336 Commits

Include renames

Author	SHA1	Message	Date
Omar SobhandClaude Fable 5.1	134ac03870	rtx-cfd: OverlapMap::region_force — momentum flux into a background region (control-volume face formula, one-sided next to holes); overset_cfd1 prints the four momentum routes (CV box, ring outer, hole boundary, wall) and their defects at the settled state ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-06 09:08:01 -07:00
Omar SobhandClaude Fable 5.1	5f780447de	rtx-cfd: PatchConvection::TvdVanAlbada — van Albada deferred correction on the curvilinear predictor (downwind-side linear weight, gradient-ratio r over the face d lengths, far-upwind across the opposite face, boundary faces upwind); annulus MMS orders 2.10/1.69 at 0.24× upwind; cylinder-flag MMS orders 1.98/1.97 (1.06× upwind — diffusion-dominated, recorded); knobs RTX_OVERSET_CFD1_TVD, RTX_OVERSET_MAX_ROUNDS, RTX_CF_SCHEME ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-06 00:21:48 -07:00
Omar SobhandClaude Fable 5.1	02c855b9c4	rtx-cfd: overset Schwarz stall rule guarded (fires only within 10× the tolerance) — unguarded it cut CFD1's transient at round 3 every step and the coupled march diverged at ny=62 (pressure 4e4 → 1e140 by step 450); CFD1 harness knobs (stall, step cap, trace) ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-05 21:01:57 -07:00
Omar SobhandClaude Fable 5.1	00f73eb70a	rtx-cfd: overset CFD1 harness — the steady-march stall rule (stall_rounds 2) actually enabled (the previous commit's edit had missed its anchor) ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-05 20:28:25 -07:00
Omar SobhandClaude Fable 5.1	9761cf2319	rtx-cfd: overset P4 — CFD1 on the composite (tests/overset_cfd1.rs): ny=41 wall drag 15.2156 (+6.46%), lift 1.0879 (−2.78%), CV drag 15.528 (+8.64%), routes 2.0% apart (staircase +10%); 28550 steps, 2099 s at dt 4.57e-4 with 5.0 Schwarz rounds mean (stall rule off) — steady-march stall rule on for the next rungs ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-05 20:27:52 -07:00
Omar SobhandClaude Fable 5.1	45ff34da8f	rtx-cfd: overset P4-0 — the O-grid around the Turek–Hron rigid body (cylinder + flag), gated ... patch_gen::{cylinder_flag_outline, cylinder_flag_patch, winslow_smooth, respace_rays} (+ convex_hull, offset_convex_polygon, nearest_on_polyline): the outline CCW then reversed to clockwise (tip semicircle 16 cells, junction fillets of a FIXED radius with 3 cells, straights graded 0.3 h → h, cylinder arc at h); the outer ring the 6 h normal offset of the body's convex hull; initial pairing by the inner point's normal offset projected onto the hull offset (an arclength-proportional pairing folded the transfinite grid at the tip: rays crossed where the curvatures differ); Winslow (TTM) smoothing of the interior with the outer nodes SLIDING along the hull offset (each re-placed at the nearest point to the extrapolated ray), then re-spacing along the smoothed rays to the across stretch. Gates (tests/patch_cylinder_flag.rs): ny = 41/62/82 → 143×12 / 183×12 / 225×12 cells, positive, wall row 0.23 h (fillet max 0.37 / 0.43 / 0.50 h), worst non-orthogonality 76.6 / 69.1 / 63.3° at the concave fillets (structural: a concave arc's normals converge at its centre), classification of the benchmark background with both donor invariants. P0 MMS on these meshes (tests/cylinder_flag_mms.rs, exact acceptors, line-implicit): Stokes orders 2.17 / 2.13, upwind 2.00 / 1.86 (cell Péclet ≈ 0.1), divergence ≤ 9e-14 — the fillet skew costs nothing measurable. Rule: a refinement ladder's geometry must be fixed in physical units — with fillet = h/2 the Stokes orders read 1.74 → 1.31, the O(h) boundary perturbation masquerading as a scheme defect; the fillet is a parameter (5 mm across the ladder). Also: the P3b knock-outs H1/H2 on the balanced default (no effect), the P3 §5.10 record in the falsifier's header. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-05 19:50:14 -07:00
Omar SobhandClaude Fable 5.1	62df6bd628	rtx-cfd: overset P3b — the reclassification impulse located (the fringe ring is a staircase of the interpolated velocities' mass defect) and removed by a converged fringe flux balance (default on): falsifier max spike 594 → 5.50 N/m at the FSI2 step (staircase 6490), 10.95 / 16.79 at dt/2 / dt/4 (12600 / 25600), rms spike 0.07% of the force, far probe 6 (7900), KE per event 4.9e-3 J/m falling with Δt (2.6 fixed) ... OverlapMap::balance_fringe_fluxes: Gauss–Seidel through the prescribed faces of every fringe cell to 1e-12 of the prescribed flux scale (≤ 50 sweeps), after every fringe stamping (3 fixed sweeps 101 N/m, 10 sweeps 5.5 — converged is the rule). OversetParameters: fringe_flux_balance (default on, RTX_OVERSET_NO_BALANCE off), fringe_balance_tolerance, refill_turned_active (measured no effect: 593.7 → 593.8; kept as the record), stall_rounds opt-in. P3b locating trace RTX_OVERSET_TRACE_SP (continuity source by class change in cell volumes/step, stored-pressure jump of turned-active cells): the flipped cells' mass source ≤ 6e-3 cell volumes/step, their stored pressure 5–10% of the range off their neighbours (4.4% on the static MMS — the meshes' discretization disagreement). Knock-outs refuted (RTX_OVERSET_H1 keep own face velocities, H4 no warm start, pressure refill): 593–597 N/m each. S4 MMS with the balance: velocity errors within 0.1% of the pinned values, the background's overlap mass defect 1e-13 by construction, pressure errors unchanged. overset_mms prints pressure diagnostics; overset_falsifier records the balanced ladder (regression guard 20 N/m at dt; RTX_OVERSET_FALSIFIER_STRICT asserts the registered gates — (ii) holds at dt, misses at dt/2, dt/4; (iv) fails: residual ∝ 1/Δt^0.8). Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-05 19:16:16 -07:00
Omar SobhandClaude Fable 5.1	6b8837301f	rtx-cfd: overset Schwarz stall detection is opt-in (stall_rounds, default 0): right for a steady march (n=64: 2.26 rounds mean, no cap hits, L2 identical), wrong for a transient (falsifier max spike 594 → 4044 N/m when on); the MMS harness sets 2 ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-05 18:35:06 -07:00
Omar SobhandClaude Fable 5.1	e2edff9b1d	rtx-cfd: overset A-P3 — the falsifier plate on the overset (FAILS the registered gates by one order less than the staircase); wall force; composite pressure-level pin; Schwarz stall detection ... patch_gen::{stadium, graded_fractions}: the falsifier plate as a stadium O-grid (semicircular ends r = half-thickness; 16 cells per end arc, straights graded 0.30 h -> h at 1.15, offset 6 h, 12 rows stretched 4x; 148x12 cells, every ray a normal, worst non-orthogonality 4 deg). CurvilinearPisoSolver::surface_force (+ PatchLoad): F = sum(-p_f S_f + mu (grad u + grad u^T)_f . S_f) on the wall faces with the wall cell's LSQ gradients (wall Dirichlet in the velocity fit); HELD on the phantom circle against the exact stress integral: 1.3e-2 / 6.3e-3 / 3.6e-3 at n = 32/64/128 (orders 1.05 / 0.81), 22x the staircase's accuracy. OversetPisoSolver: the composite p' level pinned to zero mean over the active cells every round (the coupled problem is pure Neumann; the temporal warm start handed each step's level to the next — background pressure 1e7 growing 5e4 per step on the falsifier; an unpinned level also inflated the relative Schwarz stop); stall detection (no progress over three rounds = the inner solvers' noise floor; 6560 of 150k steps burned the 20-round cap at n = 64, a 7.5 h n = 128 march); schwarz_stalled in the result. tests/overset_falsifier.rs (records; RTX_OVERSET_FALSIFIER_STRICT asserts the registered gates, _LADDER runs dt/2 and dt/4, _TRACE the top-12 spike steps): max spike 594 / 981 / 1720 N/m at dt / dt/2 / dt/4 (staircase 6490 / 12600 / 25600), rms spike 61-89 (810), far probe 502-1509 (7900), KE injection 0.16-0.21 J/m per event on the common cell set (2.6) — every large spike a ~104-cell full-row reclassification; exponent -0.77 (-1.0). The registered 5% gate (8.75 N/m) is missed 68x: the overset's own reclassification impulse is the finding (omni-cortex overset_metal_campaign.md §5.10); P3b = locate per cell, then the fringe flux balance. tests/patch_stadium.rs, curvilinear_loads.rs, overset_common::plate_patch. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-05 04:32:43 -07:00
