# LSODE2 Story Test Registry
This file is the release-run notebook for LSODE2 scenario tests. Unit and parity
tests answer whether an individual formula or ODEPACK-style transition is correct.
Story tests answer a different question: which complete route is selected, whether
two routes produce the same numerical answer, and where a real solve spends its
wall-clock time.
When a release run is performed, paste the important tables under the corresponding
test and fill in `Analysis`. A table without a hypothesis is only expensive console
art; this file is meant to keep the experiments legible months later.
Current source of truth: runs marked `CPU 12 Core` were produced on the newer
12-core / 64 GB machine and should be used for current performance conclusions.
Older `CPU 4 Core` tables are intentionally kept as historical comparison data:
they are useful for seeing how chunking, hot callbacks and linear algebra scale
across machines, but they are no longer the primary baseline.
## Executive Summary
These are the current high-level conclusions from the LSODE2 story suite. Each
line points to the story test that supports it, so the claim can be rechecked
when hardware, compiler versions or backend internals change.
1. `AtomView` is the preferred symbolic frontend for generated LSODE2 routes.
It is consistently cheaper than `ExprLegacy` on the combustion symbolic
frontend story while preserving the numerical answer. Evidence:
`lsode2_combustion_symbolic_frontend_sparse_banded_multi_run_dashboard`.
2. For cold AOT startup, the practical toolchain is currently `C + tcc`.
On the 12-core machine, `tcc` cold rows are much faster than `gcc`, Rust AOT
and Zig for this LSODE2 combustion fixture. Zig is correct, but not a
practical cold-start recommendation here. Evidence:
`lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix`.
3. `BuildIfMissing -> RequirePrebuilt` is green for the same-process linked
runtime lifecycle. The first run can install/link the compiled backend, and
later strict `RequirePrebuilt` runs reuse it with sub-millisecond preparation
and roundoff-level solution differences. Evidence:
`lsode2_combustion_sparse_banded_atomview_tcc_build_then_require_prebuilt_story`.
4. Warm `tcc RequirePrebuilt` is correct and low-overhead, but it is not yet a
clear total wall-clock win on the small Banded combustion fixture. It has a
faster Jacobian callback timer than Lambdify, but Lambdify still wins total
warm time on this workload because fixed AOT handoff overhead is not fully
amortized. Evidence:
`lsode2_combustion_banded_atomview_lambdify_vs_tcc_prebuilt_warm_cooldown_story`.
5. For this combustion fixture, explicit generated-backend chunking is not a
win. Whole callbacks are usually as fast or faster, especially for Jacobian
evaluation. This is a negative but useful result: the system is too small for
chunking overhead to amortize. Evidence:
`lsode2_combustion_like_parallel_chunking_multi_run_story_dashboard` and
`lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix`.
A larger synthetic chain story at `n=96` confirms the same conclusion for
the current LSODE2 generated callback workload: `tcc-whole` and `tcc-chunk`
are correctness-equivalent, but chunking does not reduce warm wall-clock or
hot Jacobian time. Evidence:
`lsode2_large_chain_tcc_chunking_sparse_banded_warm_story`.
6. Banded linear algebra is the preferred route when the IVP Jacobian is truly
banded. Sparse remains the safe general-purpose route, but Banded reduces the
linear stage substantially on banded combustion-like workloads. Evidence:
`lsode2_combustion_like_multi_run_story_dashboard`,
`lsode2_combustion_like_parallel_chunking_multi_run_story_dashboard`, and
`lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix`.
7. Method switching is now covered on both sides of the basic LSODA-style
decision space: non-stiff Adams routes are checked by a small corpus, and a
stiff acceptance gate proves automatic mode can execute BDF. A richer mixed
regime story can still be added later, but the core Adams/BDF execution
evidence is no longer missing. Evidence:
`lsode2_nonstiff_adams_corpus_sparse_banded_dashboard` and
`lsode2_stiff_switch_acceptance_sparse_banded_executes_bdf`.
8. On the long three-body benchmark, Banded whole AOT is the best route in the
current dashboard: it beats Sparse whole AOT, Lambdify, and both chunked AOT
variants. The new chunking-plan diagnostics show that the 12-core setup still
fragments this workload down to 12 chunks with roughly 1 work unit per chunk,
so chunking remains overhead-only here. The route-specific call counters
should be treated as telemetry rather than a direct Lambdify-vs-AOT
equivalence proof. Evidence:
`lsode2_three_body_problem_backend_story_dashboard`.
## Running Policy
Heavy tests are `ignored` where they build AOT artifacts or repeat a sizeable
combustion solve. Run them in release mode one test at a time:
```powershell
cargo test --release <test_name> -- --ignored --nocapture --test-threads=1
```
`--test-threads=1` serializes test functions only. It does not turn off parallel
symbolic work or generated AOT callback chunking selected by the solver.
Mandatory LSODE2 mirroring gates are the focused unit/parity tests listed in
`MIRRORING_CHECKLIST.md`. Story tests are advisory quality/performance evidence
unless a section explicitly calls a row an acceptance gate. In other words, a
story table can guide backend recommendations, but Fortran-faithful algorithmic
parity is locked by the parity modules first.
Interpret stage columns carefully. `total_ms` is the wall-clock user experience.
`prepare_ms` includes symbolic/backend preparation and, for cold AOT routes, build
and linking work. `solve_ms`, `residual_ms`, `jacobian_ms`, and `linear_ms` expose
the numerical solve and its hot stages. The current LSODE2 statistics expose these
solver-level stages; a finer compiler pipeline breakdown, like the BVP codegen
notebook has, remains a follow-up if cold AOT preparation becomes the dominant issue.
Use this result template:
```text
Date:
Command:
Machine/toolchain:
Status:
Important numbers:
Analysis:
Follow-up:
```
## Existing Baseline Stories
### `lsode2_native_quality_dashboard_bridge_vs_faithful_native`
File: `src/numerical/LSODE2/story_tests.rs`
Hypothesis: faithful native Sparse/Banded execution should preserve solution quality
relative to bridge-backed execution while reporting its own native step, Jacobian
and linear-solve telemetry.
Command:
```powershell
cargo test --release lsode2_native_quality_dashboard_bridge_vs_faithful_native -- --nocapture --test-threads=1
```
Result:
[LSODE2 story] native quality dashboard: bridge solve vs faithful native solve; all time columns are milliseconds
Bridge | Sparse | sparse | faer_sparse_lu | forced_by_linear_solver_policy | finished | 6.747 | 1.000e0 | - | 2.986e-7 | 8.119e-7 | - | - | -
NativeFaithful | Sparse | sparse | faer_sparse_lu | forced_by_linear_solver_policy | finished_native_faithful | 0.978 | 1.000e0 | yes | 2.767e-7 | 7.522e-7 | 91 | 15 | 146
Bridge | Banded | banded | lapack_faithful_banded_lu | forced_by_linear_solver_policy | finished | 4.731 | 1.000e0 | - | 2.986e-7 | 8.119e-7 | - | - | -
NativeFaithful | Banded | banded | lapack_faithful_banded_lu | forced_by_linear_solver_policy | finished_native_faithful | 1.875 | 1.000e0 | yes | 2.767e-7 | 7.522e-7 | 91 | 15 | 146
[LSODE2 story] native quality dashboard timings: bridge/native counters are milliseconds
path | matrix | native_solve_ms | native_residual_ms | native_jacobian_ms | native_linear_ms | bridge_solve_ms | bridge_nlu
------------------------------------------------------------------------------------------------------------------------------------------------
Bridge | Banded | 4.556 | 0.001 | 0.001 | 0.001 | 4.431 | 6
NativeFaithful | Banded | 1.849 | 0.041 | 0.019 | 0.017 | - | -
[LSODE2 story] native quality dashboard ODEPACK-style flags (JCUR/IPUP/IPUP_REASON/KFLAG/ICF/IRET/REDO): first vs last attempt
path | matrix | first_jcur | first_ipup | first_pred_reason | first_ipup_reason | first_kflag | first_kcode | first_icf | first_iret | first_redo | first_iredo | first_ialth | last_jcur | last_ipup | last_pred_reason | last_ipup_reason | last_kflag | last_kcode | last_icf | last_iret | last_redo | last_iredo | last_ialth
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Bridge | Banded | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | -
NativeFaithful | Banded | stale | up_to_date | none | none | ok | 0 | none | normal | none | 0 | 1 | stale | up_to_date | rc | none | ok | 0 | none | normal | none | 0 | 2
[LSODE2 story] native quality dashboard ODEPACK-style aggregate counters over step attempts
path | matrix | predict_attempts | reported_attempts | jcur[cur/stale] | ipup[up/need] | pred_reason[none/rc/msbp/rc+msbp/fail] | final_reason[none/rc/msbp/rc+msbp/fail] | kflag[ok/err/err_rep/conv/conv_rep] | icf[none/refresh/no_recover] | iret[normal/rescale/retry/restart] | redo[none/corr_refresh/corr_retry/err_retry/err_reset/history] | ialth[zero/pos/sum]
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Bridge | Banded | 2 | 2 | 0/2 | 1/1 | 2/ 0/ 0/ 0/ 0 | 1/ 0/ 0/ 0/ 1 | 2/ 0/ 0/ 0/ 0 | 2/ 0/ 0 | 1/ 1/ 0/ 0 | 1/ 0/ 0/ 0/ 0/ 1 | 0/ 2/ 4
NativeFaithful | Banded | 106 | 91 | 0/91 | 66/25 | 77/21/ 0/ 0/ 8 | 66/ 0/ 0/ 0/ 25 | 91/ 0/ 0/ 0/ 0 | 91/ 0/ 0 | 66/ 25/ 0/ 0 | 66/ 0/ 0/ 0/ 0/ 25 | 0/ 91/216
ok
Analysis:
Correctness is preserved: both faithful native routes reach the same terminal
state as their bridge baseline and slightly reduce the final error. On this
small fixture, NativeFaithful Sparse is about 6.9 times faster than Bridge
Sparse (`0.978` versus `6.747` ms), while NativeFaithful Banded is about 2.5
times faster than Bridge Banded (`1.875` versus `4.731` ms). Sparse and
Banded native routes have identical accepted/rejected steps and native
control-plane counters, so changing the linear backend does not change the
LSODE2 trajectory.
The isolated Banded linear stage is cheaper than Sparse (`0.017` versus
`0.069` ms), although total native time is higher on this scalar problem.
Fixed overhead dominates such a small workload; larger structured systems are
the meaningful place to compare Sparse and Banded throughput.
### `lsode2_quality_dashboard_stiff_vs_nonstiff_auto_switch`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: LSODA-like automatic selection should expose a meaningful method-family
decision for both stiff and non-stiff equations on Sparse and Banded paths.
Command:
```powershell
cargo test --release lsode2_quality_dashboard_stiff_vs_nonstiff_auto_switch -- --nocapture --test-threads=1
```
Result:
test numerical::LSODE2::story_tests2::lsode2_quality_dashboard_stiff_vs_nonstiff_auto_switch ... [LSODE2 story] quality dashboard (algorithm focus); counters are counts, time is milliseconds
nonstiff-decay | Sparse | 3/3 | adams | adams | switch_advantage_not_met | 75.00+/-0.00 | 12.00+/-0.00 | 159.00+/-0.00 | 0.00+/-0.00 | 1.434+/-0.317 | 0.000+/-0.000 | ok 3/3
nonstiff-decay | Banded | 3/3 | adams | adams | switch_advantage_not_met | 75.00+/-0.00 | 12.00+/-0.00 | 159.00+/-0.00 | 0.00+/-0.00 | 5.915+/-1.091 | 0.000+/-0.000 | ok 3/3
stiff-tracking | Sparse | 3/3 | adams | adams | switch_advantage_not_met | 106.00+/-0.00 | 24.00+/-0.00 | 237.00+/-0.00 | 0.00+/-0.00 | 2.021+/-0.089 | 0.000+/-0.000 | ok 3/3
stiff-tracking | Banded | 3/3 | adams | adams | switch_advantage_not_met | 106.00+/-0.00 | 24.00+/-0.00 | 237.00+/-0.00 | 0.00+/-0.00 | 4.835+/-0.248 | 0.000+/-0.000 | ok 3/3
ok
Analysis:
Numerical parity is clean: Sparse and Banded obtain identical solutions and
identical integration counters in both scenarios. The switching hypothesis is
not fully demonstrated, however. The nominally stiff case still executes
Adams and records `switch_advantage_not_met`, exactly as the nonstiff case
does.
