pressure-flux restriction and the existing force update.
• FluidImplicitTransport consumes those resident quantities, shared face rates and actual end-of-substep cut-cell capacities. It solves a backward-Euler upwind finite-volume system for volume and all three momentum components.
• Persistent GPU buffers, root descriptors, compute PSOs, UAV barriers, predicated iteration, existing-fence readback and PIX events follow the engine's current pressure/phase-limiter abstractions.
• The same shared transfers feed the existing independent bulk balance audit; a separate immutable snapshot reconstructs the implicit matrix, checks its residual, output, conservation, positivity and kinetic-energy dissipation.
• FluidBulkPressure, MAC projection, positional density correction and the particle-owned DXR surface remain connected as before. The new inventory is not added to particle mass or rendered as a second fluid.
• In this mode only, a partially filled coarse cell that disappears during the substep retains pressure support until its liquid can escape. Ordinary open fractional tails are still particle-defined. Capillary support and missing face weights use the actual starting fraction, not an invented full cell. The room placement regression exposed a zero transport row with nonzero resident mass when that pressure escape path was absent.

For oriented integrated volume flow Q = dt * canonicalRate, the unknown c_i has volume fraction in W and momentum/open-volume in XYZ:

(Cnew_i + sum(outgoing Q)) c_i - sum(incoming Q * c_neighbor) = qold_i
qnew_i = Cnew_i * c_i

Summing the equations cancels shared face transfers. A closing cell may have zero final capacity but positive outflow, so it remains a transport unknown and can pass its mass onward. Global mass/momentum error must match solver residual and FP32 storage error, not an invented source correction. A singular/trapped system or an exhausted solve must be reported, never silently discard mass.

This implicit update is monotone for nonnegative source mass when its transport matrix is invertible. An upper volume-fraction bound additionally needs compatible pressure/free-surface support and geometric conservation; unconditional positivity must not be mislabeled unconditional bounded VOF. First-order diffusion, partial-interface support, conservative moving particle/grid exchange, sparse physical allocation and the full adaptive specification remain required.

The implementation is original. Geometric conservation and particle/grid coupling are separate requirements; the Narrow Band FLIP authors specifically identify instability in naive coupling. This transport solver is not their full coupling scheme. FLUID_BULK_PRESSURE.md retains the previous checkpoint's unresolved temporal-preservation results.

Next numerical dependency

For this particular matrix, a sufficient upper-bound condition is A * 1 >= qold.w component-wise: the monotone inverse then implies c.w <= 1. Geometric conservation makes A * 1 equal the previous open capacity when the carrier flux is incompressible throughout the transported support. Partial particle/bulk interface support does not yet establish that condition here. Increasing the linear iteration count cannot repair this missing compatibility.

The OpenFOAM authors' semi-implicit MULES description is a useful next-stage reference: it limits a higher-order flux correction after a bounded implicit upwind predictor. That is not permission to apply our old explicit donor/receiver limiter directly to implicit transfers, or to claim boundedness when this engine's predictor is already overfilled. No OpenFOAM code is imported. A compatible bounded base state remains the prerequisite for that kind of accuracy correction and for conservative particle/grid handoff.

Validation and limits (RTX 5090)

bulk-implicit-validation.json records current executable/shader hashes, all 13 deterministic GPU scenarios, five same-build explicit-pressure references and three interleaved raw timing pairs.

• All implicit cases pass every-substep convergence/positivity, independent final-substep matrix/balance/energy checks, canonical flux restriction, pressure convergence, source admission and finite optical-input checks.
• The wake exercises 768 closing-water updates; the room placement/reset test exercises 17. Both finish with zero inventory in fully closed cells. The reference wake retains 0.000047035 m³ there.
• The five comparisons preserve particle mass exactly, stay within the existing density gate and change reconstructed surface volume by at most 0.19631%. These are physical endpoint comparisons, not a temporal image-quality proof.
• 174 Node tests and all 7 native CTests pass. The Release build and changed compute kernels compile. Three module-relative launch tests and the normal Play Lab.cmd resize/minimize/restore/shutdown test pass (189 interactive frames).
• Seven cases still exceed local endpoint capacity; e.g. the 120-frame room has 0.00254093 m³ total excess versus 2.99478e-8 m³ in the explicit-pressure reference. Upper-bound acceptance fails. No clipping hides that difference.

At 1920×1080 output, Balanced RR and frame generation off, medians across three 300-frame runs change from 22.8818 to 23.5105 ms raw and from 11.4634 to 11.8489 ms fluid time. This is a correctness prerequisite with added cost, not an optimization or default-promotion result. The implicit helper averages 0.919246 ms; the reference phase-limiter helper is not the entire explicit transport update. The first timing pair was repeated with the capture collector idle; no build/validation work overlaps the retained timing runs.

Reproduce with test-bulk-projected.ps1 -Mode implicit; obtain matched references using -Mode bounded -PressureSupport and case filter ^(calm|room|room-no-tension|wake|adaptive)$. Run profile-bulk-implicit.ps1 serially with no other GPU workload, then use record-bulk-implicit.mjs RUNTIME --profiles to validate and collect evidence. Scripts close only their owned test processes. No normal launcher, mainline native engine or portable release is changed. GPU validation is not claimed: the earlier Graphics Tools probe failed with 0x887A002D. The full adaptive-system goal remains incomplete.