Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
29 commits
Select commit Hold shift + click to select a range
eed911b
Fix native D.UPM when exactly one degree is an integer
gitRasheed Nov 19, 2026
b184171
Regenerate R parity cache for NNS 14.2
OVVO-Financial Nov 19, 2026
48ee403
Keep multiply and add separately rounded in native kernels
gitRasheed Nov 19, 2026
cc2bbee
Match partial moments and dominance to R NNS 14.2
gitRasheed Nov 19, 2026
563d4a6
Use R's finite-value errors for partial moments
gitRasheed Nov 19, 2026
04da452
Preserve numeric targets and match R validation order
gitRasheed Nov 19, 2026
251fb48
Reject categorical columns in VaR data frames
gitRasheed Nov 20, 2026
7a2ddd7
Preserve datetime target units and validate partial-moment inputs
gitRasheed Nov 20, 2026
bb05d3d
Regenerate R parity cache for NNS 14.2
OVVO-Financial Nov 21, 2026
103f0fe
Merge remote-tracking branch 'origin/rasheed/no-fp-contract' into cla…
claude Nov 21, 2026
0e5b744
Merge remote-tracking branch 'origin/rasheed/d-upm-mixed-degrees' int…
claude Nov 21, 2026
62362b7
Merge remote-tracking branch 'origin/rasheed/r142-sync' into claude/n…
claude Nov 21, 2026
60fea12
Follow R 14.2's NNS.moments, PM.matrix mean target and distance ties
claude Nov 21, 2026
9088f7a
Follow R 14.2's mode peaks, rescale rules and NNS.dep matrix and p-va…
claude Nov 21, 2026
05dee2c
Follow R 14.2's regression points, default n.best and prediction dist…
claude Nov 21, 2026
8bcdddf
Report the final classifier's predictors as NNS.reg.variables
claude Nov 21, 2026
7df4c67
Follow R 14.2's meboot bounds and variance, ARMA bands, dy.dx and NNS…
claude Nov 21, 2026
a6b95b3
Let NNS.FSD without a plot follow R 14.2's unidirectional test
claude Nov 21, 2026
048351e
Break classification stack ties as R 14.2 does
claude Nov 21, 2026
805fce8
Pick each VAR lag's tau as R 14.2 does
claude Nov 21, 2026
614eb91
Update test contracts for R NNS 14.2
claude Nov 21, 2026
1bb4ee4
Regenerate the R parity cache with NNS 14.2 (f0387ed)
claude Nov 21, 2026
c33b95c
Merge the builder's NNS 14.2 cache commit from automation/r-cache-nns…
claude Nov 21, 2026
09487c2
Record the NNS 14.2 oracle (f0387ed) in the sync manifest
claude Nov 21, 2026
41c5606
Follow R 14.2's input layer (R/inputs.R) and circular-shift rule
claude Nov 22, 2026
dabb7a9
Commit the live R sweeps to scripts/live_parity
claude Nov 22, 2026
79e1b4f
Run the live sweeps in the cache builder; pin the 14.2 oracle bundle
claude Nov 22, 2026
915a73d
Document the NNS 14.2 oracle, the live sweeps and input validation
claude Nov 22, 2026
ff41639
Restore protected workflow baseline
github-actions[bot] Nov 22, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 12 additions & 3 deletions CMakeLists.txt
Original file line number Diff line number Diff line change
Expand Up @@ -15,18 +15,22 @@ add_subdirectory(extern/NNS-core)