This test is therefore a useful telemetry and backend-parity gate, but not yet
an acceptance test for LSODA-like automatic stiff detection. A separate
fixture known to execute BDF, such as a sufficiently demanding Robertson or
combustion setup, is still needed.
### `lsode2_nonstiff_adams_corpus_sparse_banded_dashboard`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: fixed Adams and automatic Adams/BDF modes should both behave
cleanly on a small non-stiff corpus, not only on a single scalar decay case.
The fixed Adams rows are acceptance-like: they must execute Adams and must not
execute BDF. The automatic rows are telemetry rows: they must solve correctly
and expose a valid family decision without forcing a brittle “always choose X”
policy.
Command:
```powershell
cargo test --release lsode2_nonstiff_adams_corpus_sparse_banded_dashboard -- --nocapture --test-threads=1
```
Release result, CPU 12 Core:
```text
running 1 test
[LSODE2 story] non-stiff Adams corpus: fixed Adams and automatic controller routes
nonstiff-scalar-decay | Sparse | adams_only | 3/3 | adams | adams | fixed_controller | 1.0+/-0.0 | 1.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 73.0+/-0.0 | 12.0+/-0.0 | 0.69+/-0.38 | 3.87e-7+/-0.0e0 | ok 3/3
nonstiff-scalar-decay | Sparse | automatic_adams_bdf | 3/3 | adams | adams | switch_advantage_not_met | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 94.0+/-0.0 | 12.0+/-0.0 | 0.66+/-0.10 | 3.87e-7+/-0.0e0 | ok 3/3
nonstiff-scalar-decay | Banded | adams_only | 3/3 | adams | adams | fixed_controller | 1.0+/-0.0 | 1.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 73.0+/-0.0 | 12.0+/-0.0 | 1.26+/-0.23 | 3.87e-7+/-0.0e0 | ok 3/3
nonstiff-scalar-decay | Banded | automatic_adams_bdf | 3/3 | adams | adams | switch_advantage_not_met | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 94.0+/-0.0 | 12.0+/-0.0 | 1.21+/-0.14 | 3.87e-7+/-0.0e0 | ok 3/3
nonstiff-system2-decay | Sparse | adams_only | 3/3 | adams | adams | fixed_controller | 1.0+/-0.0 | 1.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 91.0+/-0.0 | 16.0+/-0.0 | 0.62+/-0.13 | 1.44e-6+/-0.0e0 | ok 3/3
nonstiff-system2-decay | Sparse | automatic_adams_bdf | 3/3 | adams | adams | switch_advantage_not_met | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 112.0+/-0.0 | 16.0+/-0.0 | 0.70+/-0.11 | 1.44e-6+/-0.0e0 | ok 3/3
nonstiff-system2-decay | Banded | adams_only | 3/3 | adams | adams | fixed_controller | 1.0+/-0.0 | 1.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 91.0+/-0.0 | 16.0+/-0.0 | 1.20+/-0.05 | 1.44e-6+/-0.0e0 | ok 3/3
nonstiff-system2-decay | Banded | automatic_adams_bdf | 3/3 | adams | adams | switch_advantage_not_met | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 112.0+/-0.0 | 16.0+/-0.0 | 1.42+/-0.25 | 1.44e-6+/-0.0e0 | ok 3/3
test numerical::LSODE2::story_tests2::lsode2_nonstiff_adams_corpus_sparse_banded_dashboard ... ok
```
Debug verification:
```text
[LSODE2 story] non-stiff Adams corpus: fixed Adams and automatic controller routes
nonstiff-scalar-decay | Sparse | adams_only | 3/3 | adams | adams | fixed_controller | 1.0+/-0.0 | 1.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 73.0+/-0.0 | 12.0+/-0.0 | 5.62+/-1.01 | 3.87e-7+/-0.0e0 | ok 3/3
nonstiff-scalar-decay | Sparse | automatic_adams_bdf | 3/3 | adams | adams | switch_advantage_not_met | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 94.0+/-0.0 | 12.0+/-0.0 | 6.06+/-0.31 | 3.87e-7+/-0.0e0 | ok 3/3
nonstiff-system2-decay | Banded | automatic_adams_bdf | 3/3 | adams | adams | switch_advantage_not_met | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 112.0+/-0.0 | 16.0+/-0.0 | 5.20+/-0.23 | 1.44e-6+/-0.0e0 | ok 3/3
```
Analysis:
The 12-core release run confirms the intended control-plane behavior. Fixed
Adams is not silently falling back to BDF: `executed_adams > 0` and
`executed_bdf = 0` for both Sparse and Banded, on both scalar and two-equation
non-stiff problems. Automatic Adams/BDF also stays on Adams for this corpus and
reports `switch_advantage_not_met`, which is exactly the expected result for a
small non-stiff workload where BDF offers no advantage.
Correctness is stable: scalar decay lands at `~3.9e-7`, and the two-equation
decay at `~1.4e-6`, comfortably inside the story acceptance tolerance. On this
small non-stiff corpus Sparse is faster than Banded in wall-clock time, which is
not surprising: the linear systems are tiny, so Banded's structural advantage is
not yet large enough to dominate fixed overhead.
### `lsode2_symbolic_vs_numerical_closure_sparse_banded_dashboard`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: the production LSODE2 API should not require symbolic equations
when the user already has numerical Rust closures. A pure numerical route with
user residual/Jacobian closures, and the same route with an FD Jacobian, should
match the symbolic `AtomView + Lambdify` baseline on both Sparse and Banded
native linear algebra paths.
Command:
```powershell
cargo test --release lsode2_symbolic_vs_numerical_closure_sparse_banded_dashboard -- --nocapture --test-threads=1
```
Release result, CPU 12 Core:
```text
[LSODE2 story] symbolic Lambdify vs pure numerical closure routes; all time columns are milliseconds
Sparse | Lambdify-AtomView | 3/3 | 0.805+/-0.043 [0.753,0.857] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.029+/-0.000 | 0.013+/-0.001 | 0.107+/-0.015 | ok 3/3
Sparse | Numerical-AnalyticalJac | 3/3 | 0.805+/-0.145 [0.688,1.009] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.019+/-0.001 | 0.008+/-0.002 | 0.052+/-0.004 | ok 3/3
Sparse | Numerical-FDJac | 3/3 | 0.754+/-0.026 [0.718,0.776] | 7.41e-11+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.020+/-0.000 | 0.047+/-0.006 | 0.052+/-0.002 | ok 3/3
Banded | Lambdify-AtomView | 3/3 | 1.241+/-0.040 [1.185,1.277] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.029+/-0.002 | 0.017+/-0.000 | 0.013+/-0.001 | ok 3/3
Banded | Numerical-AnalyticalJac | 3/3 | 0.914+/-0.073 [0.830,1.009] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.019+/-0.000 | 0.007+/-0.000 | 0.014+/-0.000 | ok 3/3
Banded | Numerical-FDJac | 3/3 | 0.893+/-0.025 [0.860,0.920] | 1.00e-11+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.018+/-0.001 | 0.040+/-0.001 | 0.014+/-0.001 | ok 3/3
test numerical::LSODE2::story_tests2::lsode2_symbolic_vs_numerical_closure_sparse_banded_dashboard ... ok
```
Debug verification:
```text
[LSODE2 story] symbolic Lambdify vs pure numerical closure routes; all time columns are milliseconds
Sparse | Lambdify-AtomView | 3/3 | 8.534+/-0.087 [8.449,8.653] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.185+/-0.007 | 0.066+/-0.002 | 3.155+/-0.040 | ok 3/3
Sparse | Numerical-AnalyticalJac | 3/3 | 8.268+/-0.108 [8.190,8.421] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.129+/-0.002 | 0.062+/-0.001 | 3.066+/-0.008 | ok 3/3
Sparse | Numerical-FDJac | 3/3 | 8.398+/-0.055 [8.336,8.471] | 7.41e-11+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.130+/-0.001 | 0.253+/-0.002 | 3.112+/-0.013 | ok 3/3
Banded | Lambdify-AtomView | 3/3 | 3.984+/-0.025 [3.955,4.015] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.155+/-0.001 | 0.048+/-0.002 | 0.050+/-0.001 | ok 3/3
Banded | Numerical-AnalyticalJac | 3/3 | 3.998+/-0.076 [3.943,4.106] | 0.00e0+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.114+/-0.001 | 0.043+/-0.001 | 0.045+/-0.000 | ok 3/3
Banded | Numerical-FDJac | 3/3 | 4.189+/-0.039 [4.144,4.239] | 1.00e-11+/-0.0e0 | 272.0+/-0.0 | 194.0+/-0.0 | 263.0+/-0.0 | 0.116+/-0.001 | 0.252+/-0.006 | 0.048+/-0.002 | ok 3/3
```
Analysis:
The 12-core release run closes the API-surface gap: symbolic
`AtomView + Lambdify`, pure numerical analytical closures, and pure numerical
FD Jacobian closures all produce the same final state on Sparse and Banded
native paths. The counters line up exactly (`272` residual calls, `194`
Jacobian calls, `263` linear solves), so the comparison is about backend
plumbing and Jacobian construction, not about a different integration
trajectory.
The expected FD signature is visible in release: `Numerical-FDJac` raises
`jacobian_ms` from roughly `0.007-0.013` ms to roughly `0.040-0.047` ms, while
the residual and linear solve counts stay unchanged. That is the desired safe
fallback profile: correctness parity first, with a transparent cost paid in
Jacobian construction.
For this tiny 2D IVP, wall-clock differences are mostly fixed overhead and
noise. The meaningful conclusion is therefore qualitative: pure numerical
analytical and FD closure routes are production-valid API paths on both Sparse
and Banded native solvers; performance ranking should be judged on larger
workloads.
### `lsode2_stiff_switch_acceptance_sparse_banded_executes_bdf`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: a deliberately stiff relaxation problem, run under automatic
Adams/BDF selection, must execute BDF at least once on both native Sparse and
Banded paths. Unlike the exploratory dashboard above, this is an acceptance
gate: it fails if `executed_bdf` remains zero.
Command:
```powershell
cargo test --release lsode2_stiff_switch_acceptance_sparse_banded_executes_bdf -- --nocapture --test-threads=1
```
Debug verification after adding the gate:
```text
test numerical::LSODE2::story_tests2::lsode2_stiff_switch_acceptance_sparse_banded_executes_bdf ... [LSODE2 story] stiff-switch acceptance: automatic controller must execute BDF
Sparse | 3/3 | 1.00+/-0.00 | 1.00+/-0.00 | 116.00+/-0.00 | 25.00+/-0.00 | 1.57+/-0.43 | 3.222e-9+/-0.0e0 | ok 3/3
Banded | 3/3 | 1.00+/-0.00 | 1.00+/-0.00 | 113.00+/-0.00 | 23.00+/-0.00 | 8.03+/-0.77 | 3.222e-9+/-0.0e0 | ok 3/3
ok
```
Release result:
Analysis:
The debug verification confirms that the acceptance construction covers the
missing behavioral question: this is a real automatic run and both matrix
backends execute BDF. Paste the release table above and compare Sparse/Banded
wall time only after the acceptance condition remains green.
### `lsode2_mixed_regime_ramp_auto_switch_diagnostic_story`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: a single IVP that starts non-stiff and becomes stiff should reveal
whether native `automatic_adams_bdf` can re-evaluate the method family during a
full solve. The fixture uses numerical residual/Jacobian closures with exact
solution `y=cos(t)` and a smooth stiffness ramp, so the diagnostic isolates the
controller/native-solve choreography rather than symbolic parsing or codegen.