Omar SobhandClaude Fable 5.1	afd1bff6ee	rtx-cfd: overset A-P2 — the patch overlaps the background (OversetPisoSolver), gated S1–S5 ... Background = the embedded solver with a mask from the overlap classification (embedded/{mod,projection}.rs: module split, projection's solve/apply halves, set_overlap, fringe p' Dirichlet by elimination into extra_diag/rhs, anchor dropped, set_inner_stop_factor, phase API begin_step/solve_correction/ apply_correction/end_step; advance rebuilt on the phases — every suite digit- identical, FSI2 default line-for-line). Patch = the curvilinear solver with an acceptor ring (set_side_velocity; set_acceptor_ring/stamp_acceptors/ set_acceptor_correction; acceptor Dirichlet by elimination into PressureSystem.links so the BiCGSTAB stop stays in flux units — identity rows measured unconverged at 2431 iterations; same phase API). overset/overlap.rs: OverlapMap — hole/fringe/active from the patch's own indices (hole = body or k <= nn-1-overlap_rows, DEFAULT_OVERLAP_ROWS = 4 from the 2.9 h depth budget), dual-quad inverse-bilinear donors patch→fringe, lattice donors →acceptors, both invariants asserted, mass-defect measures. overset/mod.rs: OversetPisoSolver — advance (exchange rebuilt BEFORE the predictors from the previous corrected field), alternating Schwarz on the acceptor p' vector with Anderson(3) (plain Schwarz measured 0.82/round: floating patch, Neumann wall) and the previous step's vector as warm start (1 round/corrector at steady state), stop relative to the STEP's p' scale (the MG absolute stop is 1e-9/dt² in pressure — the whole second correction), set_patch_mesh, snapshot/restore carrying the warm-start vector. Gates: overlap linear-exact 1e-13, quadratic orders 1.96/1.99 (acceptors), 1.40/1.91 (fringe); half-couplings: patch with exact acceptors Stokes 2.07/1.98 + 2.08/1.98, upwind 0.84/0.84, background with exact fringe 7.86e-3/2.90e-3/ 1.09e-3 (1.44/1.41); two-mesh MMS n=32/64: background 8.717e-3/4.207e-3 (1.03x/ 0.97x the embedded circle), patch 1.322e-2/6.904e-3 (1.5-1.6x), orders 1.05/ 0.94, patch div <= 5e-13, overlap mass defect 3.6e-3 -> 8.2e-4 of the overlap flux (under the registered 1e-3 from n=64; disclosed at 32); motion: stationary patch through set_patch_mesh bit-identical, snapshot/restore with a pending mesh bit-identical, translating phantom circle 1.22x/1.19x the static level over 4.5 cells. Inherited, disclosed: poisson_equivalence's no-body multigrid pin fails by 3.9e-9 at d46fb0b (M1's commit; verified in a clean worktree). Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-04 19:24:13 -07:00
Omar SobhandClaude Fable 5.1	d46fb0b7a7	rtx-cfd: overset A-P1 GATED — the curvilinear patch moves and deforms under an exact 2-D DGCL ... StepGeometry (motion.rs): time-averaged face vectors S̄_f = ½(S^n + S^{n+1}) and swept volumes δV_f = S̄_f·δc_f — exact for linear node motion on any quad, so Σ sign δV_f = V^{n+1} − V^n is algebra (1.8e-14 measured; the EndOfStep control 2.1e-3). CurvilinearPisoSolver::set_mesh(next) names the end-of-step geometry; advance swaps it in, rebuilds operators + pressure matrix on it (L_f, LSQ gradients, no mesh-velocity term in the projection), keeps the old mesh for V^n and the explicit boundary data; predictor in the conservative ALE form V^{n+1} û = V^n u^n + dt(−Σ sign (F − δV/dt) u_f + ν D + f V^n), written as u·(V^n/V^{n+1}) + … so a stationary mesh is bitwise the static path; fluxes on S̄_f; snapshot carries both meshes; swept_face_rule knob (Trapezoidal default, EndOfStep = negative control). Stokes limit keeps the mesh flux (was dropped with convection) and centres it (upwinding it cost an order: 1.06/1.00). Gates (tests/curvilinear_ale.rs, 11 tests, 83 s): uniform flow on a wiggling AND bending annulus 4.44e-15 over 400 steps, p exactly 0, 0 pressure iterations; control deviates 4.9e-5; stationary mesh through the moving path bit-identical (both diffusion variants); snapshot/restore on the moving mesh bit-identical; Taylor–Green orders unchanged — upwind 0.995/0.976 vs fixed 0.987/0.975 at 1.05× error, Stokes 1.92/1.97 vs 1.94/1.99 at 2.4×, moving annulus 2.11/2.02; linear-field falsifier 1.95/1.92 (annulus), 1.91/1.43 (square, sliding wall nodes). P0 ladders re-run identical to every digit. Rule from the diagnosis: start a moving run ON the t = 0 mesh and sweep less than a cell per step — a first step that jumped 2–4 cells imprinted an O(displacement) error no refinement removed (dt- and motion-independent). Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-04 17:00:04 -07:00
Omar SobhandClaude Fable 5.1	c63d79c300	rtx-cfd/rtx-fsi: overset A-P0 GATED + M1 precision probe — curvilinear collocated PISO: relative-reduction pressure stop (the absolute stop floored |du/dt| at 2e-4 on 64²), line-implicit-n sign fix, adjustPhi; gates: Cartesian reduction 1.37–1.40x the staggered error at orders 0.83/0.90; skewed stretched periodic annulus Stokes orders 2.30/2.06 (explicit and line-implicit), upwind 1.08/0.80; Poiseuille exact to 1e-9 on Cartesian and affine-sheared periodic channels (both diffusion variants), varying-skew channel order 2.02 (v 1.9), cell mass 1e-14; divergence ≤ 1e-11 relative every step; snapshot/restore bit-identical. M1: poisson.rs multigrid hierarchy generic over MgScalar (f32/f64), f64 CG keeps its own fine level; MgPrecision on MultigridParameters/EmbeddedParameters/PisoParameters, set_poisson_precision, harness RTX_FSI2_POISSON_F32 (march + noise probe, printed marker); f64 arm bit-identical in vivo (FSI2 default line-for-line with 08-31), f32 arm holds the noise floor and stall pins and the FSI2 band; poisson_equivalence f32 arm ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-04 12:40:43 -07:00
Omar SobhandClaude Fable 5.1	52da75a3a9	rtx-cfd: curvilinear collocated PISO on a structured patch (overset A-P0, WIP) — PatchMesh (right-handed s,n; periodic seam with shift; face metrics), patch generators (TFI, skewed annulus, sheared/varying-skew channels), CSR + Jacobi-BiCGSTAB, the Zang–Street–Koseff incremental step with the node-based 9-point L_f, LSQ gradients, explicit and line-implicit-n predictors, adjustPhi; tests: mesh metrics (5 green), operators exact on linear fields incl. the seam (green), sparse (2 green), MMS ladder (Cartesian 16/32: 1.37–1.39x the staggered error, order 0.83; n=64 stalls at a |du/dt| floor 2e-4 — open, tolerance-scaling hypothesis), annulus/Poiseuille not yet run ... Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-04 05:00:08 -07:00
Omar SobhandClaude Fable 5.1	1347bc6772	test(rtx-cfd): fresh-cell falsifier — circle body, far-field probe, kinetic energy per step, speed knob, and the per-flip kinetic-energy split by face class (interior / fresh / dead) ... The split locates the flip's velocity change: 71% (plate) / 61% (circle) on the DEAD faces — fluid faces the wall reaches, overwritten with the wall-side reconstruction in one step (a ~0.25 U_b jump per face) — 27–38% in the interior response, ~1% on fresh faces. Five treatments leave the per-flip impulse/energy unchanged on this falsifier (all measured, none kept unless already committed): substeps, the swept-volume source (40x worse), the fresh-face field extension, face apertures + wall-relative divergence in the projection (-12%, and it breaks the static invariants), and imposing the t^{n+1} wall before the predictor. The impulse is intrinsic to a binary staircase wall that stands still between flips and jumps a cell; the record is in omni-cortex docs/fresh_cell_gcl_campaign.md. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-03 11:56:04 -07:00
