Repository navigation
fix: a finite-EI step no longer converges to the root with its tangents reversed - #169
Conversation
…ts reversed A finite-EI line with little numerical damping could fail its march a few steps in. Found on the lock-in line of the long-run VIV checks (a taut 20 m line, 10 elements, EA 5e7 N, 3.8 kN, pinned in a 0.3 m/s cross current with the lift) at rho_inf 1 and 50 steps per VIV cycle (dt 0.037 s): step 2 converged in 3 quadratic iterations to a state with every nodal tangent reversed (m_x -1.13 against +1.00015), 2.4e8 J of strain energy against 3.5 J, and step 6 stagnated at 1e-3 through its 50 iterations. The same happened at rho_inf 0.8 and 0.9 (failing at steps 10 and 7), at 25 steps per cycle (steps 5, 13 and 4 at rho_inf 0.8, 0.9 and 1; at 0.5 the run went on through reversed roots to a mid-span A/D of 5.7), and without the lift (step 38). It is not the lift, the force cache or the cross-step factor reuse (the reuse is off under VIV, and the step fails the same with it disabled), and not an instability of the time integration: the energy was steady up to the jump. The cause is the Newton predictor. The line's seed has unit tangents while its elements are stretched by T/EA = 7.6e-5, so the tangents' stretch, the line's stiffest motion (EA against their small consistent mass, omega ~ 8e3 rad/s, omega dt ~ 300), starts 7.6e-5 off its balance. At rho_inf 1 nothing damps it: it rings between 1 and 1 + 2 T/EA with a committed acceleration of 4.9e3 1/s^2 on the tangents, and the constant-acceleration and Newmark guesses (dt^2/2 a_n = 3.4 and dt^2/4 a_n = 1.7 of the tangent's length) carried every tangent across zero, next to the root with each element folded into a Z along its chord. With rho_inf 0.5 the same mode carries a velocity overshoot (dt v_n = 1.7 at 25 steps per cycle). Seeded at its stretch (tangents 1 + T/EA) the line runs at every rho_inf and step. Fix: where either guess would change the length of a free nodal tangent by more than a quarter (a strain change of 0.25 in one step, which no step resolves), that tangent starts at its committed value, from which the stiff mode's true increment is a Newton step away. The test is on the length alone, since a guess that turns a tangent fast is often right: the line-control test's seabed cable turns two tangents 19 degrees in its first 0.01 s step and does not converge from held tangents. Holding only the length and keeping the turn also converged, but took 132 steps over 12 iterations against 31 on the 25-step runs. And a step that ends with a nodal tangent turned by 90 degrees or more is now a failed attempt with its own message (the next attempt or the step recovery re-solves it) instead of a converged step: with the predictor guard disabled, the line stops at step 2 with that message rather than marching on through reversed roots. The line now runs at rho_inf 0.5 to 1 and 25 and 50 steps per cycle. Over 20 cycles its largest mid-span A/D is that of the line seeded at its stretch within 0.75 % and its largest strain energy within 1.2 times; at 50 steps per cycle and rho_inf 1 it runs the 400 cycles of the long-run check (largest A/D 0.568 against 0.567 seeded at the stretch; energy mismatch 7.3e-4 of the VIV work, 8.3e-5 at 100 steps per cycle). The guard changes only a step that would stretch a tangent by a quarter: every example and test-data deck gives bit-identical outputs on the old and new driver (226 files over 115 decks, the console logs equal up to their progress clock), the full test suite passes on the default and the Prescott OpenBLAS kernel (427 tests), and the 66 example decks run on both kernels with no fallback note. Test: test_hermite_tangent_predictor, the taut line seeded with unit tangents against the same line seeded at its stretch, at rho_inf 0.5, 0.8, 0.9 and 1 and 25 and 50 steps per cycle over 20 cycles, and without the lift at rho_inf 1. Before, 7 of the 9 runs failed and one ran through reversed roots (9 assertions); now every step converges (at most 5.9 iterations per step on average, 23 at most), the strain energy stays within 1.5 times the reference's and the largest A/D within 2 %. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