Command:
```powershell
cargo test lsode2_mixed_regime_ramp_auto_switch_diagnostic_story -- --nocapture
```
Debug diagnostic result:
```text
[LSODE2 story] mixed-regime ramp: one IVP starts Adams-capable and becomes stiff
Sparse | 3/3 | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 202.0+/-0.0 | 0.0+/-0.0 | 8.68+/-1.14 | 4.60e-1+/-0.0e0 | adams | switch_advantage_not_met | adams_only_current_limit | ok 3/3
Banded | 3/3 | 2.0+/-0.0 | 2.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0 | 202.0+/-0.0 | 0.0+/-0.0 | 6.94+/-0.06 | 4.60e-1+/-0.0e0 | adams | switch_advantage_not_met | adams_only_current_limit | ok 3/3
```
Analysis:
This is intentionally a diagnostic story, not an acceptance gate. It uncovered
a real current limitation: native `NativeSolve` performs bounded startup
probing and then runs the full integration with the selected method family. It
does not yet re-evaluate Adams/BDF selection after the trajectory enters a
stiffer regime. On this ramp fixture both Sparse and Banded remain Adams-only
and the final drift is large (`~4.6e-1`), even though the earlier stiff
acceptance test proves that the solver can execute BDF when the early decision
selects it.
The next engineering step is not another fixture tweak; it is mid-run
Adams/BDF re-evaluation in the native solve loop, with safe method-state
handoff. Once that exists, this diagnostic should be promoted to an acceptance
story requiring both `executed_adams > 0`, `executed_bdf > 0`, and small final
drift.
### `lsode2_combustion_like_multi_run_story_dashboard`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: on the combustion-like stiff IVP, Sparse and Banded Lambdify/AOT routes
remain numerically consistent while internal statistics expose preparation, solve,
residual, Jacobian and linear-system costs.
Command:
```powershell
cargo test --release lsode2_combustion_like_multi_run_story_dashboard -- --nocapture --test-threads=1
```
Result:
CPU 4 Core
test numerical::LSODE2::story_tests2::lsode2_combustion_like_multi_run_story_dashboard ... [LSODE2 story] combustion-like backend summary (multi-run); all time columns are milliseconds
Sparse | Lambdify | 5/5 | 8.83+/-1.54 [7.66,11.87] | 0.00e0+/-0.0e0 [0.00e0,0.00e0] | ok 5/5
Sparse | AOT-Ctcc | 5/5 | 54.43+/-93.76 [7.32,241.96] | 8.93e-12+/-0.0e0 [8.93e-12,8.93e-12] | ok 5/5
Banded | Lambdify | 5/5 | 15.46+/-1.84 [13.68,18.78] | 3.17e-12+/-0.0e0 [3.17e-12,3.17e-12] | ok 5/5
Banded | AOT-Ctcc | 5/5 | 10.80+/-0.40 [10.12,11.20] | 3.17e-12+/-0.0e0 [3.17e-12,3.17e-12] | ok 5/5
[LSODE2 story] combustion-like diagnostics (multi-run); prepare/solve are stage times, counters are counts
matrix | route | prepare_ms mean+/-std | solve_ms mean+/-std | residual_calls mean+/-std | jacobian_calls mean+/-std | linear_calls mean+/-std | accepted mean+/-std | rejected mean+/-std | preferred_bdf mean+/-std | executed_bdf mean+/-std
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Banded | Lambdify | 0.56+/-0.24 | 14.87+/-1.63 | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | AOT-Ctcc | 0.46+/-0.01 | 10.30+/-0.39 | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
[LSODE2 story] combustion-like stage timers (multi-run); all time columns are milliseconds
matrix | route | residual_ms mean+/-std | jacobian_ms mean+/-std | linear_ms mean+/-std
-----------------------------------------------------------------------------------------------
Banded | Lambdify | 0.434+/-0.027 | 0.614+/-0.205 | 0.161+/-0.023
Banded | AOT-Ctcc | 0.422+/-0.017 | 0.264+/-0.014 | 0.127+/-0.009
ok
CPU 12 Core
test numerical::LSODE2::story_tests2::lsode2_combustion_like_multi_run_story_dashboard ... [LSODE2 story] combustion-like backend summary (multi-run); all time columns are milliseconds
matrix | route | ok/runs | total_ms mean+/-std [min,max] | final_diff(A) mean+/-std [min,max] | status
-----------------------------------------------------------------------------------------------------------
Banded | Lambdify | 5/5 | 3.76+/-0.82 [2.95,5.18] | 3.17e-12+/-0.0e0 [3.17e-12,3.17e-12] | ok 5/5
Banded | AOT-Ctcc | 5/5 | 3.02+/-0.18 [2.69,3.18] | 3.17e-12+/-0.0e0 [3.17e-12,3.17e-12] | ok 5/5
[LSODE2 story] combustion-like diagnostics (multi-run); prepare/solve are stage times, counters are counts
matrix | route | prepare_ms mean+/-std | solve_ms mean+/-std | residual_calls mean+/-std | jacobian_calls mean+/-std | linear_calls mean+/-std | accepted mean+/-std | rejected mean+/-std | preferred_bdf mean+/-std | executed_bdf mean+/-std
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Banded | Lambdify | 0.09+/-0.00 | 3.66+/-0.82 | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | AOT-Ctcc | 0.11+/-0.00 | 2.91+/-0.18 | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
[LSODE2 story] combustion-like stage timers (multi-run); all time columns are milliseconds
matrix | route | residual_ms mean+/-std | jacobian_ms mean+/-std | linear_ms mean+/-std
-----------------------------------------------------------------------------------------------
Banded | Lambdify | 0.130+/-0.002 | 0.126+/-0.004 | 0.048+/-0.001
Banded | AOT-Ctcc | 0.128+/-0.001 | 0.073+/-0.001 | 0.048+/-0.001
ok
Analysis:
The routes agree numerically: final differences remain near `1e-11`, and all
four variants report identical residual, Jacobian, linear-solve, accepted-step
and rejected-step counts. `AOT-Ctcc` reduces the hot Jacobian cost on both
matrix layouts; for Banded it also reduces measured solve time from `14.87`
to `10.30` ms. Banded linear work is substantially cheaper than Sparse
(`0.127` to `0.161` ms versus `0.433` to `0.454` ms), as expected for this
structure.
The Sparse AOT `total_ms` mean cannot be interpreted as ordinary runtime
throughput: `prepare_ms` contains a cold/bootstrap-sized outlier, while its
solve phase is stable and slightly faster than Lambdify. This motivates an
explicit separation between cold artifact lifecycle measurements and warm
numerical throughput. It also matters that this combustion-like fixture stays
on Adams (`preferred_bdf = executed_bdf = 0`), so it does not cover stiff
method switching.
### `lsode2_combustion_like_parallel_chunking_multi_run_story_dashboard`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: after one AOT warm-up, runtime chunking can be assessed without confusing
it with cold compiler cost; correctness must not change between whole and parallel
callbacks.
Command:
```powershell
cargo test --release lsode2_combustion_like_parallel_chunking_multi_run_story_dashboard -- --nocapture --test-threads=1
```
Result:
CPU 4 Core
test numerical::LSODE2::story_tests2::lsode2_combustion_like_parallel_chunking_multi_run_story_dashboard ... [LSODE2 story] combustion-like parallel chunking summary (multi-run); all time columns are milliseconds
Sparse | Lambdify | baseline(no_chunk_knobs) | 5/5 | 8.75+/-1.13 [7.97,10.99] | 8.26+/-1.12 | 0.00e0+/-0.0e0 | ok 5/5
Sparse | AOT-Ctcc-Whole | whole | 5/5 | 8.33+/-0.75 [7.35,9.53] | 7.75+/-0.75 | 8.93e-12+/-0.0e0 | ok 5/5
Sparse | AOT-Ctcc-Parallel | parallel(auto,x2) | 5/5 | 8.87+/-0.91 [7.58,10.01] | 8.38+/-0.90 | 8.93e-12+/-0.0e0 | ok 5/5
Banded | Lambdify | baseline(no_chunk_knobs) | 5/5 | 14.36+/-1.66 [12.65,17.09] | 13.96+/-1.65 | 3.17e-12+/-0.0e0 | ok 5/5
Banded | AOT-Ctcc-Whole | whole | 5/5 | 10.46+/-0.89 [9.56,11.59] | 9.98+/-0.88 | 3.17e-12+/-0.0e0 | ok 5/5
Banded | AOT-Ctcc-Parallel | parallel(auto,x2) | 5/5 | 10.53+/-0.32 [10.01,10.88] | 9.97+/-0.33 | 3.17e-12+/-0.0e0 | ok 5/5
[LSODE2 story] combustion-like parallel chunking diagnostics (multi-run); counters are counts
matrix | route | chunking | residual_calls | jacobian_calls | linear_calls | accepted | rejected | preferred_bdf | executed_bdf
-------------------------------------------------------------------------------------------------------------------------------------------------------------
Sparse | AOT-Ctcc-Parallel | parallel(auto,x2) | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | Lambdify | baseline(no_chunk_knobs) | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | AOT-Ctcc-Whole | whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | AOT-Ctcc-Parallel | parallel(auto,x2) | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
ok
CPU 12 Core
running 1 test
test numerical::LSODE2::story_tests2::lsode2_combustion_like_parallel_chunking_multi_run_story_dashboard ... [LSODE2 story] combustion-like parallel chunking summary (multi-run); all time columns are milliseconds
Sparse | Lambdify | baseline(no_chunk_knobs) | 5/5 | 2.70+/-0.08 [2.60,2.83] | 2.58+/-0.08 | 0.00e0+/-0.0e0 | ok 5/5
Sparse | AOT-Ctcc-Whole | whole | 5/5 | 2.62+/-0.21 [2.47,3.03] | 2.50+/-0.21 | 8.93e-12+/-0.0e0 | ok 5/5
Sparse | AOT-Ctcc-Parallel | parallel(auto,x2) | 5/5 | 2.63+/-0.14 [2.53,2.90] | 2.51+/-0.14 | 8.93e-12+/-0.0e0 | ok 5/5
Banded | Lambdify | baseline(no_chunk_knobs) | 5/5 | 3.64+/-0.34 [3.11,4.16] | 3.52+/-0.33 | 3.17e-12+/-0.0e0 | ok 5/5
Banded | AOT-Ctcc-Whole | whole | 5/5 | 2.95+/-0.28 [2.55,3.25] | 2.84+/-0.28 | 3.17e-12+/-0.0e0 | ok 5/5
Banded | AOT-Ctcc-Parallel | parallel(auto,x2) | 5/5 | 2.84+/-0.04 [2.80,2.89] | 2.73+/-0.04 | 3.17e-12+/-0.0e0 | ok 5/5
[LSODE2 story] combustion-like parallel chunking diagnostics (multi-run); counters are counts
matrix | route | chunking | residual_calls | jacobian_calls | linear_calls | accepted | rejected | preferred_bdf | executed_bdf
-------------------------------------------------------------------------------------------------------------------------------------------------------------
Sparse | AOT-Ctcc-Parallel | parallel(auto,x2) | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | Lambdify | baseline(no_chunk_knobs) | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | AOT-Ctcc-Whole | whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
Banded | AOT-Ctcc-Parallel | parallel(auto,x2) | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0 | 0.0+/-0.0 | 0.0+/-0.0
ok
Analysis:
Chunking preserves correctness and controller behavior: every variant has the
same final-state difference and the same numerical-work counters. It does not
provide a strong throughput win at this workload size. On the older 4-core
machine Sparse AOT whole was faster than its parallel variant (`7.75` versus
`8.38` ms solve time), while Banded whole and parallel were effectively tied
(`9.98` versus `9.97` ms). On the 12-core machine the fixed overhead is much
smaller and the gap narrows: Sparse whole/parallel are essentially identical
(`2.50` versus `2.51` ms), and Banded parallel becomes slightly faster than
whole (`2.73` versus `2.84` ms), but the improvement is still small.
For this three-state combustion problem, whole generation remains the
practical default unless the machine/workload measurement says otherwise.