Omar SobhandClaude Fable 5.1	c9492d3e1e	test(rtx-cfd): fresh-cell falsifier extended (circle body, far-field probe, kinetic energy, speed knob) + print-only divergence trace; two candidate fixes REFUTED on it (swept-volume source 40x worse at either sign; fresh-face field extension no effect), both kept default-off with their verdicts ... Phase 1 of omni-cortex docs/fresh_cell_gcl_campaign.md. The mechanism of the moving-body force spikes is measured from four directions: per-flip force amplitude ∝ 1/dt, kinetic energy injected per flipped cell 0.048 J/m independent of dt and body shape (plate row vs circle), felt at a far-field pressure probe, and ∝ U^2 (2.60 / 0.64 / 0.15 J/m at U = 1 / 0.5 / 0.25). A binary mask's wall position jumps by one cell at every flip and the fluid answers with a fixed impulse. Neither the swept-volume source (the wall faces already carry the swept volume — the source double-counts it) nor the fresh-face velocity is where it lives. Next: the virtual cut cell in the projection (apertures + the wall-relative divergence), registered in the campaign doc. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-03 11:40:28 -07:00
Omar SobhandClaude Fable 5.1	23eb996a9f	test(rtx-cfd): the fresh-cell falsifier — an oscillating flag-sized plate on the FSI2 grid and dt; force spikes track fresh cells (90% at dt) and their per-event amplitude scales as dt^-1.0 (rms dt^-0.44) ... Phase 0 of omni-cortex docs/fresh_cell_gcl_campaign.md (Seo & Mittal's oscillating-body test on this embedded solver). Plate 0.35 x 0.02 m, 1 m/s peak, 80 mm amplitude, still fluid, h = 1/152, dt = 3.24e-4 and its halvings; force sampled as the coupling samples it. Measured: at rest exactly zero; moving, 648 fresh cells per sweep at every dt, rms spike 810 / 1080 / 1490 N/m and MAX spike 6.5e3 / 1.26e4 / 2.56e4 at dt, dt/2, dt/4 — the per-event impulse doubles with each halving (a fixed whole-cell volume error per fresh cell delivered in one step, the raw (dV/dt)|1 - CFL_b| source), against a physical added-mass force of ~1.3e3. RTX_FRESHCELL_LADDER=1 runs the ladder, RTX_FRESHCELL_CSV=<dir> dumps per-step records. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-03 11:24:24 -07:00
Omar SobhandClaude Fable 5.1	19d8307ae4	test(rtx-fsi): closure schedule by tip speed (CRESCUE_SPEED=<f>), default off ... Route 1 of omni-cortex docs/coupling_rescue_campaign.md §12, registered in docs/closure_scheduling_campaign.md: whenever the last committed tip jump exceeds f x its trailing-2000 peak, open a coarse episode (the s=2 interpolated closure, CRESCUE_COARSE=<M> steps, re-evaluated at each end) BEFORE the step — preventive, because the s=1 closure incubates the crossing instability inside the healthy envelope and no signature-triggered rescue acts early enough (rungs A, C, A' refuted). Speed episodes are expected twice per period: not counted against the runaway-episode cap; the coarse fraction is printed at the end of the march. Verified: FSI2 committed default digit-identical knob-off vs the same-day baseline; fmt + clippy clean. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-03 09:57:40 -07:00
Omar SobhandClaude Fable 5.1	b66f73aeef	test(rtx-fsi): coupling rescue rung A' — increment trigger (CRESCUE_INC=<K>), default off ... Trigger (c) of omni-cortex docs/coupling_rescue_campaign.md §11: a predictor increment above K x its trailing-2000 median (non-rescued steps; median refreshed every 10 steps) opens a coarse episode (rung C) BEFORE the step's first pass — no pass is spent on a step the episode replaces. Calibrated on the INCTRACE dumps: the anchor u=1.00 r=1 stays under 1.72x for its whole march (K = 2..8 never fire); the (1.27, 2.0) death crosses 3x at 145 steps before its panic and never during resonant growth (the tip-jump trigger fires at 54). Registered K = 3, fallback 4. Needs CRESCUE=1 and CRESCUE_COARSE=<M>. Verified: FSI2 committed default digit-identical knob-off vs the same-day baseline; fmt + clippy clean on the touched files. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-02 23:01:12 -07:00
Omar SobhandClaude Fable 5.1	7e0f159097	test(rtx-fsi): coupling rescue rung C (CRESCUE_COARSE, burst-local s=2 coarsening) + INCTRACE per-step increment dump — both default off ... Coupling-rescue campaign, continued (omni-cortex docs/coupling_rescue_campaign.md §11). Rung A was refuted 4/4 by mechanism (the substeps reproduce the rejected motion — the runaway is in the converged coupled load at the crossing). Diagnostics on the same death: SUBCYCLE=2 marches GREEN to t=16 (zero bursts), HYST=0.25 dies EARLIER. So: - RTX_{prefix}_CRESCUE_COARSE=<M> (with CRESCUE=1): on a trigger, reject the step and take 2dt coupled steps with the fluid subcycled at 2x (fluid dt unchanged = the s=2 interpolated closure) for M coupled steps, then resume; episodes counted, cap 5 per second of march (loud). rescue.rs: coarse_step / attempt_with generalisation; march loop is now a while loop (a coarse step consumes two indices, the series carries a linear midpoint). VERDICT: refuted 2/2 — the coarse steps themselves cannot close once the state is 10x wild; both rungs act too late (the kinematic trigger is the limitation). - RTX_{prefix}_INCTRACE=<csv>: reporting-only per-step dump (step, t, predictor increment, tol_step, passes, residual, stalled, tip jump) — rung A''s calibration data (healthy anchor vs death). Verified: the FSI2 committed default digit-identical knob-off after each change (same-day baseline); fmt + clippy clean on the touched files. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-02 22:09:49 -07:00
Omar SobhandClaude Fable 5.1	e76271ac67	test(rtx-fsi): coupling-level rescue rung A (RTX_FSI{2,3}_CRESCUE, default off) + TRACE_FROM autopsy window ... Coupling-rescue campaign (omni-cortex docs/coupling_rescue_campaign.md). Every recorded FSI2 s=1 coupling death is the END of a multi-step runaway (74-march burst scan: green marches never exceed 2.4x their own p95 tip jump; 27/28 deaths burst 5-37 steps first). The traced autopsy of the (1.27, 2.0) death (replay digit-identical, CSV byte-identical) shows a growing period-2 instability of the CONVERGED coupled scheme at the tip's max-velocity crossing: increment x33 and converged load 1e4 -> 1e6 N over 200 steps with the coupling converging on 111 of the first 112 steps. - march.rs: RTX_{prefix}_TRACE_FROM (print-only autopsy window: per-pass residual/load + one line per step with increment, tol_step, retry_at, acceptable, outcome, committed tip jump); RTX_{prefix}_CRESCUE (default off): on a fatal stall or a committed tip jump > 3x the running p95 (trailing 2000 non-rescued steps), reject the step and repeat the interval as 2/4/8/16/32 coupled substeps of dt/n (Mayr-Wall-Gee reduced-step repetition, five repetitions); per-rescue record printed, MarchResult.coupling_rescues/_failures/rescue_records, rate cap 20 rescues per second of march (dies loudly). - rescue.rs (new): the substep ladder — each substep a complete coupled step at dt/n with its own predictor, fresh coupler, C1 velocity chaining, tolerances at the substep's increment; the march's own step path is NOT routed through it (digit identity by construction). - rtx-fea NonlinearDynamicStepper::step_with_dt (step at an explicit dt; step() delegates float-for-float); Fsi2Harness::advance_subcycled_with (explicit fluid dt; advance_subcycled delegates). Verified: FSI2 + FSI3 committed defaults and the noise probe digit-identical knob-off vs same-day / 2026-08-31 baselines; knob ON on the anchor u=1.00 r=1: green, zero rescues, CSV byte-identical to TWIN-1's u1.00.csv; fmt + clippy clean on the touched files. Verdict of rung A on the provocation set: REFUTED 3/3 by mechanism — every rescued interval was carried but reproduced the rejected step's motion (dt/2..dt/32 give the same jump), so the runaway lives in the coupled LOAD at the crossing, not in the time integration; the SUBCYCLE=2 closure marches through the same crossing. The knob stays, default off, as the instrument that measured this. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-02 21:36:06 -07:00
Omar SobhandClaude Fable 5.1	0f578087ce	feat(rtx-interpret): drift-gated decoder renormalisation (SAEConfig::normalize_gate) ... `normalize_gate: Option<(lo, hi)>` — when set, a decoder column is rescaled to unit norm only if its norm has left the band; columns inside are left exactly as the gradient step made them (divisor 1.0). `None` keeps per-step renormalisation, bit for bit. Why (omni-cortex D629/D630): per-step rescaling was measured doing two opposite things on the same 32-unit SAE. With it off, two runs descended cleanly to floors 3-7x LOWER than with it on — it was fighting descent. Two other runs (lr 0.01, seeds 7 and 99) diverged outright without it — it was also the clamp holding an unstable rate finite, turning a blow-up into a slow oscillation that looked like a healthy dictionary drifting. The band keeps the second role and drops the first; D630 measures whether it does both. The cold-start exemption is applied after the gate, unchanged. Test pins: in-band columns bit-identical before/after, out-of-band pulled to unit, gate None == per-step. Fixture norms sit strictly off the band edge — a hand-scaled 2.0 came out 2.0000002 in f32 and was, correctly, treated as outside. Co-Authored-By: Claude Fable 5.1 <[email protected]>	2026-09-02 20:30:25 -07:00