|
@codex review |
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 631a81751f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… tangent The guard skipped every node with a fixed tangent component, so on a planar model (out-of-plane tangent component fixed at every node) it never acted, and the reversed-tangent check then stopped the run with no guarded predictor to recover from. The guard now acts where any tangent component is free: the length change is that of the free components' increment, and only they are held. Test: G6, the no-lift line of G5 kept in its plane (r_z and m_z fixed at every node), at rho_inf 1 and 50 steps per cycle. It now runs as G5 (1000 steps, the same energy, deflection and iterations); before, it stopped at step 2 on the reversed-tangent check. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
|
@codex review |
|
Codex Review: Didn't find any major issues. Swish! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
…int phase The mid-point VIV phase update changes every coupled-VIV result, so the validation case L3-W and the numbers the documents quote from coupled VIV runs are recomputed with it, each as its text describes (deck, options, window, output flag). L3-W (validation/experiments/wang_scr), rerun as its README says, and its two sensitivity runs: the largest r.m.s. strain over the measured largest 1.0729 -> 1.0733, the station mean 1.0362 -> 1.0361, the r.m.s. log error 0.339 -> 0.338, the dominant frequency 0.992 Hz (unchanged), 7 r.m.s. peaks along the span (unchanged); 400 elements 1.0753 -> 1.0759 and half the time step 1.0742 -> 1.0747 (0.24 % and 0.13 % from the base run). Every number and sentence of the README, VALIDATION.md and the CHANGELOG still holds at its printed precision; summary.json, the comparison CSV and provenance.json (commit, date, sensitivity) are updated. Numbers changed: - doc/outputs.rst, examples/viv_taut_cable_inline.dat: the VIV loads' power over 60-90 s 8.52 -> 8.45 W, the drag (and damping) 8.56 -> 8.48 W (balance within 0.5 %, unchanged). - doc/python_postprocessing.rst: the travelling-wave index 0.05 -> 0.04. - doc/theory.rst: the EI = 0 wake-oscillator line locks in at 0.283 Hz (was "the same 0.282 Hz"); the amplitude estimate within 0.04 % (0.1 % on the EI = 0 line; was 0.03 %); Thorsen's damping A/D 1.02 -> 1.01; the moving/held pipe table (peak 0.819/0.819 -> 0.811/0.818, RMS 0.446/0.447 -> 0.446/0.446; without the row 0.711/0.850 -> 0.700/0.851 and 0.349/0.450 -> 0.345/0.450; Cl alone 0.713/0.714 -> 0.713/0.712 and 0.372/0.371 -> 0.372/0.370; without 0.711/0.736 -> 0.700/0.733 and 0.349/0.375 -> 0.345/0.374), the RMS agreement within 0.2 % -> 0.6 % (0.02 % with the table); the towed cable 0.7905 / 0.7906 -> 0.7902 / 0.7903, 0.665 -> 0.663, 0.6764 / 0.6764 -> 0.6758 / 0.6758, Tf on the held cable at most 0.23 % -> 0.24 %. - doc/driver_format.md: towed within 0.01 % -> 0.02 %, at most 0.23 % -> 0.24 %. - VALIDATION.md: L2-VIV zero seed 0.701 -> 0.699 against 0.699; L2-VIV-EI 0.662/0.421 against 0.663/0.421 -> 0.661/0.420 against 0.661/0.420; the coupled VIV row 0.6645 / 0.6646 -> 0.6631 / 0.6625 (standalone 0.665 -> 0.663). The L2-VIV band against MoorDyn-F, rerun over 150-300 s at U 0.06-0.42 m/s: mode 1 at 0.09-0.18 m/s within 0.2 %, mode 3 at 0.39-0.42 m/s within 1.4 %, and at 0.36 m/s MoorDyn-F settles in mode 2 and CableDyn in mode 3 (the earlier phase update gives the same branches, its A/D within 0.5 % of the new one at every speed, so the discrepancy predates this change); the row said "mode 3 at U >= 0.36 m/s within 1.2 %". - CHANGELOG.md: the wake entry's