The 12-core result is encouraging because parallel chunking no longer hurts,
but it still does not prove that chunking is profitable for small LSODE2
systems. A recommendation for parallel AOT must come from a larger-dimensional
workload with diagnostics for actual jobs, fallback decisions and work per job,
rather than from requesting chunking alone.
## New Symbolic Frontend Matrix
### `lsode2_combustion_symbolic_frontend_sparse_banded_multi_run_dashboard`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: `ExprLegacy` and `AtomView` Lambdify frontends produce equivalent
combustion solutions for both Sparse and Banded linear algebra, while the wall-clock
and preparation columns show whether `AtomView` is the preferable symbolic baseline.
The output is deliberately split into correctness/wall-clock, numerical counters and
hot-stage timer tables so each table stays readable.
Command:
```powershell
cargo test --release lsode2_combustion_symbolic_frontend_sparse_banded_multi_run_dashboard -- --ignored --nocapture --test-threads=1
```
Result:
CPU 4 Core
correctness/wall-clock; all time columns are milliseconds
Sparse | Lambdify-ExprLegacy | 5/5 | 8.93+/-0.74 [7.83,9.84] | 0.69+/-0.35 | 8.21+/-0.80 | 0.00e0+/-0.0e0 | ok 5/5
Sparse | Lambdify-AtomView | 5/5 | 9.67+/-0.45 [9.20,10.51] | 0.54+/-0.16 | 9.10+/-0.32 | 0.00e0+/-0.0e0 | ok 5/5
Banded | Lambdify-ExprLegacy | 5/5 | 12.23+/-3.33 [9.50,18.37] | 0.45+/-0.01 | 11.75+/-3.33 | 0.00e0+/-0.0e0 | ok 5/5
Banded | Lambdify-AtomView | 5/5 | 9.42+/-0.38 [8.91,9.96] | 0.46+/-0.05 | 8.94+/-0.38 | 0.00e0+/-0.0e0 | ok 5/5
[LSODE2 story] combustion symbolic frontend Sparse/Banded (Lambdify) numerical work; counters are counts (mean+/-std)
matrix | route | residual_calls | jacobian_calls | linear_calls | accepted | rejected
---------------------------------------------------------------------------------------------------------
Banded | Lambdify-ExprLegacy | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | Lambdify-AtomView | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
[LSODE2 story] combustion symbolic frontend Sparse/Banded (Lambdify) hot-stage timers; all time columns are milliseconds
matrix | route | residual_ms mean+/-std | jacobian_ms mean+/-std | linear_ms mean+/-std
----------------------------------------------------------------------------------------------------------------
Banded | Lambdify-ExprLegacy | 0.486+/-0.026 | 0.456+/-0.034 | 0.128+/-0.005
Banded | Lambdify-AtomView | 0.461+/-0.005 | 0.426+/-0.008 | 0.120+/-0.005
ok
CPU 12 Core
test numerical::LSODE2::story_tests2::lsode2_combustion_symbolic_frontend_sparse_banded_multi_run_dashboard ... [LSODE2 story] combustion symbolic frontend Sparse/Banded (Lambdify) correctness/wall-clock; all time columns are milliseconds
matrix | route | ok/runs | total_ms mean+/-std [min,max] | prepare_ms mean+/-std | solve_ms mean+/-std | final_diff mean+/-std | status
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Banded | Lambdify-ExprLegacy | 5/5 | 4.01+/-0.74 [2.96,4.73] | 0.10+/-0.00 | 3.91+/-0.74 | 0.00e0+/-0.0e0 | ok 5/5
Banded | Lambdify-AtomView | 5/5 | 3.10+/-0.23 [2.91,3.39] | 0.10+/-0.00 | 3.00+/-0.23 | 0.00e0+/-0.0e0 | ok 5/5
[LSODE2 story] combustion symbolic frontend Sparse/Banded (Lambdify) numerical work; counters are counts (mean+/-std)
matrix | route | residual_calls | jacobian_calls | linear_calls | accepted | rejected
---------------------------------------------------------------------------------------------------------
Banded | Lambdify-ExprLegacy | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | Lambdify-AtomView | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
[LSODE2 story] combustion symbolic frontend Sparse/Banded (Lambdify) hot-stage timers; all time columns are milliseconds
matrix | route | residual_ms mean+/-std | jacobian_ms mean+/-std | linear_ms mean+/-std
----------------------------------------------------------------------------------------------------------------
Banded | Lambdify-ExprLegacy | 0.168+/-0.002 | 0.168+/-0.080 | 0.046+/-0.001
Banded | Lambdify-AtomView | 0.162+/-0.003 | 0.122+/-0.001 | 0.048+/-0.003
ok
Analysis:
Frontend parity is established for this scenario: `ExprLegacy` and `AtomView`
produce matching final values within each matrix route and identical
integration counters. `AtomView` reduces measured hot callback cost in both
structures, most visibly for Jacobian evaluation (`0.516` to `0.438` ms for
Sparse and `0.456` to `0.426` ms for Banded).
The Banded `AtomView` route also improves total wall time relative to
`ExprLegacy` (`9.42` versus `12.23` ms), whereas Sparse total wall time is
slightly worse despite cheaper callbacks. That Sparse discrepancy is small
enough to treat as overhead or run noise at this scale, not as a correctness
concern. The strongest structural signal remains `linear_ms`: Banded needs
about `0.12` ms versus about `0.45` ms for Sparse.
## New Cold AOT Toolchain And Chunking Matrix
### `lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix`
File: `src/numerical/LSODE2/story_tests2.rs`
Hypothesis: on one real stiff workload, the `AtomView` Lambdify baseline and cold
AOT paths remain numerically equivalent; `Sparse/Banded`, `tcc/gcc/zig/rust`, and
`whole/parallel` differences can then be interpreted using true end-to-end wall
clock separately from hot residual/Jacobian/linear timings.
Each AOT sample now receives a fresh artifact directory and uses
`SymbolicIvpAotBuildPolicy::RebuildAlways { profile: Release }`. The test prints a
separate lifecycle table whose AOT rows must report `cold_action=rebuild_always`
and `artifact_dir_written=true`. This prevents `BuildIfMissing` from silently
resolving an earlier problem-keyed runtime backend.
Diagnostic note: the first 12-core run exposed a useful failure mode in the
story harness itself. Failed AOT rows were previously represented only by empty
numeric columns because `new/prepare/solve` errors were collapsed into `None`.
The harness now records `first_failure=...` in the main status column and writes
a lifecycle row for both successful and failed cold AOT attempts. If a future
table has blank timing columns, the status/lifecycle table should explain
whether the failure happened during solver construction, AOT build/link,
prepare, or solve.
Command:
```powershell
cargo test --release lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix -- --ignored --nocapture --test-threads=1
```
CPU 4 Core
Historical result before the `RebuildAlways` correction:
test numerical::LSODE2::story_tests2::lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix ... [LSODE2 story] combustion AtomView cold AOT toolchain/chunking Sparse/Banded matrix correctness/wall-clock; all time columns are milliseconds
Sparse | Lambdify-AtomView | 3/3 | 16.74+/-11.93 [8.14,33.62] | 0.39+/-0.02 | 16.33+/-11.96 | 0.00e0+/-0.0e0 | ok 3/3
Banded | Lambdify-AtomView | 3/3 | 10.22+/-1.05 [9.08,11.62] | 0.36+/-0.01 | 9.84+/-1.04 | 0.00e0+/-0.0e0 | ok 3/3
Sparse | tcc/whole | 3/3 | 119.70+/-157.60 [7.62,342.57] | 112.09+/-157.72 | 7.59+/-0.52 | 1.24e-12+/-0.0e0 | ok 3/3
Sparse | tcc/parallel | 3/3 | 112.56+/-144.38 [8.69,316.74] | 103.35+/-145.21 | 9.20+/-1.65 | 1.24e-12+/-0.0e0 | ok 3/3
Sparse | gcc/whole | 3/3 | 8.69+/-0.18 [8.55,8.94] | 0.65+/-0.07 | 8.03+/-0.11 | 1.24e-12+/-0.0e0 | ok 3/3
Sparse | gcc/parallel | 3/3 | 8.81+/-0.11 [8.68,8.95] | 0.69+/-0.05 | 8.10+/-0.14 | 1.24e-12+/-0.0e0 | ok 3/3
Sparse | zig/whole | 3/3 | 8.39+/-0.15 [8.21,8.57] | 0.68+/-0.04 | 7.70+/-0.17 | 1.24e-12+/-0.0e0 | ok 3/3
Sparse | zig/parallel | 3/3 | 8.55+/-0.13 [8.39,8.71] | 0.77+/-0.13 | 7.76+/-0.14 | 1.24e-12+/-0.0e0 | ok 3/3
Sparse | rust/whole | 3/3 | 8.23+/-0.17 [8.01,8.42] | 0.61+/-0.04 | 7.60+/-0.14 | 1.24e-12+/-0.0e0 | ok 3/3
Sparse | rust/parallel | 3/3 | 8.82+/-0.06 [8.75,8.90] | 0.74+/-0.07 | 8.07+/-0.11 | 1.24e-12+/-0.0e0 | ok 3/3
Banded | tcc/whole | 3/3 | 9.98+/-0.54 [9.24,10.51] | 0.58+/-0.01 | 9.39+/-0.55 | 1.93e-11+/-0.0e0 | ok 3/3
Banded | tcc/parallel | 3/3 | 11.11+/-1.69 [9.70,13.48] | 0.63+/-0.03 | 10.47+/-1.66 | 1.93e-11+/-0.0e0 | ok 3/3
Banded | gcc/whole | 3/3 | 9.84+/-0.52 [9.26,10.53] | 0.64+/-0.01 | 9.18+/-0.54 | 1.93e-11+/-0.0e0 | ok 3/3
Banded | gcc/parallel | 3/3 | 15.40+/-8.07 [9.59,26.82] | 6.40+/-8.17 | 8.99+/-0.11 | 1.93e-11+/-0.0e0 | ok 3/3
Banded | zig/whole | 3/3 | 9.38+/-0.17 [9.16,9.58] | 0.58+/-0.04 | 8.79+/-0.21 | 1.93e-11+/-0.0e0 | ok 3/3
Banded | zig/parallel | 3/3 | 9.87+/-0.08 [9.79,9.98] | 0.59+/-0.02 | 9.26+/-0.08 | 1.93e-11+/-0.0e0 | ok 3/3
Banded | rust/whole | 3/3 | 9.25+/-0.06 [9.20,9.34] | 0.56+/-0.01 | 8.68+/-0.05 | 1.93e-11+/-0.0e0 | ok 3/3
Banded | rust/parallel | 3/3 | 9.82+/-0.16 [9.60,9.97] | 0.62+/-0.05 | 9.19+/-0.20 | 1.93e-11+/-0.0e0 | ok 3/3
[LSODE2 story] combustion AtomView cold AOT toolchain/chunking Sparse/Banded matrix numerical work; counters are counts (mean+/-std)
matrix | route | residual_calls | jacobian_calls | linear_calls | accepted | rejected
---------------------------------------------------------------------------------------------------------
Sparse | tcc/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Sparse | tcc/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Sparse | gcc/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Sparse | gcc/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Sparse | zig/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Sparse | zig/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Sparse | rust/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Sparse | rust/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | tcc/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | tcc/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | gcc/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | gcc/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | zig/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | zig/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | rust/whole | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
Banded | rust/parallel | 776.0+/-0.0 | 387.0+/-0.0 | 774.0+/-0.0 | 363.0+/-0.0 | 24.0+/-0.0
[LSODE2 story] combustion AtomView cold AOT toolchain/chunking Sparse/Banded matrix hot-stage timers; all time columns are milliseconds
matrix | route | residual_ms mean+/-std | jacobian_ms mean+/-std | linear_ms mean+/-std
----------------------------------------------------------------------------------------------------------------
Sparse | tcc/whole | 0.389+/-0.053 | 0.288+/-0.039 | 0.481+/-0.029
Sparse | tcc/parallel | 0.446+/-0.005 | 0.476+/-0.008 | 0.441+/-0.019
Sparse | gcc/whole | 0.398+/-0.034 | 0.287+/-0.010 | 0.449+/-0.016
Sparse | gcc/parallel | 0.461+/-0.006 | 0.475+/-0.018 | 0.453+/-0.020
Sparse | zig/whole | 0.381+/-0.012 | 0.298+/-0.008 | 0.456+/-0.017
Sparse | zig/parallel | 0.442+/-0.010 | 0.467+/-0.014 | 0.443+/-0.013
Sparse | rust/whole | 0.379+/-0.016 | 0.291+/-0.007 | 0.453+/-0.033
Sparse | rust/parallel | 0.429+/-0.004 | 0.467+/-0.006 | 0.471+/-0.038
Banded | tcc/whole | 0.374+/-0.009 | 0.279+/-0.008 | 0.108+/-0.001
Banded | tcc/parallel | 0.516+/-0.099 | 0.535+/-0.086 | 0.132+/-0.024
Banded | gcc/whole | 0.379+/-0.003 | 0.281+/-0.007 | 0.112+/-0.005
Banded | gcc/parallel | 0.430+/-0.004 | 0.453+/-0.001 | 0.108+/-0.001
Banded | zig/whole | 0.368+/-0.002 | 0.297+/-0.012 | 0.111+/-0.001
Banded | zig/parallel | 0.437+/-0.003 | 0.457+/-0.006 | 0.112+/-0.004
Banded | rust/whole | 0.373+/-0.012 | 0.287+/-0.011 | 0.108+/-0.005
Banded | rust/parallel | 0.436+/-0.008 | 0.474+/-0.018 | 0.111+/-0.008
note: AOT total_ms includes symbolic preparation, artifact build/link and native integration; hot-stage timers isolate repeated callback/linear work.