Omar SobhandClaude Opus 5	9575b84803	style: clear the fmt gate and two lib clippy warnings ... Deferred deliberately while the TWIN-2B/2C campaign had live marches: each march is a fresh `cargo test` invocation, so reformatting `turek_hron_fsi2.rs` mid-campaign would have forced a test-binary rebuild and cost comparability for a cosmetic gate. The family closed, so this is now free. - `cargo fmt --all` across 8 files that had drifted (including the FSI2/FSI3 harnesses touched by the UMEAN/ES override commits). - `rtx-feature-store/tests/integration_tests.rs` had trailing whitespace rustfmt refused to format around ("left behind trailing whitespace" internal error), so the whole file was being skipped; stripped it and the file formats now. - Two `unnecessary_parentheses` warnings in the rtx-transformers lib (`continual/progressive.rs`, `curriculum/mod.rs`) — these were the only rustytorch warnings surfacing through omni-cortex's workspace clippy gate, which is how they were found. No behaviour change. rtx-fsi test binaries still build. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01B1feFAQxjbCRHePUdxuNra	2026-09-02 19:15:53 -07:00
Omar SobhandClaude Fable 5	045e145962	test(rtx-fsi): RTX_FSI{2,3}_ES stiffness override for the TWIN-2 parameter sweep ... The second knob of the TWIN-2 campaign (omni-cortex docs/twin2_stiffness_campaign.md): case_from_env now also reads RTX_{prefix}_ES over the case's benchmark Young's modulus, mirroring UMEAN exactly — digit-identical with the knob unset (verified in vivo on BOTH committed defaults against same-day baselines: FSI2 and FSI3 physics lines diff-clean, only compile/wall timings moved), a loud override line naming the benchmark value and E/E0 when set. The benchmark-inflow physics-band guard in both turek_hron tests widens to a benchmark-CASE guard (u_mean AND e_s bit-identical) so no measured band ever asserts off-benchmark; machinery invariants stay asserted at every inflow and stiffness. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-09-01 20:50:58 -07:00
Omar SobhandClaude Fable 5	c08abbdc4f	test(rtx-fsi): RTX_FSI{2,3}_UMEAN inflow override for the TWIN-1 parameter sweep ... case_from_env applies the knob over the case's benchmark inflow; unset, the same f64 flows and both committed defaults are verified digit-identical in vivo (physics lines diff-clean vs same-day baselines; only wall-split timing percentages moved). Physics bands in both tests are guarded to benchmark inflow — an off-benchmark march (the sweep) asserts machinery invariants only; its bands live in the TWIN-1 composition harness (omni-cortex docs/twin_composition_campaign.md). The override prints loudly so sweep logs are self-describing. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-31 21:25:30 -07:00
Omar SobhandClaude Fable 5	4aec4e589d	rtx-cfd/rtx-fsi: the FlowField->clawview exporter — real FSI fields through the viewer, end to end ... The march gains RTX_FSI{2,3}_FFLD (MarchConfig::ffld_dir, off by default): every snap_every committed steps (10 when SNAPEVERY is 0), dump the committed FlowField (FlowField::save, bit-exact), the solver's own fluid-cell mask and the interface polygon as text sidecars, plus an index.csv. Reporting-only after acceptance; the FSI2 default with the knob off reproduces every printed digit of the warm-start baseline (uy 3.4921 +- 3.5109, conservation 8.25e-12). The exporter (rtx-cfd examples/ffld_to_vtk): a dump directory -> clawview-readable legacy VTK — per-snapshot 2D triangle meshes (one selectable point scalar --field p|umag|vort, the 0/1 fluid mask as integer CELL_DATA) and an optional space-time volume (--spacetime: frames stacked along z = time, prisms split to tets, POINT_DATA phi), which clawview's slice-plane animation plays as a transient movie. Verified end to end with REAL fields, not synthetic: a 29-frame FSI3 release-transient dump (t 4.0 -> 4.1, 64 MB) exported to 29 snapshots + a 984,312-tet space-time volume; the clawview server loaded both (176,320 nodes volume; 23,877-node snapshots) and served live cross-sections (time slices), vorticity isosurfaces and contours from them. The check caught two real viewer-contract constraints now encoded in the exporter: clawview's legacy-VTK path parses CELL_DATA scalars as INTEGER markers only, and supports exactly ONE point scalar (all POINT_DATA blocks append into `phi`) — multi-field snapshots need the .clwv route, out of scope here. Closes the top open thread of the fifteenth-session handoff (the FSNP/FlowField->viewer exporter); the mesh-repo push and the claw-gds cargo feature remain clawview-side items. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-30 14:45:54 -05:00
Omar SobhandClaude Fable 5	4d2cede7bc	rtx-cfd: warm-start the first corrector's pressure-correction solve — measured 2.5x fewer PCG iterations where it counts ... The MG smoother (neighbour_sum + PCG, 34-40% of the fluid) was the remaining fluid cost. Measured in vivo (FSI2 default, temporary iteration counters): 2.7-3.0 PCG iterations/solve from a zero initial guess. The correction field is temporally correlated step to step, so project() now seeds the FIRST corrector's solve from the previous step's p' (current fluid cells only — the p'=0 invariant on non-fluid cells survives the copy-back); later correctors and the SOR fallback start from zero exactly as before. The corrector restriction is measured, not guessed: the all-correctors draft cut the rigid phase 2.8 -> 1.27 iters/solve but cost 3.9/solve in the coupled phase (baseline 2.56) — corrector 2 solves for a much smaller correction and corrector 1's full-magnitude p' is a WORSE guess than zero there. First-corrector-only: rigid 2.8 -> ~1.1 (best of the three variants), coupled 2.62 ~= baseline. Total PCG iterations on the FSI2 default: 55,088 -> 32,836 (1.68x fewer). Wall: FSI2 default 225 -> 172 s, FSI3 default 343 -> 249 s. Session cumulative (banded LU + indexed SDF + warm start): FSI2 524 -> 172 s (3.0x), FSI3 944 -> 249 s (3.8x). This is a TOLERANCE-LEVEL solver-path change (each projection reaches the same true-residual stop from a different start), and the FSI3 release-window pins fired for the third and fourth time across the drafts — completing the picture: EVERY windowed observable of the [4.0, 4.2] release transient is branch-sensitive (four measured branches now recorded in the test: uy mid 10.77/6.02/2.91/8.88, amp 23.6/25.2/24.6/19.9, ux mid -2.90/-2.91/-2.56/-1.82, retries 2/0/1/1). The release bands are re-pinned as gross-physics tripwires around the measured scatter; the load-bearing regression pins for solver changes are the settled-cycle study bands, whose re-verification under this change is launched (verdicts to solver_status.md). Protocol: FSI2 default green (uy 3.4921 in-band), FSI3 default green under the re-pinned release bands (deterministic across two runs), FSI1 green, noise-probe floors identical, rtx-cfd suite + rtx-fsi quick tests green. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-30 08:54:27 -05:00
Omar SobhandClaude Fable 5	328f65233e	rtx-interpret: decoder cold-start window for reinitialized SAE units ... reinitialize_encoder_neuron_cold / SAETrainer::reinitialize_neuron_cold exempt a freshly reset unit's decoder column from per-step normalize_decoder for cold_steps training steps, so its small random init is not blown up to unit norm before it has learned anything — the mechanism omni-cortex's D605 refutation left as the prime suspect, now testable (D620 downstream). Runtime-only state, not carried through checkpoints (documented); cold_steps 0 is exactly the plain reinit, and existing entry points delegate with 0. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01B1feFAQxjbCRHePUdxuNra	2026-08-30 08:31:08 -05:00