EI = 0 frequency, the amplitude estimate 0.03 % -> 0.04 %, the frame entry's pipe and tow numbers as in the theory, and the #169 entry's largest A/D within 0.75 % -> 0.1 % except at rho_inf 1 and 25 steps per cycle. - tests/test_viv_frame_deck.f90 and tests/test_hermite_viv_frame.f90 (header comments only): the measured values of their gates. Every other quoted coupled-VIV number still holds at its precision (the example decks' A/D 0.66, 0.82, 0.19 and 0.03, their frequencies, the in-line and higher-harmonic gates, the wake, diameter and Thorsen gates, the riser case). The Delta Flume case was rerun before; the NDP case is on another branch. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
…tep (#171) A taut line locked in by the synchronisation model changed its lock-in frequency by 0.74 %, 0.41 % and 0.27 % per halving of the step from 25 to 200 steps per cycle (the long-run VIV checks), worse than first order, while the same phase update under prescribed motion was second order. The phase sampled the committed generalised-alpha acceleration a_n for its cross-flow phase (and the wake oscillator for its forcing). That acceleration is the line's at t_n + (alpha_m - alpha_f) dt only, a first-order error at rho_inf < 1, and at rho_inf = 1 it carries an undamped alternating mode of the integrator: on the 20 m line the mid-span a_n alternated by +-11 m/s^2 against a physical acceleration of about 1.3 m/s^2. The lock-in at rho_inf = 1 therefore converged to a frequency 3.4 % low (3.940 at 3200 steps per cycle against 4.0797 rad/s), and at rho_inf 0.8 and 0.4 it was first order (0.06/n and 0.25/n at n steps per response period). Prescribed motion imposes the exact acceleration, which is why it did not show there. Candidates switched off one at a time on that line, 25 to 3200 steps per cycle: a second-order running mean square (exact exponential with trapezoidal samples) changed nothing; a fixed normalisation in place of the RMS left it first order; a tighter Newton tolerance changed nothing; the lift at a fixed shedding rate (no phase feedback) was second order (ratios 3.8 to 4.0); the acceleration recomputed from the equations of motion after every step moved the limit to 4.0797 rad/s; and the phase on the step's mid-point motion with the acceleration from the converged velocities restored second order (ratios 3.6 to 5.0, then 4.01 to 4.04 from 400 steps on) at rho_inf 1 and 0.8 alike. No amplitude table, flow frame or in-line VIV is involved in that case. The phase now advances by the implicit mid-point rule, one fixed-point pass, on the node's cross-flow motion at the step's mid-point: the mean position, the velocity (q_n+1 - q_n)/dt and the acceleration (v_n+1 - v_n)/dt, the step means of the velocity and of its derivative (second order at every rho_inf, free of the committed acceleration's lag and alternating mode, and keeping the phase of a harmonic motion exactly). The step's lift takes the phase predicted with the rate of the previous step's mid-point, a new row of the VIV motion statistics (two with in-line VIV), so the step recovery, sub-step slots, snapshots and checkpoints carry it and a restart stays bit-identical; a newly installed VIV model seeds it from the present state. The running mean squares advance by the mid-point rule (1 - exp(-dt/T_w) on the mid-point sample), the in-line phase in the same way slaved to the mid-point cross-flow rate, the flow frame's and the wake filter's filtered velocities exactly for the velocity linear over the step, and the wake oscillator under the mid-point acceleration (held over the step, its integral exact) at the mean of the start and end wake frequencies, that acceleration being its predictor (row 2 of its statistics). The time discretisation of the VIV phase therefore changes on purpose; this is not a round-off change. Measured on the same line from 25 to 3200 steps per cycle (EI = 0 line and