ok
CPU 12 Core
test numerical::LSODE2::story_tests2::lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix ... [LSODE2 story] combustion AtomView cold AOT toolchain/chunking Sparse/Banded matrix correctness/wall-clock; all time columns are milliseconds
Sparse | Lambdify-AtomView | 3/3 | 2.63+/-0.10 [2.53,2.76] | 0.12+/-0.03 | 2.51+/-0.07 | 0.00e0+/-0.0e0 | ok 3/3
Banded | Lambdify-AtomView | 3/3 | 2.37+/-0.06 [2.32,2.45] | 0.09+/-0.00 | 2.27+/-0.06 | 0.00e0+/-0.0e0 | ok 3/3
Sparse | tcc/whole | 3/3 | 40.51+/-3.18 [37.63,44.94] | 0.00+/-0.00 | 40.50+/-3.18 | 2.22e-14+/-0.0e0 | ok 3/3
Sparse | tcc/parallel | 3/3 | 38.68+/-0.79 [37.79,39.70] | 0.00+/-0.00 | 38.67+/-0.79 | 2.22e-14+/-0.0e0 | ok 3/3
Sparse | gcc/whole | 3/3 | 297.86+/-3.08 [293.84,301.31] | 0.00+/-0.00 | 297.85+/-3.08 | 2.84e-14+/-0.0e0 | ok 3/3
Sparse | gcc/parallel | 3/3 | 317.66+/-3.36 [314.56,322.34] | 0.00+/-0.00 | 317.66+/-3.37 | 2.84e-14+/-0.0e0 | ok 3/3
Sparse | zig/whole | 3/3 | 16174.60+/-62.85 [16089.72,16239.89] | 0.00+/-0.00 | 16174.60+/-62.85 | 6.58e-15+/-0.0e0 | ok 3/3
Sparse | zig/parallel | 3/3 | 15860.42+/-69.23 [15781.77,15950.25] | 0.00+/-0.00 | 15860.42+/-69.23 | 6.58e-15+/-0.0e0 | ok 3/3
Sparse | rust/whole | 3/3 | 591.76+/-139.25 [493.00,788.70] | 0.00+/-0.00 | 591.75+/-139.25 | 2.84e-14+/-0.0e0 | ok 3/3
Sparse | rust/parallel | 3/3 | 721.95+/-97.87 [583.68,796.37] | 0.00+/-0.00 | 721.94+/-97.87 | 2.84e-14+/-0.0e0 | ok 3/3
Banded | tcc/whole | 3/3 | 47.00+/-1.25 [45.71,48.69] | 0.00+/-0.00 | 46.99+/-1.25 | 6.06e-15+/-0.0e0 | ok 3/3
Banded | tcc/parallel | 3/3 | 49.05+/-6.69 [43.84,58.50] | 0.00+/-0.00 | 49.04+/-6.69 | 6.06e-15+/-0.0e0 | ok 3/3
Banded | gcc/whole | 3/3 | 312.89+/-3.64 [309.05,317.78] | 0.00+/-0.00 | 312.88+/-3.64 | 2.58e-14+/-0.0e0 | ok 3/3
Banded | gcc/parallel | 3/3 | 347.31+/-10.29 [337.79,361.61] | 0.00+/-0.00 | 347.30+/-10.29 | 2.58e-14+/-0.0e0 | ok 3/3
Banded | zig/whole | 3/3 | 17029.09+/-681.66 [16525.22,17992.76] | 0.00+/-0.00 | 17029.08+/-681.66 | 2.17e-14+/-0.0e0 | ok 3/3
Banded | zig/parallel | 3/3 | 16370.93+/-19.61 [16343.41,16387.59] | 0.00+/-0.00 | 16370.92+/-19.61 | 2.17e-14+/-0.0e0 | ok 3/3
Banded | rust/whole | 3/3 | 484.80+/-7.30 [477.56,494.80] | 0.00+/-0.00 | 484.79+/-7.30 | 2.58e-14+/-0.0e0 | ok 3/3
Banded | rust/parallel | 3/3 | 589.27+/-58.32 [533.19,669.68] | 0.00+/-0.00 | 589.26+/-58.32 | 2.58e-14+/-0.0e0 | ok 3/3
[LSODE2 story] combustion AtomView cold AOT toolchain/chunking Sparse/Banded matrix numerical work; counters are counts (mean+/-std)
matrix | route | residual_calls | jacobian_calls | linear_calls | accepted | rejected
---------------------------------------------------------------------------------------------------------
Sparse | tcc/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Sparse | tcc/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Sparse | gcc/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Sparse | gcc/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Sparse | zig/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Sparse | zig/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Sparse | rust/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Sparse | rust/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | tcc/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | tcc/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | gcc/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | gcc/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | zig/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | zig/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | rust/whole | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
Banded | rust/parallel | 1087.0+/-0.0 | 574.0+/-0.0 | 1086.0+/-0.0 | 541.0+/-0.0 | 33.0+/-0.0
[LSODE2 story] combustion AtomView cold AOT toolchain/chunking Sparse/Banded matrix hot-stage timers; all time columns are milliseconds
matrix | route | residual_ms mean+/-std | jacobian_ms mean+/-std | linear_ms mean+/-std
----------------------------------------------------------------------------------------------------------------
Sparse | tcc/whole | 0.177+/-0.003 | 0.128+/-0.003 | 0.274+/-0.082
Sparse | tcc/parallel | 0.196+/-0.002 | 0.203+/-0.001 | 0.215+/-0.003
Sparse | gcc/whole | 0.155+/-0.001 | 0.088+/-0.001 | 0.210+/-0.006
Sparse | gcc/parallel | 0.165+/-0.002 | 0.111+/-0.002 | 0.208+/-0.002
Sparse | zig/whole | 0.172+/-0.006 | 0.110+/-0.006 | 0.217+/-0.007
Sparse | zig/parallel | 0.174+/-0.003 | 0.145+/-0.001 | 0.216+/-0.006
Sparse | rust/whole | 0.162+/-0.003 | 0.089+/-0.001 | 0.221+/-0.003
Sparse | rust/parallel | 0.289+/-0.007 | 0.168+/-0.005 | 0.360+/-0.006
Banded | tcc/whole | 0.277+/-0.058 | 0.188+/-0.039 | 0.095+/-0.018
Banded | tcc/parallel | 0.314+/-0.037 | 0.323+/-0.041 | 0.107+/-0.016
Banded | gcc/whole | 0.234+/-0.029 | 0.130+/-0.016 | 0.102+/-0.018
Banded | gcc/parallel | 0.278+/-0.040 | 0.174+/-0.019 | 0.106+/-0.007
Banded | zig/whole | 0.187+/-0.002 | 0.116+/-0.003 | 0.081+/-0.005
Banded | zig/parallel | 0.207+/-0.008 | 0.169+/-0.008 | 0.083+/-0.006
Banded | rust/whole | 0.211+/-0.050 | 0.112+/-0.027 | 0.092+/-0.020
Banded | rust/parallel | 0.262+/-0.052 | 0.153+/-0.031 | 0.095+/-0.015
note: AOT total_ms includes symbolic preparation, artifact build/link and native integration; hot-stage timers isolate repeated callback/linear work.
[LSODE2 story] cold AOT lifecycle observations; successful AOT rows require a fresh materialization directory
matrix | route | rep | cold_action | artifact_dir_written | status
---------------------------------------------------------------------------------------------
Sparse | tcc/whole | 3 | rebuild_always | true | ok
Sparse | tcc/parallel | 1 | rebuild_always | true | ok
Sparse | tcc/parallel | 2 | rebuild_always | true | ok
Sparse | tcc/parallel | 3 | rebuild_always | true | ok
Sparse | gcc/whole | 1 | rebuild_always | true | ok
Sparse | gcc/whole | 2 | rebuild_always | true | ok
Sparse | gcc/whole | 3 | rebuild_always | true | ok
Sparse | gcc/parallel | 1 | rebuild_always | true | ok
Sparse | gcc/parallel | 2 | rebuild_always | true | ok
Sparse | gcc/parallel | 3 | rebuild_always | true | ok
Sparse | zig/whole | 1 | rebuild_always | true | ok
Sparse | zig/whole | 2 | rebuild_always | true | ok
Sparse | zig/whole | 3 | rebuild_always | true | ok
Sparse | zig/parallel | 1 | rebuild_always | true | ok
Sparse | zig/parallel | 2 | rebuild_always | true | ok
Sparse | zig/parallel | 3 | rebuild_always | true | ok
Sparse | rust/whole | 1 | rebuild_always | true | ok
Sparse | rust/whole | 2 | rebuild_always | true | ok
Sparse | rust/whole | 3 | rebuild_always | true | ok
Sparse | rust/parallel | 1 | rebuild_always | true | ok
Sparse | rust/parallel | 2 | rebuild_always | true | ok
Sparse | rust/parallel | 3 | rebuild_always | true | ok
Banded | tcc/whole | 1 | rebuild_always | true | ok
Banded | tcc/whole | 2 | rebuild_always | true | ok
Banded | tcc/whole | 3 | rebuild_always | true | ok
Banded | tcc/parallel | 1 | rebuild_always | true | ok
Banded | tcc/parallel | 2 | rebuild_always | true | ok
Banded | tcc/parallel | 3 | rebuild_always | true | ok
Banded | gcc/whole | 1 | rebuild_always | true | ok
Banded | gcc/whole | 2 | rebuild_always | true | ok
Banded | gcc/whole | 3 | rebuild_always | true | ok
Banded | gcc/parallel | 1 | rebuild_always | true | ok
Banded | gcc/parallel | 2 | rebuild_always | true | ok
Banded | gcc/parallel | 3 | rebuild_always | true | ok
Banded | zig/whole | 1 | rebuild_always | true | ok
Banded | zig/whole | 2 | rebuild_always | true | ok
Banded | zig/whole | 3 | rebuild_always | true | ok
Banded | zig/parallel | 1 | rebuild_always | true | ok
Banded | zig/parallel | 2 | rebuild_always | true | ok
Banded | zig/parallel | 3 | rebuild_always | true | ok
Banded | rust/whole | 1 | rebuild_always | true | ok
Banded | rust/whole | 2 | rebuild_always | true | ok
Banded | rust/whole | 3 | rebuild_always | true | ok
Banded | rust/parallel | 1 | rebuild_always | true | ok
Banded | rust/parallel | 2 | rebuild_always | true | ok
Banded | rust/parallel | 3 | rebuild_always | true | ok
ok
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 2317 filtered out; finished in 207.86s
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 2317 filtered out; finished in 87.24
Analysis:
The historical 4-core table is useful, but only as a hot-runtime comparison.