Omar SobhandClaude Fable 5	6c48e53998	rtx-cfd: indexed polygon SDF — bit-identical queries, the fluid's measured hot function cut ... The 2026-08-30 fluid profile (symbolized samples, rigid AND coupled phases of the FSI3 default) attributed the fluid step to the function: polygon_signed_distance 51% rigid / 35% coupled — the embedded mask rebuild and its ghost reconstruction walk every edge of the ~150-vertex interface polygon for every cell-centre and face query, every step. (Also measured, refuting the parked consolidation: Level::new — the MG hierarchy build — is 0.5-0.7% in BOTH phases; caching it would buy nothing. The MG smoother at 34-40% is the honest remaining fluid cost.) PolygonSdf (solvers/incompressible/polygon_sdf.rs): a binned edge index whose query is BIT-IDENTICAL to polygon_signed_distance by construction — per-edge distances use the same float ops, the ring search provably visits a superset of the argmin (convex-projection lower bound sqrt(d_out^2 + ((r-1)b)^2)), and parity XORs the same ray tests over exactly the straddling edges (y-binned). Equality is ASSERTED, not assumed: tests compare to_bits against the brute force over ~40k adversarial points (flag-like walks, random polygons with degenerate zero-length edges, horizontal-edge/vertex-y rays). Wired into EmbeddedBody::polygon and the FSI harness's shared geometry (rebuilt per set_geometry, ~microseconds for 150 edges). Verification — the bar for a bit-exact change is digit identity, and it holds: FSI2 and FSI3 committed defaults reproduce EVERY printed digit of the banded-LU baseline logs (uy 3.7732±3.7920 / 6.0229± 25.2190 mm, conservation 8.26e-12 / 1.49e-12, rigid drags 121.4 / 426.9); rtx-cfd full suite 0 failures; rtx-fsi lib/piston/transfer/ FSI1 green. The study pins need no re-run: the trajectories are unchanged by construction and confirmed by measurement. Wall clock: FSI2 rigid 323 -> 167 s (1.93x), whole default 400 -> 225 s; FSI3 rigid 420 -> 250 s (1.68x), whole default 539 -> 343 s. Cumulative with the banded LU this session: FSI3 default 944 -> 343 s (2.75x), FSI2 524 -> 225 s (2.33x). Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-30 07:47:42 -05:00
Omar SobhandClaude Fable 5	0b4f306ed1	rtx-fea: reduced Newmark (mor::dynamic) + the phase-4a offline replay — the ≥10x gate is REFUTED by measurement at the validated resolution ... The dynamic layer over ReducedNonlinearModel: reduced consistent mass V'MV (full element sum, never ECSW-sampled — ECSW weights are trained on internal-force virtual work and would conserve the wrong inertia), reduced_force_and_jacobian exposed (solve() refactored onto it), and ReducedNewmark mirroring NonlinearDynamicStepper::newmark_newton in reduced coordinates (same predictor, residual, tangent shape; no rescue ladder by design — a reduced Newton death is a finding). TDD (tests/reduced_newmark.rs): identity-basis march reproduces the full stepper to 2.4e-14 over 15 steps (both Newton loops tightened to 1e-10 so only solver rounding separates them); rigid-translation reduced mass = rho*A to 1e-9; a 6-mode POD basis tracks its training trajectory at 4.2e-4 rms against a 1.0e-4 projection floor. Phase 4a (fsi3_ecsw_offline.rs, fsi3_reduced_newmark_replay, env-gated): reduced Newmark replay of the harvested FSI3 trajectory at record cadence (dt_rec = 5x march dt), driven by the recorded end-of-step loads. Measured, m=12/20: - COST (dt-independent, the verdict): 3,068/3,580 us/step at 4.6/5.0 Newton iters — 2.0-2.3x the banded full-order structural step (7,200 us/pass, bandedlu_fsi3_ny62_t85). The >=10x gate needs <=720 us/step; one reduced eval alone costs ~640 us because phase 2 refuted hyperreduction (every eval loops all 70 elements). The gate arithmetic is closed: reduced Newton needs >=2 evals, capping the ROM at ~5x. THE CAMPAIGN GATE (pinned cycle bands at >=10x structural speedup) CANNOT BE MET at the validated resolution. - TRACKING at record cadence diverges in the release transient (dies t=4.35-4.45) — and the RTX_REPLAY_IDENTITY control dies EARLIER (t=4.13) in the exact subspace: the death is the 5x-coarse integration + aliased loads, NOT the reduction. The record-cadence replay cannot judge subspace dynamics; the projection floor (1.1e-3 at m=12) remains the honest subspace statement. Campaign verdict to be recorded in omni-cortex in the pre-registered words. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-30 07:07:49 -05:00
Omar SobhandClaude Fable 5	10c779e96e	rtx-fea: banded LU replaces the dense factorization on the Newton tangent — the march's cost center, fixed ... The 2026-08-29 profile attributed 98% of the structural step (79% of a coupled FSI pass) to LuDirect::factorize — nalgebra's dense full-pivot LU on the 560-DOF tangent, every Newton iteration. The tangent is banded (half-bandwidth ~26: the flag mesh numbers the short direction innermost). BandedLu (solvers/banded.rs): LAPACK dgbtrf-style column-major band storage, partial pivoting with kl fill rows, band limits measured from the CSR pattern per factorize, O(n·kl·(kl+ku)). Swapped into NonlinearDynamicStepper (tangent + rest-state mass solve); LuDirect untouched elsewhere. TDD: 10 manufactured-system tests green first run (recovery to 1e-12 vs exact and vs LuDirect across band shapes incl. full-bandwidth degeneration; zero-diagonal pivoting; indefinite shifted-stiffness tangent; singularity; per-solve refactorization). Solver-path change — full verification protocol run: - rtx-fea 29 binaries 0 failures; rtx-fsi lib/piston/transfer green. - FSI2 committed default: every printed digit IDENTICAL to the 2026-08-28 baseline (uy 3.7732±3.7920 mm, f 2.547, conservation 8.26e-12). FSI1 identical. Noise-probe floors reproduced. - Wall clock: FSI2 coupled phase 233 s -> 77 s (3.0x, 0.60 -> 0.20 s/step); FSI3 coupled 517 s -> 119 s (4.3x). Structure is no longer the cost center; the fluid's MG-caching consolidation is next. Finding 1: newton_rescue's vacuousness guard fired — the 2026-08-24 killer (symmetric 1e4 N mid-swing reversal) converges on the PLAIN path under partial-pivot rounding at every probed combo to 1e5 N. Re-provoked: asymmetric 1e4 -> +1e5 N reversal defeats plain Newton at swing steps 3, 4 AND 5 (not knife-edge); pinned at steps 4, whose coarse-vs-fine gap (0.66x of scale) sits inside the pre-registered 0.75 band — the band is untouched. Finding 2: the FSI3 release pin fired and the PIN was the finding. uy_mid (windowed mean over [4.0,4.2]) moved 44% (10.7684 -> 6.0229 mm) while amplitude (+7%), ux mid (+0.3%) and 5.2x growth all held; the baseline's 2 IQN history-reset retries became 0 — a rounding-level branch flip at unit density ratio (the traced bistable-mask sensitivity). The windowed mean of a growing 5-Hz oscillation is not a rounding-robust observable; its band now covers both measured branches (both recorded in the assertion), amp/ux re-centered at ±35%. New trajectory re-verified deterministic digit-for-digit twice before re-pinning; green in vivo under the new pins. Study-tier pins (FSI3 sticky-mask cycle, FSI2 s=1 benchmark cycle) re-verification launched; results to be recorded in solver_status.md. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-29 23:20:50 -05:00
Omar SobhandClaude Fable 5	8a8da2383d	rtx-fsi: wall-split instrumentation — the STRUCTURE is 79% of the coupled phase ... Per-pass timers (reporting-only, default digits verified identical) around the three pass components. Measured on the FSI3 committed default: fluid 99 s (19.7%), structure 397 s (79.2%), load sampling 0.9%, state save 0.0%. The flag's Newton step costs ~195 ms/pass against ~0.5 ms of element assembly and a ~10 ms-scale 570-DOF solve — either the stepper carries ~20x implementation overhead or the ROM upside is enormous; the ECSW campaign's re-scope decision inverts accordingly (profile the stepper next). Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-29 16:50:28 -05:00