rho_inf 0.8 unless noted), the ratio of successive changes per halving of the step, before and after: frequency, rho_inf 1: limit 3.4 % low; 3.6-4.9 frequency, rho_inf 0.8 / 0.4: 2.0-3.4 / 1.5-3.2; 3.6-4.8 / 3.8-4.5 mid-span A/D, in-line VIV: 2.2-2.3; 2.6-4.3 frequency, flow frame: 1.6-2.0; 3.1-4.5 frequency, amplitude table: 1.9-2.3; 3.0-4.5 wake oscillator, frequency: 2.0-2.5; 3.8-4.0 finite-EI cable, frequency: 1.7-2.3; 3.1-4.1 The Richardson limits at rho_inf 1, 0.8 and 0.4 agree to 1e-9 (4.0797304 rad/s), and the errors are about 2.2/n^2 (rho_inf 1, 0.8) and 3.4/n^2 (0.4) in frequency and 1.6/n^2 and 3.4/n^2 in mid-span A/D at n steps per response period, 3.5/n^2 on the finite-EI cable. With the cable seeded at its static stretch, rho_inf 1 reaches the rho_inf 0.8 limit; from unit tangents the undamped tangent-stretch mode rings and moves it at fine steps (the seed, not the VIV update). Under prescribed motion the phase error of the long-run check after 400 cycles at 100 steps per cycle is 0.0167 rad against 0.0334 before, still second order (3.99). The energy closure of that check at 100 steps per cycle is 2.0e-4 of the VIV work against 8.3e-5 before, second order (ratio 4.0): the lift's predicted phase now takes the rate of the previous step's mid-point, half a step behind the old step-start rate. The two VIV example decks change at their default dtM 0.02 s (about 170 steps per response period, rhoInf 0.4) and move toward their results at dtM 0.00125 s. viv_taut_line_crossflow, against dtM 0.00125 s: the mid-span RMS 0.31 % high before, 0.02 % low after; the lock-in frequency 0.15 % low before, 0.018 % after; the fairlead tension RMS 0.29 % high before, 0.027 % low after. viv_taut_cable_inline: the quarter-span in-line RMS 1.56 % low before, 0.34 % after; the cross-flow frequency 0.10 % high before, 0.022 % after; the largest fairlead tension 0.70 % high before, 0.23 % low after; the mid-span RMS 0.21 % and 0.20 % low. Over the 90 s the phase moves, so the time series differ pointwise by up to 12 % (crossflow, fairlead tension) and 36 % (inline, quarter-span y) of the channel's range. Every other example and test-data deck (132 of 134) gives bit-identical output files and console logs. The full test suite (430 tests, on dev with #169 and #170) passes on the default and the Prescott OpenBLAS kernel, and the 66 example decks run with no fallback note on both. Test: viv_step_order, the lock-in frequency of the line second order from 100 to 400 steps per cycle at rho_inf 1 (ratio 3.99) and 0.4 (3.67) and within 1e-4 of each other (4.3e-5), the wake oscillator's crossing frequency (3.90) and the finite-EI cable from 200 to 800 steps per cycle (4.39); on the previous code every check fails (ratios 2.44, 2.27, 2.18 and 2.76, the two rho_inf 3.1 % apart). The VIV state layouts in test_model, test_system, test_viv_coupled, test_hermite_viv_frame and test_hfmf_mirror carry the predictor rows. With the mid-point VIV update the predictor check of the lifted taut line fails one cell, rho_inf 1 at 25 steps per cycle (21 per response period): from the unit-tangent seed z max/D 0.5903 against the balanced seed's 0.5641 (+4.6 %) and Es max 1.51 times it. Every step converges (6.15 iterations on average, at most 21). At rho_inf 1 nothing damps the unit seed's stretch ringing, and at that resolution it leaves a near-Nyquist velocity in the trajectory: the mid-node (v_n+1 - v_n)/dt alternates by 4.8 m/s^2 against a VIV acceleration of 3.2 m/s^2 (a velocity of about 0.09 m/s, 27 % of the VIV velocity; the committed a_n alternates by 32 against 17, 19 against 10.5 on the old update), 10 to 20 times less at 50 steps per cycle. The response then depends on how the VIV update treats that velocity: four equally consistent variants (the last two mid-point accelerations averaged, a three-point second-order Nyquist-free acceleration, an averaged predictor rate, both