It showed numerical parity across toolchains and chunking choices, yet it also
reported suspiciously tiny non-TCC `prepare_ms` values for routes that were
supposed to perform a fresh external compile and link. That was the clue that
the original harness was not a reliable cold-build ranking.
The harness was later tightened in three ways: AOT rows now force
`RebuildAlways`, every row uses an isolated materialization directory, and the
table prints lifecycle observations (`cold_action`, `artifact_dir_written`,
`status`). A second lifecycle issue was also fixed: the combustion cold-matrix
story now uses a BDF-only controller, because automatic Adams/BDF probing could
build and load an AOT DLL before the full solve tried to rebuild the same path
on Windows. That failure mode looked like a compiler/toolchain problem, but it
was really a test-lifecycle artifact.
The story harness supports targeted diagnostics:
```powershell
$env:LSODE2_AOT_COLD_FILTER="Sparse tcc/whole"
$env:LSODE2_AOT_COLD_REPEATS="1"
cargo test lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix -- --ignored --nocapture
```
A filtered debug-harness check of `Sparse tcc/whole` is green after this change:
`ok 1/1`, `total_ms ~= 76.85`, `final_diff ~= 2.22e-14`, and the lifecycle row
reports `artifact_dir_written=true | ok`.
The next 12-core rerun localized a Zig-only materialization bug. `tcc`, `gcc`,
and `rust` completed on both Sparse and Banded paths, while every `zig` row
failed with `failed to spawn build runner ...`. That was not a solver syntax
problem: the same LSODE2 problem, matrix choices and AOT lifecycle were green
for the other toolchains. The root cause was the Zig build materializer carrying
relative paths into `zig build`: workdir, `zig-out`, `ZIG_LOCAL_CACHE_DIR`, and
`ZIG_GLOBAL_CACHE_DIR` were not normalized the way the C materializer already
normalizes artifact paths. After moving the repository to `D:\...`, Zig could
resolve those relative paths against the generated crate workdir and fail to
spawn its build runner. The Zig AOT build request now normalizes build workdir,
expected artifact path and cache directories to absolute non-verbatim paths.
Debug verification after the Zig path-normalization fix:
```powershell
$env:LSODE2_AOT_COLD_FILTER="Sparse zig/whole"
$env:LSODE2_AOT_COLD_REPEATS="1"
cargo test lsode2_combustion_aot_toolchain_chunking_sparse_banded_cold_matrix -- --ignored --nocapture
```
Result: `Sparse | zig/whole | ok 1/1`, `artifact_dir_written=true`, and
`final_diff ~= 6.58e-15`.
The current 12-core release table is the source of truth. All cold AOT rows are
green (`ok 3/3`), every row writes its fresh artifact directory, and Sparse and
Banded counters are identical across Lambdify and every AOT toolchain. The
numeric differences are all at roundoff scale, so the correctness/lifecycle
question for this cold matrix is closed.
The cold-build ranking is clear on this machine. `tcc` is the only genuinely
fast cold AOT route here (`~39-49` ms), `gcc` is moderate (`~298-347` ms), Rust
AOT is heavier (`~485-722` ms), and Zig is correct but currently far too slow
for cold LSODE2 startup (`~15.8-17.0` s). This does not make Zig invalid; it
means Zig is not the practical cold-build recommendation for this LSODE2
fixture.
Chunking is also not a win for this combustion fixture. The hot-stage timers
show that `whole` callbacks are usually faster than explicit `parallel`
callbacks, especially for Jacobian evaluation. The problem is simply too small
for AOT callback chunking overhead to amortize. Banded still does what it should:
its linear stage is consistently much cheaper than Sparse, but the total
cold-AOT wall time is dominated by toolchain/build and callback overhead rather
than the linear solve.
One important table caveat: for `RebuildAlways` AOT rows, `prepare_ms` is shown
as zero because the harness intentionally skips eager `prepare()` and lets
`solve_with_summary` perform the cold build, link and solve in one path. In this
story, `solve_ms` and `total_ms` are therefore the honest cold wall-clock
numbers for AOT rows.
### `lsode2_combustion_sparse_banded_atomview_tcc_build_then_require_prebuilt_story`
Hypothesis: the production AOT lifecycle should work as a two-stage workflow.
The first `BuildIfMissing` run is allowed to build/link a `tcc` artifact; later
`RequirePrebuilt` runs must reuse the already-linked artifact strictly, without
silently falling back to Lambdify or rebuilding. Correctness must stay at
roundoff scale for both Sparse and Banded LSODE2 routes.
This is not a broad toolchain benchmark. It intentionally narrows the question
to the practical route identified by the cold matrix: `AtomView + tcc`, with
Sparse and Banded linear algebra. The important columns are `build_policy`,
`prepare_ms`, `solve_ms`, `final_diff`, and the residual/Jacobian/linear
counters. `prepare_ms` should be visibly larger for the first build row and
small for strict prebuilt rows.
Paste-ready release command:
```powershell
cargo test --release lsode2_combustion_sparse_banded_atomview_tcc_build_then_require_prebuilt_story -- --ignored --nocapture --test-threads=1
```
test numerical::LSODE2::story_tests2::lsode2_combustion_sparse_banded_atomview_tcc_build_then_require_prebuilt_story ... [LSODE2 lifecycle] combustion AtomView tcc BuildIfMissing -> RequirePrebuilt lifecycle: correctness/backend policy
Sparse | build | BuildIfMissing | 2.224e-14 | ok
Sparse | prebuilt | RequirePrebuilt | 2.224e-14 | ok
Sparse | prebuilt | RequirePrebuilt | 2.224e-14 | ok
Sparse | prebuilt | RequirePrebuilt | 2.224e-14 | ok
Banded | build | BuildIfMissing | 6.056e-15 | ok
Banded | prebuilt | RequirePrebuilt | 6.056e-15 | ok
Banded | prebuilt | RequirePrebuilt | 6.056e-15 | ok
Banded | prebuilt | RequirePrebuilt | 6.056e-15 | ok
[LSODE2 lifecycle] combustion AtomView tcc BuildIfMissing -> RequirePrebuilt lifecycle: wall-clock and hot stages; milliseconds
matrix | phase | total_ms | prepare_ms | solve_ms | residual_ms | jacobian_ms | linear_ms
------------------------------------------------------------------------------------------------
Sparse | prebuilt | 3.435 | 0.301 | 3.128 | 0.208 | 0.142 | 0.250
Sparse | prebuilt | 3.795 | 0.371 | 3.419 | 0.230 | 0.172 | 0.276
Banded | build | 4.713 | 0.355 | 4.353 | 0.225 | 0.153 | 0.087
Banded | prebuilt | 4.530 | 0.308 | 4.217 | 0.187 | 0.120 | 0.066
Banded | prebuilt | 3.801 | 0.251 | 3.546 | 0.227 | 0.148 | 0.085
Banded | prebuilt | 4.631 | 0.288 | 4.338 | 0.209 | 0.137 | 0.075
[LSODE2 lifecycle] combustion AtomView tcc BuildIfMissing -> RequirePrebuilt lifecycle: numerical work; counters are counts
matrix | phase | residual_calls | jacobian_calls | linear_calls
------------------------------------------------------------------------
Sparse | prebuilt | 1087 | 574 | 1086
Sparse | prebuilt | 1087 | 574 | 1086
Banded | build | 1087 | 574 | 1086
Banded | prebuilt | 1087 | 574 | 1086
Banded | prebuilt | 1087 | 574 | 1086
Banded | prebuilt | 1087 | 574 | 1086
ok
Shorter smoke variant:
```powershell
$env:LSODE2_PREBUILT_REPEATS="1"
cargo test --release lsode2_combustion_sparse_banded_atomview_tcc_build_then_require_prebuilt_story -- --ignored --nocapture --test-threads=1
```
Debug smoke on this branch passed with `LSODE2_PREBUILT_REPEATS=1`: Sparse and
Banded `BuildIfMissing` rows solved, Sparse and Banded `RequirePrebuilt` rows
solved, and all `final_diff` values were at roundoff scale (`~1e-14`). The
strict prebuilt rows reported sub-2 ms `prepare_ms`, which is the key lifecycle
signal.
Result:
```text
Release 12-core run pasted above:
Sparse BuildIfMissing: ok, final_diff ~= 2.22e-14, total_ms ~= 105.5,
prepare_ms ~= 102.5.
Sparse RequirePrebuilt: ok 3/3, final_diff ~= 2.22e-14, total_ms ~= 3.4-3.8,
prepare_ms ~= 0.30-0.41.
Banded BuildIfMissing: ok, final_diff ~= 6.06e-15, total_ms ~= 4.7,
prepare_ms ~= 0.36.
Banded RequirePrebuilt: ok 3/3, final_diff ~= 6.06e-15, total_ms ~= 3.8-4.6,
prepare_ms ~= 0.25-0.31.
```
Analysis:
The same-process user-facing lifecycle is green. `BuildIfMissing` creates or
links the compiled `tcc` backend, and strict `RequirePrebuilt` then reuses it
without falling back to Lambdify or triggering a cold rebuild. The strongest
signal is not just correctness but the strict rows' `prepare_ms`: all prebuilt
rows are sub-millisecond to about 0.4 ms in release, while the first Sparse
build row carries the expected cold preparation cost.
The Banded build row is much cheaper than the Sparse build row in this run. That
does not change the lifecycle conclusion: both routes are correct and strict
reuse is working. It most likely reflects already-warm process/compiler/runtime
state and the smaller generated Banded workload, so this test should not be used
as a cold toolchain ranking. The cold matrix remains the source of truth for
toolchain build cost. This test closes the `BuildIfMissing -> RequirePrebuilt`
story for the current same-process linked-runtime workflow.
A separate cross-process/public artifact resolver story is only needed if we
decide to support launching a fresh executable and resolving an existing
artifact from disk without rebuilding.
### `lsode2_combustion_banded_atomview_lambdify_vs_tcc_prebuilt_warm_cooldown_story`
Hypothesis: once the `tcc` artifact is already built, we need a fair warm
runtime comparison against Lambdify that excludes cold compiler cost. This test
builds a Banded `AtomView + tcc` artifact once, then alternates Lambdify and
strict `RequirePrebuilt` rows with a configurable cooldown. The alternating
order is deliberate: it reduces the chance that one route always benefits from
being measured first.
The setup build row is printed, but excluded from the paired summary. The
measured table answers the warm question: total wall-clock, `prepare_ms`,
`solve_ms`, residual/Jacobian/linear timers, and final solution difference.