Omar SobhandClaude Fable 5	50e382c046	rtx-fsi: ECSW phase-2 offline study — POD subspace confirmed, hyperreduction refuted at this resolution ... fsi3_ecsw_offline.rs (env-gated on RTX_ECSW_SNAP; committed default is a no-op) measures the three pre-registered quantities on the harvested run-7 flag trajectory, with a numbering self-check (clamped-DOF displacement exactly 0.0 across all 3,294 snapshots). Measured: (1) the flapping manifold compresses 560 free DOFs to 12-20 POD modes at ~0.5% held-out rms projection error — the subspace exists. (2) ECSW has NOTHING TO EXPLOIT on the 70-element flag: the NNLS residual sits at single-element scale until nearly every element joins (31 el -> 10%, 51 -> 10%, 67 -> 2.5%, 70 -> 1e-15; held-out == training everywhere; cycle-only manifold identical) — each macroscopic Quad8 carries non-redundant virtual work, so no sub-percent sample smaller than the mesh exists. (3) full reduced assembly 548 us/eval; ECSW at best 1.4x at a useless 10% residual. ReducedNonlinearModel gains assemble_reduced_force (the measured quantity). The campaign's >=10x structural gate cannot come from hyperreduction at the validated resolution — recorded in the campaign doc with the re-scope options. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-29 15:06:33 -05:00
Omar SobhandClaude Fable 5	d9d8801f1a	rtx-fea: mor gains total-Lagrangian operators and a held-out ECSW residual ... The mor scope was small-strain only ("geometrically linear") — a basis trained on total-Lagrangian trajectories (the FSI flag marches with_total_lagrangian) sampled through small-strain operators would conserve the virtual work of the wrong force. ElementOperator now carries a Formulation (SmallStrain | TotalLagrangian), the TL branch mirroring NonlinearDynamicAnalysis exactly (SVK from Lame parameters, total_lagrangian::internal_force_and_tangent); train_ecsw / ReducedNonlinearModel::new keep their behavior and delegate, with _formulated variants added. ecsw_residual evaluates a trained model's ||Cw - b||/||b|| on arbitrary snapshots — the held-out generalization measurement; on the training set it reproduces training_residual to 1e-12 (pinned). Verified sharply: identity-basis reduced TL solve vs a tight-tolerance full TL solve agrees to 4.4e-15 (machine precision) while small-strain operators land 1.1e-2 away at the same load — the switch is exercised and exact. (At the default 1e-6 convergence criteria the reference itself stops 1.7e-4 short; measured and recorded in the test comment.) Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-29 15:01:09 -05:00
Omar SobhandClaude Fable 5	9fe9d7f74a	rtx-cfd + rtx-fsi: ECSW campaign phase 1 — snapshot dump + FlowField save/load ... FlowField::save/load serialize the complete field state bit-exact (all twelve matrices including *_old, predictors and sources, so a load is a true restart state), with a roundtrip test asserting to_bits equality on every value and rejection of truncated/corrupt files. The march gains an ECSW snapshot knob (RTX_FSI{2,3}_SNAP path, SNAPEVERY, default off): every N committed steps it appends an FSNP record — t, full-DOF displacement/velocity/acceleration (what rtx_fea::mor's pod_basis/train_ecsw consume, plus what the phase-4 dynamic reduction will need) and the committed sparse nodal load for the offline full-vs-reduced replay. Reporting-only: reads committed state after acceptance, no float ops on the solver path. Verified: smoke run's FSNP parsed by an independent reader (570 DOFs, correct record count, physical values); FSI2 committed default digit-identical with the knob off. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-28 21:56:37 -05:00
Omar SobhandClaude Fable 5	366a46b471	rtx-interpret: seeded SAE init + coupled per-unit optimizer reset ... SparseAutoencoder::new_seeded draws encoder/decoder weights via randn_seeded with per-tensor SplitMix64-derived seeds (same derivation as MambaBlock::new_seeded), so identical (config, seed) gives bit-exact SAEs — without it, cross-instance loss comparisons are noise (measured downstream: 0.09 vs 0.65 starts on identical data). SAETrainer::reinitialize_neuron couples the encoder-unit weight reinit with zeroing that unit's optimizer moment rows (encoder row, bias slot, decoder column), so external generate-and-test callers can't reset weights while leaving optimizer state stale — previously only the trainer's internal dead-neuron resampling did both. Note train_step's update rule is plain SGD today, so the moment reset is inert until the Adam path is switched on; the coupling is the contract either way, and a doctored-checkpoint test pins the row/column semantics. Also drops a vacuous assert!(true) smoke test that failed clippy's assertions_on_constants. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01B1feFAQxjbCRHePUdxuNra	2026-08-28 20:49:11 -05:00
Omar SobhandClaude Fable 5	a49602cf16	rtx-fsi: pin FSI3's release response and settled cycle ... Two tiers in the FSI2 pattern. The committed default (t_end 4.2) pins the deterministic release response at the digit-stable measurement (uy 10.7684 +- 23.5726 mm, ux -2.8982; +-35% bands, 2x onset-growth floor against a measured 4.6x). The sticky-mask study configuration (HYST>0, subcycle 2, t_end >= 8.4, ny 62 or 82) pins the settled flapping cycle measured across two grids at t_end 8.5 / window(2.0): uy amp in [38, 72] mm (measured 50.2 / 57.6; the un-flapped release and every pre-hysteresis dead march sit far below, and the reference +-34.9 sits OUTSIDE — landing inside the reference band is a loud finding, not a regression), mids, crossing-frequency in [4.4, 6.8] (covers the estimator's measured beat-scatter — the DFT puts both grids at 5.4-5.6 vs ref 5.46), and drag MEDIAN in [380, 530] (measured 454.0 vs ref 460.2). WindowStats gains drag/lift medians — the honest central loads (the extreme-based mids are noise-dominated at large deformation). Band anchors computed exactly as the test computes them, replayed offline against the run 7/8/10 CSVs (deterministic trajectories). Verified: both committed defaults green with the release pins live, trajectories digit-identical (FSI3's loads line gains the medians, FSI2's output byte-identical). The t_end 8.5 study pin runs at both grids are in flight. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-28 07:56:51 -05:00
Omar SobhandClaude Fable 5	137a62c4ea	rtx-fsi: time-resolved-honest load record — per-step sampling, interval median ... The reported drag/lift were one instantaneous measure_force sample every 10th coupled step, and the embedded-boundary surface force carries zero-mean sign-flipping fresh-cell pressure transients at step scale (measured: +-4,000-scale swings against a +-78 reference while displacements matched the benchmark to 0.1%, and the window MEDIANS sat near-physical). The record now samples the committed field every step and records the interval median — the estimator the CSV analysis supports; ten steps span 2-5% of a flap period, so nothing physical is smeared. Reporting only: measure_force reads the committed state. Verified on both committed defaults against pre-change logs — every displacement, conservation, and coupling digit identical; only the load lines moved, and toward the benchmark: FSI3 drag 628.66 +- 545.11 -> 443.26 +- 148.50 (ref 460.2 +- 27.47, mid now within 3.7%), lift amp 2289 -> 668; FSI2 lift amp 52 -> 28. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-27 14:58:17 -05:00
Omar SobhandClaude Fable 5	bff84ccdcc	rtx-cfd: mask hysteresis — sticky cell classification against a reference mask ... The measured FSI3 killer is the bistable mask: one geometry (|d| identical to 4 digits) samples two load branches (60 vs 120 kN), and the traced s=1 death is the secant walking a 20x load cliff (68,886 -> 1,307,938 N over a 1e-3 candidate change). A sticky band makes the load map single-valued at the crossing: formerly-fluid cells flip only at phi < -band, formerly-solid at phi > band, classified against the mask held at rebuild time — in a coupling loop, the restored committed step-start mask. Band 0 is structurally bit-identical (phi > -0.0 <=> phi > 0.0) and verified digit-for-digit on both committed defaults (FSI2 and FSI3, every physics digit). Measured cost on the translating-circle MMS at band 0.25h: +0.5% field error (u/p ratios 1.17/2.12 vs the no-hysteresis moving levels 1.16/2.11); flip delay = band/(v dt), deterministic. Exposed as RTX_FSI{2,3}_HYST in multiples of h_min. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01X2GmJXeQ2njUecEKiJZ1G2	2026-08-26 15:27:09 -05:00