of the latter) give z max/D 0.623, 0.579, 0.589 and 0.578, +0.8 % to +10 % from the converged 0.5639 with the old update's +0.8 % among them. That value is the fine-step one (200 to 800 steps per cycle), the same at rho_inf 0.8 with either seed and at rho_inf 1 with the balanced seed. The new update reaches it from 50 steps per cycle with either seed (unit seed 0.5632 at 50, 0.5641 at 100, 0.5639 from 200 on), while the old update settled at rho_inf 1 0.23 % high from the unit seed and 0.1 % low from the balanced one. The result is deterministic (Newton tolerance 1e-6 to 1e-10 and a 1e-7 seed tilt change it by under 0.03 %). So that cell measures the VIV update's treatment of an unresolved spurious velocity, not the predictor. It keeps G1 (every step converges) and G4 (iterations) and skips G2 and G3, and a new gate G7 holds the convergence claim: the unit seed at rho_inf 1 and 50 steps per cycle within 0.5 % of the converged z max/D 0.5639 (measured 0.5632). The other cells keep G2 and G3 (measured at most 1.07 times and 0.09 %). doc/theory.rst and the CHANGELOG note that at rho_inf 1 from an unbalanced seed the VIV response is resolution-sensitive below about 50 steps per cycle: keep rhoInf at or below 0.95 or seed the tangents at the stretch. The nine cases and their sensitivity runs (half the time step, 128 elements, 80 s for cases 5, 6, 8 and 9, the 38.1 kN/m top spring) rerun on the second-order VIV phase update. The counts are unchanged: the dominant mode right in 5 of 9 and within one in all, the mean in-line displacement within 17 % in 7 of 9 (-25 % to +34 %). The cross-flow amplitude is -45 % to +4 % against the maxima (-40 % to +18 % against the envelopes; within the 12 % measurement scatter in cases 2, 4, 6 and 7), the cross-flow curvature -39 % to +26 %, the in-line curvature -53 % to +111 %. Half the time step now changes the peak amplitude of cases 1-4, 7 and 8 by at most 2.7 % (4.5 % before) and their in-line curvature by at most 1.2 % (10 % before). Case 5 keeps its peak amplitude within 0.6 % (it moved by up to 20 % before) while its in-line curvature moves by up to 25 %. The peak amplitude of case 6 now depends on the step and the mesh: -11 % at half the time step, kept over an 80 s run, and -13 % with 128 elements, its other metrics within 2 % (in-line curvature 12 %). Case 9 still settles in mode 7 at half the time step. README, provenance.json (new 80 s runs of cases 6 and 8 at both steps), summary.json, response_comparison.csv, VALIDATION.md and the changelog entry follow. The mid-point VIV phase update changes every coupled-VIV result, so the validation case L3-W and the numbers the documents quote from coupled VIV runs are recomputed with it, each as its text describes (deck, options, window, output flag). L3-W (validation/experiments/wang_scr), rerun as its README says, and its two sensitivity runs: the largest r.m.s. strain over the measured largest 1.0729 -> 1.0733, the station mean 1.0362 -> 1.0361, the r.m.s. log error 0.339 -> 0.338, the dominant frequency 0.992 Hz (unchanged), 7 r.m.s. peaks along the span (unchanged); 400 elements 1.0753 -> 1.0759 and half the time step 1.0742 -> 1.0747 (0.24 % and 0.13 % from the base run). Every number and sentence of the README, VALIDATION.md and the CHANGELOG still holds at its printed precision; summary.json, the comparison CSV and provenance.json (commit, date, sensitivity) are updated. Numbers changed: - doc/outputs.rst, examples/viv_taut_cable_inline.dat: the VIV loads' power over 60-90 s 8.52 -> 8.45 W, the drag (and damping) 8.56 -> 8.48 W (balance within 0.5 %, unchanged). - doc/python_postprocessing.rst: the travelling-wave index 0.05 -> 0.04. - doc/theory.rst: the EI = 0 wake-oscillator line locks in at 0.283 Hz (was "the same 0.282 Hz"); the amplitude estimate within 0.04 % (0.1 % on the EI = 0 line; was 0.03 %); Thorsen's damping A/D 1.02 -> 1.01; the moving/held pipe