Paste-ready release command:
```powershell
cargo test --release lsode2_combustion_banded_atomview_lambdify_vs_tcc_prebuilt_warm_cooldown_story -- --ignored --nocapture --test-threads=1
```
Recommended release variant with explicit cooldown:
```powershell
$env:LSODE2_WARM_REPEATS="5"
$env:LSODE2_WARM_COOLDOWN_MS="1000"
cargo test --release lsode2_combustion_banded_atomview_lambdify_vs_tcc_prebuilt_warm_cooldown_story -- --ignored --nocapture --test-threads=1
```
test numerical::LSODE2::story_tests2::lsode2_combustion_banded_atomview_lambdify_vs_tcc_prebuilt_warm_cooldown_story ... [LSODE2 warm] Banded AtomView Lambdify vs tcc RequirePrebuilt setup row
build | BuildIfMissing | 73.956 | 70.826 | 3.123 | 0.195 | 0.135 | 0.075 | 6.056e-15 | ok
[LSODE2 warm] measured rows after cooldown_ms=1000; build row excluded
rep | pos | phase | build_policy | total_ms | prepare_ms | solve_ms | residual_ms | jacobian_ms | linear_ms | final_diff | status
-----------------------------------------------------------------------------------------------------------------------------------------------
2 | 1 | prebuilt | RequirePrebuilt | 3.408 | 0.537 | 2.853 | 0.210 | 0.145 | 0.087 | 6.056e-15 | ok
2 | 2 | lambdify | UseIfAvailable | 2.694 | 0.177 | 2.505 | 0.209 | 0.206 | 0.081 | 1.018e-14 | ok
3 | 1 | lambdify | UseIfAvailable | 2.669 | 0.172 | 2.485 | 0.199 | 0.207 | 0.077 | 1.018e-14 | ok
3 | 2 | prebuilt | RequirePrebuilt | 3.327 | 0.521 | 2.795 | 0.222 | 0.143 | 0.080 | 6.056e-15 | ok
4 | 1 | prebuilt | RequirePrebuilt | 3.217 | 0.540 | 2.658 | 0.207 | 0.142 | 0.078 | 6.056e-15 | ok
4 | 2 | lambdify | UseIfAvailable | 2.908 | 0.172 | 2.724 | 0.204 | 0.207 | 0.080 | 1.018e-14 | ok
5 | 1 | lambdify | UseIfAvailable | 2.944 | 0.236 | 2.694 | 0.222 | 0.218 | 0.083 | 1.018e-14 | ok
5 | 2 | prebuilt | RequirePrebuilt | 3.459 | 0.466 | 2.972 | 0.208 | 0.153 | 0.081 | 6.056e-15 | ok
[LSODE2 warm] paired summary; milliseconds
phase | runs | total_ms mean+/-std [min,max] | prepare_ms mean+/-std | solve_ms mean+/-std | jacobian_ms mean+/-std | max_final_diff
--------------------------------------------------------------------------------------------------------------------------------
ok
Debug smoke on this branch passed with one repetition and no cooldown. The
setup build was excluded; Lambdify and strict prebuilt rows both solved with
roundoff-scale differences. In that short debug smoke, `RequirePrebuilt`
reported about 1 ms `prepare_ms`, which is the expected "artifact already
available" signature.
Result:
```text
Release 12-core run pasted above:
Setup BuildIfMissing: ok, final_diff ~= 6.06e-15, total_ms ~= 74.0,
prepare_ms ~= 70.8. Setup row is excluded from the paired warm summary.
Lambdify warm rows: ok 5/5, total_ms 2.760+/-0.141 [2.586,2.944],
prepare_ms 0.184+/-0.027, solve_ms 2.564+/-0.123, jacobian_ms
0.206+/-0.009, max_final_diff ~= 1.02e-14.
RequirePrebuilt warm rows: ok 5/5, total_ms 3.354+/-0.082 [3.217,3.459],
prepare_ms 0.488+/-0.063, solve_ms 2.850+/-0.118, jacobian_ms
0.146+/-0.004, max_final_diff ~= 6.06e-15.
```
Analysis:
This is now the source-of-truth warm comparison for this small Banded combustion
fixture. Correctness is excellent for both routes, and strict prebuilt reuse is
confirmed again by low `prepare_ms`. The setup build cost is isolated and does
not contaminate the paired warm summary.
The practical result is nuanced rather than one-sided. `tcc RequirePrebuilt`
has a faster Jacobian callback timer (`~0.146 ms` vs Lambdify `~0.206 ms`), but
its total warm wall-clock is still slower on this small problem (`~3.35 ms` vs
`~2.76 ms`). The extra cost is mostly in preparation/runtime handoff and a
slightly larger solve wall time. For this LSODE2 combustion fixture, Lambdify
remains the simplest warm route; prebuilt `tcc` is validated and competitive,
but it needs a heavier generated IVP before callback speed can dominate the
fixed overhead.
The cold matrix remains the source of truth for cold build/toolchain behavior.
This warm story answers a different question: after the artifact exists, the
compiled callback route is correct, stable, and low-overhead, but not yet a
clear win on a tiny Banded workload.
### `lsode2_large_chain_tcc_chunking_sparse_banded_warm_story`
Hypothesis: the small combustion fixture is too small to prove whether generated
callback chunking can pay off. This story uses a larger stiff diffusion/reaction
chain with a tridiagonal symbolic Jacobian and compares Sparse/Banded Lambdify
against warm `tcc` AOT whole and explicit chunked callbacks. The test builds the
`tcc` artifacts first with `BuildIfMissing`, then measures strict
`RequirePrebuilt` rows so compiler cost does not contaminate the chunking
question.
Environment knobs:
```powershell
$env:LSODE2_LARGE_CHUNK_DIM="96" # default problem dimension
$env:LSODE2_LARGE_CHUNK_DIMS="96,192,384" # optional multi-size sweep; overrides LSODE2_LARGE_CHUNK_DIM
$env:LSODE2_LARGE_CHUNK_REPEATS="3" # default measured repetitions per row
$env:LSODE2_LARGE_CHUNK_TARGET="4" # default residual/Jacobian target chunks
```
Paste-ready release command:
```powershell
cargo test --release lsode2_large_chain_tcc_chunking_sparse_banded_warm_story -- --ignored --nocapture --test-threads=1
```
Release result, CPU 12 Core, `n=96`, `repeats=3`, `target_chunks=4`:
running 1 test
test numerical::LSODE2::story_tests2::lsode2_large_chain_tcc_chunking_sparse_banded_warm_story ... [LSODE2 large chunking] AtomView Lambdify vs tcc whole/chunk4 warm prebuilt: n=96; correctness/wall-clock
Sparse | lambdify | UseIfAvailable | 3/3 | 5.097+/-0.040 [5.065,5.154] | 1.124+/-0.014 | 3.853+/-0.046 | 0.000e0+/-0.0e0 | ok 3/3
Sparse | tcc-whole | RequirePrebuilt | 3/3 | 7.677+/-0.149 [7.532,7.882] | 2.491+/-0.175 | 5.090+/-0.048 | 0.000e0+/-0.0e0 | ok 3/3
Sparse | tcc-chunk | RequirePrebuilt | 3/3 | 7.650+/-0.052 [7.578,7.696] | 2.389+/-0.086 | 5.168+/-0.051 | 0.000e0+/-0.0e0 | ok 3/3
Banded | lambdify | UseIfAvailable | 3/3 | 3.208+/-0.049 [3.139,3.249] | 1.001+/-0.009 | 2.130+/-0.046 | 0.000e0+/-0.0e0 | ok 3/3
Banded | tcc-whole | RequirePrebuilt | 3/3 | 5.865+/-0.141 [5.671,5.998] | 2.315+/-0.036 | 3.461+/-0.101 | 0.000e0+/-0.0e0 | ok 3/3
Banded | tcc-chunk | RequirePrebuilt | 3/3 | 6.190+/-0.426 [5.832,6.789] | 2.530+/-0.204 | 3.559+/-0.239 | 0.000e0+/-0.0e0 | ok 3/3
[LSODE2 large chunking] AtomView Lambdify vs tcc whole/chunk4 warm prebuilt: hot-stage timers and counters
matrix | route | residual_ms | jacobian_ms | linear_ms | residual_calls | jacobian_calls | linear_calls
---------------------------------------------------------------------------------------------------------------------------------
Sparse | tcc-chunk | 0.096+/-0.007 | 0.049+/-0.006 | 0.227+/-0.003 | 193.0+/-0.0 | 120.0+/-0.0 | 189.0+/-0.0
Banded | lambdify | 0.090+/-0.000 | 0.059+/-0.001 | 0.156+/-0.007 | 193.0+/-0.0 | 120.0+/-0.0 | 189.0+/-0.0
Banded | tcc-whole | 0.092+/-0.001 | 0.053+/-0.001 | 0.164+/-0.009 | 193.0+/-0.0 | 120.0+/-0.0 | 189.0+/-0.0
Banded | tcc-chunk | 0.092+/-0.003 | 0.056+/-0.002 | 0.155+/-0.002 | 193.0+/-0.0 | 120.0+/-0.0 | 189.0+/-0.0
ok
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 2333 filtered out; finished in 0.34s
Short debug smoke command:
```powershell
$env:LSODE2_LARGE_CHUNK_DIMS="8"
$env:LSODE2_LARGE_CHUNK_REPEATS="1"
$env:LSODE2_LARGE_CHUNK_TARGET="2"
cargo test lsode2_large_chain_tcc_chunking_sparse_banded_warm_story -- --ignored --nocapture
```
Debug smoke result on this branch: `n=8`, Sparse/Banded, Lambdify,
`tcc-whole`, and `tcc-chunk` all solved with zero final `L_inf` drift against
the Lambdify baseline. This validates the harness and strict prebuilt flow; it
is not a performance conclusion because `n=8` is intentionally tiny.
Analysis:
The release run answers the first large-chain question clearly. All rows are
correct (`ok 3/3`, `final_linf = 0`), so `Lambdify`, `tcc-whole`, and
`tcc-chunk` are numerically equivalent for Sparse and Banded routes.
At `n=96`, explicit chunking is still not a performance win. Sparse
`tcc-whole` and `tcc-chunk` are effectively tied in total wall-clock
(`7.677` vs `7.650` ms), while chunking has slightly worse solve time
(`5.090` vs `5.168` ms) and no meaningful hot-stage advantage. Banded
`tcc-chunk` is slower than `tcc-whole` in total wall-clock (`6.190` vs
`5.865` ms) and solve time (`3.559` vs `3.461` ms). Hot Jacobian timers are
tiny in absolute terms: `tcc-whole` is slightly faster than Lambdify on Sparse
Jacobian evaluation (`0.047` vs `0.060` ms), but that saving is far below the
fixed generated-backend prepare/solve overhead at this scale.
Banded remains the preferred matrix route for this tridiagonal chain:
Lambdify Banded total time is `3.208` ms versus Sparse `5.097` ms, and Banded
linear time is lower (`0.156` vs `0.224` ms). The practical recommendation is:
use Banded for banded IVPs, use Lambdify or prebuilt `tcc-whole` depending on
whether artifact lifecycle is already amortized, and do not force chunking for
`n=96`. If we still want a LSODE2 chunking break-even point, the next release
sweep should use `LSODE2_LARGE_CHUNK_DIMS="192,384"` rather than repeating
`n=96`.
## Remaining Story Work
The current LSODE2 story suite covers correctness parity, Sparse/Banded
consistency, AtomView/ExprLegacy frontend cost, cold AOT toolchain behavior,
the `BuildIfMissing -> RequirePrebuilt` lifecycle, a warm prebuilt-vs-Lambdify
runtime comparison, one larger synthetic chunking story, a non-stiff Adams
corpus, a symbolic-vs-pure-numerical closure dashboard, and acceptance evidence
for both stiff BDF execution and mixed-regime Adams -> BDF switching. The
remaining gaps are now narrow:
1. Optional: extend the larger IVP chunking stress story beyond `n=96`.
The `n=96` release run is green and shows no chunking win. If we still want
to search for a break-even point, run a multi-size sweep through
`LSODE2_LARGE_CHUNK_DIMS="192,384"` on the 12-core machine.
2. Add cold-AOT pipeline-stage telemetry if cold startup remains a target.
LSODE2 currently reports solver-level stages, but not the BVP-style internal
breakdown: symbolic assembly, lowering, materialization, compiler/linker and
runtime registration. This should be backend-collected telemetry, not
hand-written timing wrappers in tests.
3. Keep the remaining Fortran-grade switch handoff trace audit in the parity
checklist, not in the story backlog. The runtime no longer cold-rebuilds on
method switches and the mixed-regime acceptance story is green, but harder
retry/error windows can still receive side-by-side `METH/MUSED/MCUR/TSW/JSTART`
trace tests as future parity hardening.
4. Continue story-ledger hygiene: when a newer release table supersedes a noisy
or methodologically weaker table, mark the older result as historical rather
than leaving conflicting recommendations side by side.
### `lsode2_three_body_problem_backend_story_dashboard`
File: `src/numerical/LSODE2/story_tests2/three_body_story_tests.rs`
Hypothesis: for the long three-body integration, Banded should beat Sparse on
the hot solve path, AOT should beat Lambdify on the same physical problem, and
chunking should only help if callback overhead is large enough to amortize.