Omar SobhandClaude Fable 5	3207f2d4c6	rtx-fsi: decouple the retry threshold from the widened stall acceptance ... The first STALLX draft used one threshold for both, which silently disabled the history-reset retry for the widened band: a stall in [5x, STALLX x) was accepted unretried where runs 2/3 retried it — caught by trajectory divergence at t = 5.7 (runs 2/3 digit-identical there; the run-5 variant killed at 2.3 h, artifacts preserved as *_unretried_variant_*). Now `retry_at` stays the old 5x-or-increment window (retry semantics bit-identical to the pre-knob code at any STALLX), and `stall_accept` widens only the post-retry acceptance. Default STALLX=5 keeps both thresholds equal — committed defaults unchanged. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-25 07:41:34 -05:00
Omar SobhandClaude Fable 5	992cd76449	rtx-fsi: stall acceptance as a knob (STALLX) — the floor must stay tight; only the rare-event window widens ... FSI3's developed cycle stalls (bistable mask flips at max-velocity crossings) sit at 3.5e-4 against the measured floor 6e-5: run 3 died 16% over the 5x window at t = 7.70. Riding the FLOOR up instead was measured to make it worse — run 4 at floor 1e-4 died EARLIER (t = 6.30) at a HIGHER stall (1.2e-3, 2.4x its window), because the accepted-step scatter IS the wall-velocity noise (tol/dt_c): a looser floor feeds the flip noise it is trying to pass. Keep the floor tight and widen only the counted, bounded stall window: `stall_accept` (RTX_<CASE>_STALLX, default 5 — both committed defaults bit-identical). Precedent: FSI2's s = 1 benchmark run accepted a worst stall of 5.4e-4 the same way and measured 0.1% in amplitude. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-25 05:00:56 -05:00
Omar SobhandClaude Fable 5	647f247601	rtx-fsi: report Newton rescue counts in the coupling-failure panic ... FSI3 run 3 died on a coupling stall 0.2 s after the rescue carried it past run 2's Newton death, and the panic path printed nothing about how many rescues had engaged. Observability only — the panic message now carries rescue_counts. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-25 01:07:49 -05:00
Omar SobhandClaude Fable 5	d2c82af91b	rtx-fea: rescue the nonlinear Newmark Newton — line search, then step subdivision ... Both FSI3 study deaths were the flag's SVK Newton returning ConvergenceFailed{60} inside a coupling pass at a violent mid-cycle load. The plain full-step Newton stays float-op identical (FSI2 and FSI3 committed defaults re-verified digit-for-digit, Newton rescues (0,0)); only on failure does the stepper retry: a backtracking line search on ||R|| (Armijo, alpha down to 2^-29 — 2^-8 was measured too shallow when the tangent K_T + M/(beta dt^2) is near singular and the solved direction enormous and inexact), then 2/4/8/16 Newmark substeps of dt/n, each line-searched. Rescues are counted and surfaced through MarchResult and both FSI test printouts. Measured before writing (tests/newton_rescue.rs): a static tip load from rest NEVER defeats plain Newton (1e6 N converges in 19 its — from a quiescent state the predictor is the current configuration and M/(beta dt^2) regularizes the walk) — pinned as a negative result; the killer is a mid-swing load REVERSAL (1e4 N tip load, 3 steps of swing at dt 5e-3, then reversed: dead in 60 its), the FSI3 turning-point shape — now rescued by the line search alone and consistent with a dt/32 reference march of the same interval (-0.341 vs -0.195 m, same branch), with determinism (bit-identical re-step) and march-continuation pinned alongside. Also: pin the s = 1 FSI2 benchmark cycle at study horizons (iqn / subcycle 1, t_end >= 16: f in [1.85, 2.0], uy amp in [70e-3, 92e-3] — the mode-2 s = 2 cycle fails both bands, so losing the benchmark cycle stays loud). Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-24 19:14:13 -05:00
Omar SobhandClaude Fable 5	555a72cbc0	rtx-fsi: spike-clamp the force-measurement probe ... measure_force integrated raw traction samples while the coupling loads carried the 20x-median clamp — so the s = 1 benchmark run's REPORTED drag/lift were +-4,000-scale garbage against a +-78 reference while its displacements matched the benchmark to 0.1%. Collect, clamp, then integrate, both probes. Reporting only: the committed FSI2 default reproduces its trajectory to every printed digit (uy 3.7732 / 3.7920 mm) and its rigid-phase drag (121.4) with the clamp in. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-24 05:59:18 -05:00
Omar SobhandClaude Fable 5	c1bcdc408e	rtx-fsi: FSI3 floor 6e-5 — the floor rides with the motion; measure it through the cycle ... The first subcycle-2 study (t = 4 -> 8) ran 2.2 s of cycle — over [5, 6] s: +-42.8 mm at 6.2 Hz, trending toward the reference +-34.9 mm at 5.46 Hz from above — then stalled at t = 6.18 s: residual 1.9e-4 against an accept window of 1.5e-4 (5 x the 3e-5 floor the rest-state probe suggested), at a turning point of the cycle where the step increment, and with it the increment-relative part of the window, collapses while the flip noise does not. Floor set from the cycle's measured stall level (window 3e-4). Default re-verified: 581 steps, 3.5 subit/step (max 5), 0 stalls, conservation 1.5e-12, ux -2.90 +- 2.93 mm vs reference -2.86 +- 2.70. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-23 23:41:02 -05:00
Omar SobhandClaude Fable 5	c0666bf22a	rtx-fsi: C3 — FSI3 opened: the added-mass regime, its impulse artefact, and the coupler hygiene it demanded ... The march is extracted from the FSI2 test into a shared tests/fsi2_harness/march.rs (MarchConfig from env / MarchResult / window statistics), the harness parameterised by BenchmarkCase (FSI2, FSI3: inflow, solid density, modulus, rigid-flag drag reference), and turek_hron_fsi3.rs written on top (Re 200, density ratio 1, E 5.6e6). The committed FSI2 default is bit-identical through all of it — every new knob defaults to FSI2's behaviour; re-verified twice to every printed digit (uy 3.7732 / 3.7920 mm). FSI3's first contact, traced pass by pass (RTX_FSI3_TRACE), exposed the added-mass instability in its purest form and, one measured mechanism at a time, what a partitioned coupling at unit density ratio needs: - C^1 interface motion (MarchConfig::c1_interface): a constant per-step interface velocity is a velocity JUMP at each step start, and the incompressible fluid answers with an impulsive added-mass load ~ rho L dv / dt_fluid (8x the physical reaction under subcycling): 1,600 N at release, 48,000 N and a 59 mm response one step later, the flag's Newton dead the pass after. Constant acceleration from the previous end velocity to 2 dd/dt - v_n removes the impulse (loads 1,700-2,400 N). - IQN first-pass relaxation as a knob (initial_relaxation, FSI3 0.05): |1 - omega (1 + g)| must contract; 0.5 diverges past gain 3. - The divergence verdict waits for the secant (IqnIls): the exploratory second pass on a high-gain map legitimately overshoots 10x before the first secant column exists. Pinned by a gain-40 model test. - Kinematic predictor (predictor: "kinematic", velocity only): the structure-alone predictor ignores an added mass comparable to the flag's and overshoots 2-5x, drawing 5-6x loads every first pass; and NOT with the acceleration — Newmark average acceleration carries an inconsistent initial acceleration as a sign-alternating mode (d + dt v + dt^2 a / 2 predicted 22 mm at release; converged 0.14). - Quiescent release (quiescent_release): the structure-alone "consistent" initial acceleration M^-1 F ignores the added mass. - Coupler hygiene (IqnIls): a stalled or diverged step's secant columns are no longer retained (a bistable mask flip's columns extrapolated a 30 mm interface jump on the next step); two-window stagnation detection reports a plateau early instead of bouncing to the budget (a single-window test misjudged a slowly converging step and is recorded as such); trust region tightened to 10x the residual. A noise-column filter at the tolerance was measured to HURT (stalled a converging step at 5.5e-4) and is disabled (threshold 0). - The floor measured, not borrowed (fsi2_interface_noise.rs gains RTX_NOISE_CASE=fsi3): flip jumps 3.6e-4 (12x FSI2's), the subcycle-8 release map stalling near 1e-6, the subcycle-2 map converging to 9e-10; and through the release transient (flag at ~0.3 m/s) the subcycle-8 floor rides up to ~1e-3 with the motion — which moved the FSI3 default to subcycle 2 at a 3e-5 floor. FSI3 committed default (ny 62, release t = 4, t_end 4.2, subcycle 2): 581 steps, 3.8 subit/step (max 8), 0 stalls, 0 retries, conservation 4.3e-13; uy 10.1 +- 27.5 mm and ux -3.1 +- 3.2 mm over the first 0.2 s (reference cycle 1.45 +- 34.90, -2.86 +- 2.70). Machinery invariants pinned; physics bands await the study record. Rigid-flag drag 426.9 vs CFD3's 439.45 (-2.9%). The FSI2 mode-2 study pin (IQN / subcycle 2, t_end >= 16) was measured with the pre-hygiene coupler; re-verify on the next s = 2 study run. 48 lib tests green, clippy clean. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-23 19:43:02 -05:00