table (peak 0.819/0.819 -> 0.811/0.818, RMS 0.446/0.447 -> 0.446/0.446; without the row 0.711/0.850 -> 0.700/0.851 and 0.349/0.450 -> 0.345/0.450; Cl alone 0.713/0.714 -> 0.713/0.712 and 0.372/0.371 -> 0.372/0.370; without 0.711/0.736 -> 0.700/0.733 and 0.349/0.375 -> 0.345/0.374), the RMS agreement within 0.2 % -> 0.6 % (0.02 % with the table); the towed cable 0.7905 / 0.7906 -> 0.7902 / 0.7903, 0.665 -> 0.663, 0.6764 / 0.6764 -> 0.6758 / 0.6758, Tf on the held cable at most 0.23 % -> 0.24 %. - doc/driver_format.md: towed within 0.01 % -> 0.02 %, at most 0.23 % -> 0.24 %. - VALIDATION.md: L2-VIV zero seed 0.701 -> 0.699 against 0.699; L2-VIV-EI 0.662/0.421 against 0.663/0.421 -> 0.661/0.420 against 0.661/0.420; the coupled VIV row 0.6645 / 0.6646 -> 0.6631 / 0.6625 (standalone 0.665 -> 0.663). The L2-VIV band against MoorDyn-F, rerun over 150-300 s at U 0.06-0.42 m/s: mode 1 at 0.09-0.18 m/s within 0.2 %, mode 3 at 0.39-0.42 m/s within 1.4 %, and at 0.36 m/s MoorDyn-F settles in mode 2 and CableDyn in mode 3 (the earlier phase update gives the same branches, its A/D within 0.5 % of the new one at every speed, so the discrepancy predates this change); the row said "mode 3 at U >= 0.36 m/s within 1.2 %". - CHANGELOG.md: the wake entry's EI = 0 frequency, the amplitude estimate 0.03 % -> 0.04 %, the frame entry's pipe and tow numbers as in the theory, and the #169 entry's largest A/D within 0.75 % -> 0.1 % except at rho_inf 1 and 25 steps per cycle. - tests/test_viv_frame_deck.f90 and tests/test_hermite_viv_frame.f90 (header comments only): the measured values of their gates. Every other quoted coupled-VIV number still holds at its precision (the example decks' A/D 0.66, 0.82, 0.19 and 0.03, their frequencies, the in-line and higher-harmonic gates, the wake, diameter and Thorsen gates, the riser case). The Delta Flume case was rerun before; the NDP case is on another branch. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
Summary
A finite-EI line with little numerical damping could fail its march a few steps in: Newton converged quadratically to a spurious root with every nodal tangent reversed, with each element folded into a Z along its chord. This PR guards the Newton predictor against it and turns such a root into a failed attempt instead of an accepted step. Steps that converge today are unchanged: every example and test-data deck is bit-identical.
How it was found
The case is the lock-in line of the long-run VIV checks:
What happened:
Without the lift, the$\rho_\infty = 1$ run at 50 steps per cycle fails at step 38.
Ruled out:
Cause: the predictor
The line's seed has unit tangents, while its elements are stretched by$T/EA = 7.6\times10^{-5}$ . The length of a nodal tangent is the line's stretch, its stiffest motion: EA against the tangent's small consistent mass, $\omega \approx 8\times10^3$ rad/s, $\omega\Delta t \approx 300$ .
That mode starts$7.6\times10^{-5}$ off its balance. At $\rho_\infty = 1$ nothing damps it, so it rings between 1 and $1 + 2T/EA$ with a committed acceleration of $4.9\times10^3$ 1/s² on the tangents. Both guesses then carry every tangent across zero:
Newton lands on the root with the element folded back. At$\rho_\infty = 0.5$ the same mode carries a velocity overshoot instead ($\Delta t, v_n = 1.7$ at 25 steps per cycle). Seeded at its stretch (tangents $1 + T/EA$ ), the line runs at every $\rho_\infty$ and step size tested.
Fix (
dyn_step_core)PRED_STRETCH_MAX = 0.25, a strain change of 0.25 in one step, which no step resolves), that tangent starts at its committed value. The stiff mode's true increment is then one Newton step away. The guard reads committed state only, so restarts stay exact.Both are documented in
doc/theory.rst(Newton iteration).Results
doc/options.rstalready advisesrhoInf≤ 0.95.Test
tests/test_hermite_tangent_predictor.f90(ctesthermite_tangent_predictor, 1.6 s in Release).Checks
pre-commit.