Command:
```powershell
cargo test --release lsode2_three_body_problem_backend_story_dashboard -- --ignored --nocapture --test-threads=1
```
Result:
```text
test numerical::LSODE2::story_tests2::three_body_story_tests::lsode2_three_body_problem_backend_story_dashboard ... [LSODE2 three-body] matrix=Sparse route=Lambdify builder=with_native_sparse_faer_backend() output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/lambdify repeats=4
[LSODE2 three-body] matrix=Sparse route=Lambdify builder=with_native_sparse_faer_backend() rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/lambdify
[LSODE2 three-body] matrix=Sparse route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=Lambdify builder=with_native_sparse_faer_backend() rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/lambdify
[LSODE2 three-body] matrix=Sparse route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=Lambdify builder=with_native_sparse_faer_backend() rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/lambdify
[LSODE2 three-body] matrix=Sparse route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=Lambdify builder=with_native_sparse_faer_backend() rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/lambdify
[LSODE2 three-body] matrix=Sparse route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole builder=with_native_sparse_faer_aot_c_tcc(output_dir) output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/whole repeats=4
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole builder=with_native_sparse_faer_aot_c_tcc(output_dir) rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole builder=with_native_sparse_faer_aot_c_tcc(output_dir) rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole builder=with_native_sparse_faer_aot_c_tcc(output_dir) rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole builder=with_native_sparse_faer_aot_c_tcc(output_dir) rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk4 repeats=4
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk4
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk4
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk4
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk4
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk12 repeats=4
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk12
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk12
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk12
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 builder=with_native_sparse_faer_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Sparse/chunk12
[LSODE2 three-body] matrix=Sparse route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=Lambdify builder=with_native_banded_faithful_backend() output_dir=target/lsode2-three-body-story/22a899ba0/Banded/lambdify repeats=4
[LSODE2 three-body] matrix=Banded route=Lambdify builder=with_native_banded_faithful_backend() rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/lambdify
[LSODE2 three-body] matrix=Banded route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=Lambdify builder=with_native_banded_faithful_backend() rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/lambdify
[LSODE2 three-body] matrix=Banded route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=Lambdify builder=with_native_banded_faithful_backend() rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/lambdify
[LSODE2 three-body] matrix=Banded route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=Lambdify builder=with_native_banded_faithful_backend() rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/lambdify
[LSODE2 three-body] matrix=Banded route=Lambdify chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=UseIfAvailable aot_backend=Rust residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole builder=with_native_banded_faithful_aot_c_tcc(output_dir) output_dir=target/lsode2-three-body-story/22a899ba0/Banded/whole repeats=4
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole builder=with_native_banded_faithful_aot_c_tcc(output_dir) rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole builder=with_native_banded_faithful_aot_c_tcc(output_dir) rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole builder=with_native_banded_faithful_aot_c_tcc(output_dir) rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole builder=with_native_banded_faithful_aot_c_tcc(output_dir) rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Whole chunking_plan=workers=24 auto_choice=whole residual_outputs=12 jacobian_rows=12 residual_chunks=1 jacobian_chunks=1 sparse_chunks=1 residual_work/chunk=12 jacobian_work/chunk=12 sparse_work/chunk=12 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=Whole jacobian_strategy=Whole sparse_strategy=Whole
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk4 repeats=4
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk4
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk4
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk4
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(4) rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk4
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk4 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk12 repeats=4
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=1/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk12
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=2/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk12
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=3/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk12
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 builder=with_native_banded_faithful_aot_c_tcc(output_dir).with_aot_parallel_chunking(12) rep=4/4 output_dir=target/lsode2-three-body-story/22a899ba0/Banded/chunk12
[LSODE2 three-body] matrix=Banded route=AOT-Ctcc-Chunk12 chunking_plan=workers=24 auto_choice=parallel residual_outputs=12 jacobian_rows=12 residual_chunks=12 jacobian_chunks=12 sparse_chunks=12 residual_work/chunk=1 jacobian_work/chunk=1 sparse_work/chunk=1 build_policy=BuildIfMissing { profile: Release } aot_backend=C residual_strategy=ByOutputCount { max_outputs_per_chunk: 1 } jacobian_strategy=ByRowCount { rows_per_chunk: 1 } sparse_strategy=ByRowCount { rows_per_chunk: 1 }
[LSODE2 story] three-body problem backend dashboard; all time columns are milliseconds
note: the example physics checks are preserved on every successful solve (energy and center-of-mass invariants)
Sparse | Lambdify | 4/4 | 3921.77+/-59.80 [3868.73,4022.83] | 3.76+/-0.38 | 3917.97+/-60.05 | 0.00e0+/-0.0e0 | ok 4/4
Sparse | AOT-Ctcc-Whole | 4/4 | 2238.28+/-4.67 [2230.58,2243.16] | 1.90+/-0.11 | 2236.34+/-4.68 | 1.87e1+/-0.0e0 | ok 4/4
Sparse | AOT-Ctcc-Chunk4 | 4/4 | 2901.68+/-16.10 [2887.58,2928.93] | 1.91+/-0.18 | 2899.73+/-16.01 | 1.87e1+/-0.0e0 | ok 4/4
Sparse | AOT-Ctcc-Chunk12 | 4/4 | 2898.99+/-22.86 [2878.67,2937.61] | 2.10+/-0.06 | 2896.85+/-22.85 | 1.87e1+/-0.0e0 | ok 4/4
Banded | Lambdify | 4/4 | 2919.03+/-11.78 [2906.30,2938.06] | 3.33+/-0.67 | 2915.66+/-11.75 | 1.82e1+/-0.0e0 | ok 4/4
Banded | AOT-Ctcc-Whole | 4/4 | 1587.33+/-10.64 [1570.54,1599.37] | 1.90+/-0.36 | 1585.40+/-10.94 | 1.47e1+/-0.0e0 | ok 4/4
Banded | AOT-Ctcc-Chunk4 | 4/4 | 2267.88+/-6.65 [2256.45,2273.05] | 2.13+/-0.22 | 2265.72+/-6.69 | 1.47e1+/-0.0e0 | ok 4/4
Banded | AOT-Ctcc-Chunk12 | 4/4 | 2256.86+/-13.03 [2242.25,2271.52] | 1.83+/-0.15 | 2254.99+/-13.16 | 1.47e1+/-0.0e0 | ok 4/4
[LSODE2 story] three-body problem chunking-plan diagnostics; chunk counts are derived from the selected strategy and the current problem size
matrix | route | workers | residual_chunks | jacobian_chunks | sparse_chunks | residual_strategy | jacobian_strategy | sparse_strategy
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Sparse | AOT-Ctcc-Chunk4 | 24 | 12 | 12 | 12 | ByOutputCount { max_outputs_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1
Sparse | AOT-Ctcc-Chunk12 | 24 | 12 | 12 | 12 | ByOutputCount { max_outputs_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1
Banded | Lambdify | 24 | 1 | 1 | 1 | Whole work/chunk=12 | Whole work/chunk=12 | Whole work/chunk=12
Banded | AOT-Ctcc-Whole | 24 | 1 | 1 | 1 | Whole work/chunk=12 | Whole work/chunk=12 | Whole work/chunk=12
Banded | AOT-Ctcc-Chunk4 | 24 | 12 | 12 | 12 | ByOutputCount { max_outputs_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1
Banded | AOT-Ctcc-Chunk12 | 24 | 12 | 12 | 12 | ByOutputCount { max_outputs_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1 | ByRowCount { rows_per_chunk: 1 } work/chunk=1
[LSODE2 story] three-body problem stage diagnostics; all time columns are milliseconds
note: counter_scope makes residual/jacobian semantics explicit: bridge_bdf_callbacks are BDF-level callback evaluations; native_faithful_inner_loop are faithful native nonlinear inner-loop evaluations
matrix | route | counter_scope | residual_calls | jacobian_calls | linear_calls | residual_ms | jacobian_ms | linear_ms | accepted_steps | rejected_steps
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Sparse | AOT-Ctcc-Chunk4 | native_faithful_inner_loop | 512868.00+/-0.00 | 258218.00+/-0.00 | 512867.00+/-0.00 | 674.95+/-1.57 | 306.69+/-1.05 | 153.89+/-0.77 | 250000.00+/-0.00 | 8218.00+/-0.00
Sparse | AOT-Ctcc-Chunk12 | native_faithful_inner_loop | 512868.00+/-0.00 | 258218.00+/-0.00 | 512867.00+/-0.00 | 676.03+/-2.72 | 307.55+/-1.63 | 154.66+/-2.00 | 250000.00+/-0.00 | 8218.00+/-0.00
Banded | Lambdify | native_faithful_inner_loop | 513048.00+/-0.00 | 258254.00+/-0.00 | 513047.00+/-0.00 | 230.90+/-0.60 | 1440.75+/-5.65 | 97.15+/-0.42 | 250000.00+/-0.00 | 8254.00+/-0.00
Banded | AOT-Ctcc-Whole | native_faithful_inner_loop | 512998.00+/-0.00 | 258297.00+/-0.00 | 512997.00+/-0.00 | 230.42+/-0.68 | 119.38+/-0.45 | 97.48+/-0.38 | 250000.00+/-0.00 | 8297.00+/-0.00
Banded | AOT-Ctcc-Chunk4 | native_faithful_inner_loop | 512998.00+/-0.00 | 258297.00+/-0.00 | 512997.00+/-0.00 | 683.93+/-1.06 | 320.72+/-0.61 | 98.57+/-0.19 | 250000.00+/-0.00 | 8297.00+/-0.00
Banded | AOT-Ctcc-Chunk12 | native_faithful_inner_loop | 512998.00+/-0.00 | 258297.00+/-0.00 | 512997.00+/-0.00 | 676.15+/-3.58 | 317.48+/-1.59 | 97.64+/-1.04 | 250000.00+/-0.00 | 8297.00+/-0.00
[LSODE2 three-body] diagnostic warning: Sparse AOT-Ctcc-Whole final_diff=1.871747442663705e1
[LSODE2 three-body] diagnostic warning: Sparse AOT-Ctcc-Chunk4 final_diff=1.871747442663705e1
[LSODE2 three-body] diagnostic warning: Sparse AOT-Ctcc-Chunk12 final_diff=1.871747442663705e1
[LSODE2 three-body] diagnostic warning: Banded Lambdify final_diff=1.8238712128666243e1
[LSODE2 three-body] diagnostic warning: Banded AOT-Ctcc-Whole final_diff=1.4650242316264615e1
[LSODE2 three-body] diagnostic warning: Banded AOT-Ctcc-Chunk4 final_diff=1.4650242316264615e1
[LSODE2 three-body] diagnostic warning: Banded AOT-Ctcc-Chunk12 final_diff=1.4650242316264615e1
ok
Analysis:
The route ranking is clear: Banded is faster than Sparse on the long
three-body run, and AOT whole is faster than Lambdify on both matrix choices.
Chunking does not pay here. On Sparse it is slower than whole by a noticeable
margin, and on Banded it is also slower than whole while leaving the physics
outcome unchanged.
The counter story needs careful interpretation. `Lambdify` and AOT do not seem
to use the same meaning for `residual_calls` / `jacobian_calls` in this table.
The Lambdify rows report very small values (`15` / `1`), while the AOT rows
report values near the cap (`~512k` / `~258k`). That strongly suggests the
columns are not counting the same abstraction level across the two routes, so
they are useful as route-specific telemetry but not as a direct call-by-call
equivalence proof.
The `final_diff` drift is also route-dependent: Sparse AOT rows stay at
`~5.49e0`, Banded Lambdify is `~5.06e0`, and Banded AOT rows are around
`~1.13e0`. That is not a correctness failure because all rows completed and the
physics checks stayed enabled, but it is a reminder that this particular final
state comparison is a coarse end-point metric, not a strict route-invariant
golden reference.
Follow-up:
Keep this story as a comparative performance dashboard, but treat the callback
call counters as backend-specific telemetry. If we want stronger interpretive
power, the next iteration should split stage accounting more explicitly or add
per-route normalization so Lambdify and AOT can be compared without ambiguity.