Omar SobhandClaude Fable 5	be04e0e233	rtx-fsi: pin the mode-2 flapping cycle under the IQN / subcycle-2 configuration ... Study horizons (t_end >= 16) under RTX_FSI2_COUPLER=iqn RTX_FSI2_SUBCYCLE=2 now assert the measured mode-2 cycle: frequency in [2.05, 2.45] Hz and uy amplitude in [45, 70] mm — bands that admit the settled cycle at either grid (ny=62: 57.6 mm / 2.227 Hz at t=16, 55.9 / 2.224 at t=30; ny=82: 57.1 / 2.236 at t=20) and reject both the wake attractor (3.73 Hz, ±17 mm) and the reference itself (1.93 Hz, ±82 mm) as material changes. The default configuration's pins are unchanged. Module docs brought up to the eleventh-session record. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-23 16:54:12 -05:00
Omar SobhandClaude Fable 5	a342e703a4	rtx-fsi: trust-region cap on the IQN step + increment-scaled stall acceptance ... The s = 1 FSI2 march (deepest rung: 1.955 Hz, ±73 mm mid-growth) found two coupler failure modes at peak motion: - run 1 (budget 12): NotConverged at residual 6.1e-4 = 9% of the step's own increment, after the history-reset retry — killed at t = 11.1 s. - run 2 (budget 30): the deeper budget let an ill-conditioned secant model extrapolate the locally violent map into a candidate interface that swept to the domain wall and crashed the mask build BEFORE any residual guard could fire (t ~ 9.9 s). Fixes, both scale-relative per the absolute-threshold rule: - STEP_CAP = 50: the full quasi-Newton step r + W alpha is capped at 50x the current residual norm, direction kept. Legitimate large Newton steps (near-marginal gains) pass; thousand-fold geometric extrapolations cannot. Pinned by a noisy-map test asserting every iterate's step stays within the cap. - The march accepts a stalled step at residual < max(5 x tolerance, 0.1 x the step's own increment) — the rare violent step near peak motion carries an order-below-increment error, counted like every stall and bounded by the existing stall-fraction assert; the retry trigger mirrors the same bound. 46 lib tests green, clippy clean. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-21 19:14:53 -07:00
Omar SobhandClaude Fable 5	d5f19ea497	rtx-fsi: IqnIls::reset_history — stale secant columns overshoot in rapid transients ... The s = 2 ladder march (resonant growth finally open: 2.3 Hz, amplitude through ±46 mm — 2.7x the wake attractor the loose coupling locked) died at t = 11.12 s with CouplingDiverged at iteration 2: the predictor left a 1e-4 first residual, and the first quasi-Newton update drove it to 1.1e-3. Cross-step secant history assumes the interface Jacobian drifts slowly; columns recorded at much smaller amplitude steered the least-squares direction wrong at 2.7x that amplitude, and the divergence guard read the overshoot as added mass. The map itself converges deeply from a clean start (the noise probe's stall measurement), so the recovery is: reset the history, retry the step once from the predictor. - IqnIls::reset_history(), with a cold-start-equivalence test. - turek_hron_fsi2: on an unaccepted coupling verdict under the IQN coupler, reset + one retry from the predictor (Aitken carries no history — a retry would repeat the identical iteration, so the path is IQN-only); retried steps counted and reported like stalls. Verified end-to-end: the re-run crossed the killing step and marched on (t = 11.18: uy +56 mm, window amp ±50 mm, 1.4 subit/step). 45 lib tests green, clippy clean. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-21 09:34:10 -07:00
Omar SobhandClaude Fable 5	140310b223	rtx-fsi: the noise floor interrogated — smoothing refuted, IQN-ILS lands, tight coupling reopened ... The tenth session ended on "lower the interface noise floor". This builds the levers and measures them, and the measurements overturn the diagnosis: - smooth_tractions: arclength moving average over the wetted surface, area-weighted, smooth normal-similarity factor so corners do not mix and the smoothed load stays continuous in the geometry. Nine unit tests. MEASURED NEGATIVE RESULT: the flip-scan floor is unchanged to 0.2% at radii 1-3h — the flip's load jump is coherent through the fluid field (mask rebuild shifts the pressure around the flipped cell), and a surface average preserves coherent shifts. Default off; the probe pins the attribution so nobody re-reaches for this lever. - IqnIls: interface quasi-Newton with inverse least squares (Degroote 2009) — filtered MGS least squares over secant columns (filter RELATIVE to column norm), cross-step history reuse, per-step set_tolerance. Model-map tests: exact on anisotropic linear maps within dim+2 passes (scalar Aitken provably cannot be), scale invariant, history reuse shortens the next step, stalls at the noise scale instead of diverging (fixture lesson: per-pass noise, not state-dependent noise — the latter has a genuine fixed point). - tests/fsi2_harness/: the FSI2 machinery extracted shared; verified pure code motion (committed release response reproduced to every printed digit). March gains RTX_FSI2_SMOOTH / RTX_FSI2_COUPLER=iqn / RTX_FSI2_REUSE knobs; pinned bands guard the default configuration. - tests/fsi2_interface_noise.rs: the probe. Flip-scan floor at subcycle 8: 3.05e-5 (pinned); smoothing attribution pinned; the cross-subcycle scan recorded but unpinned (the fixed geometry increment's wall-velocity trend, increment/dt_c, swamps the flip signal at small dt_c — a dt_c^2 scaling hypothesis died in that operationalization). THE OPERATIONAL FLOOR — the real release step subiterated at tolerance 1e-9 with residuals traced — converges DEEP at both subcycles: s8 aitken 3.4e-9 / iqn 1.6e-9, s2 both ~6.4e-10 in 5-6 passes. The flip jumps are events at specific geometries, not a floor under every step: the tenth session's subcycle-2 blowup was tolerance mis-budgeting (2e-4 held fixed while dt_c shrank), not an impassable floor. Probe bug found and fixed on the way: stale shared geometry leaked a 4.5e-5 phantom first residual into the first stall run; every measurement now resets the geometry on entry. All 924+17 tests green: lib 44 (was 27), piston 2, curved edge 1, FSI1, the committed FSI2 march (release response identical), the probe. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq	2026-08-21 06:42:05 -07:00
quantumandClaude Fable 5	c36cf2f8a7	rtx-backend-metal: fix swap_dims returning a corrupt strided view ... swap_dims copied the buffer but wrote elements back at their ORIGINAL positions (new_idx was computed with the swapped strides), returning a stride-swapped non-contiguous tensor. Every other op in this backend — elementwise kernels, MPS matmul, to_vec — reads raw buffers and ignores strides, so any transpose consumer (notably the autograd matmul backward, grad_a = grad_c @ b^T) silently computed on untransposed data. Found via CPU-vs-Metal gradient parity on the DigiGraph HetGAT: forward matched, gradients were ~2x off. swap_dims now physically permutes into a contiguous result (reading through the input's strides + offset), and reshape asserts contiguity instead of silently reinterpreting a non-contiguous buffer. 3 new parity tests incl. matmul-after-transpose (25 total pass on-device). Co-Authored-By: Claude Fable 5 <[email protected]>	2026-08-21 06:11:38 -07:00