# Preserve the accepted NNS-core vendor tree; explicitly overlay the R14
# central-tendency kernels that its R13.3 snapshot does not yet supply, the
# multivariate prediction kernel with R's own nmath densities, and R's NNS.seas
# lag scan (kept out of the nanobind module, which is compiled for size).
# multivariate prediction kernel with R's own nmath densities, the partial
# moments with R's D.UPM degree handling, and R's NNS.seas lag scan (kept out of
# the nanobind module, which is compiled for size).
get_target_property(NNS_CORE_SOURCES nnscore SOURCES)
set(NNS_R14_CORE_SOURCES "")
foreach(source IN LISTS NNS_CORE_SOURCES)
if(NOT source STREQUAL "src/central_tendencies.cpp" AND NOT source STREQUAL "src/mreg_predict.cpp")
if(NOT source STREQUAL "src/central_tendencies.cpp" AND NOT source STREQUAL "src/mreg_predict.cpp"
AND NOT source STREQUAL "src/partial_moments.cpp")
list(APPEND NNS_R14_CORE_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/extern/NNS-core/${source}")
endif()
endforeach()
list(APPEND NNS_R14_CORE_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/src/nns/_r14_central_tendencies.cpp")
list(APPEND NNS_R14_CORE_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/src/nns/_r14_mreg_predict.cpp")
list(APPEND NNS_R14_CORE_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/src/nns/_r14_partial_moments.cpp")
list(APPEND NNS_R14_CORE_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/src/nns/_r14_seas.cpp")
list(APPEND NNS_R14_CORE_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/src/nns/_r14_pm_norm.cpp")
set_property(TARGET nnscore PROPERTY SOURCES "${NNS_R14_CORE_SOURCES}")

# The Python extension is a shared object. The vendored NNS-core target builds
Expand All @@ -35,6 +39,11 @@ set_property(TARGET nnscore PROPERTY SOURCES "${NNS_R14_CORE_SOURCES}")
set_target_properties(nnscore PROPERTIES POSITION_INDEPENDENT_CODE ON)

nanobind_add_module(_nnscore src/nns/_nnscore_bindings.cpp)
# The reference loops round the product before adding it to the accumulator.
if(CMAKE_CXX_COMPILER_ID MATCHES "^(GNU|Clang|AppleClang)$")
target_compile_options(nnscore PRIVATE -ffp-contract=off)
target_compile_options(_nnscore PRIVATE -ffp-contract=off)
endif()
target_link_libraries(_nnscore PRIVATE nnscore)
target_compile_features(_nnscore PRIVATE cxx_std_17)
install(TARGETS _nnscore LIBRARY DESTINATION nns)
2 changes: 1 addition & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -125,7 +125,7 @@ uv run python examples/run_all_vignettes.py

## Design boundaries

NNS Python prioritizes stable public behavior from R NNS 14.0, matched to the
NNS Python prioritizes stable public behavior from R NNS 14.2, matched to the
last bit where verified (see the parity page), not private helper parity. The
package returns NumPy arrays and plain dictionaries rather than R `data.table`
objects and uses explicit Python errors for unsafe R coercions.
Expand Down
2 changes: 1 addition & 1 deletion docs/api_status.md
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,7 @@ invariant, and property coverage.
| Stack: `nns_stack` | implemented | medium-high | Numeric/class paths, intervals, factor predictors, `ts_test`, and R's seeded CV/balance draws (bit-for-bit) are covered; test columns are matched by name and selection outputs carry predictor names. |
| Boost: `nns_boost` | implemented | medium | Learner trials, epochs, and the final stack match R's seeded stream; feature weights/frequencies are named by column. |
| Seasonality: `nns_seas` | implemented | high | Non-plotting installed-R path is implemented and cached defensively. |
| ARMA and VAR: `nns_arma`, `nns_arma_optim`, `nns_var` | implemented | high | Forecasts, `nns_arma_optim` (including its smoothed `nns_reg` paths) and every VAR dimension-reduction path (`cor`, `NNS.dep`, `NNS.caus`, `all`) match R NNS 14.0 to the last bit on the live sweeps, including own-lag weights, `objective="max"`, vector and list `tau`, and interior/trailing missing values. ARMA prediction intervals reproduce R's `set.seed` stream through `random_seed`. VAR's `relevant_variables` are the regressors the cross-validated dimension reduction threshold keeps, and `multivariate` is `nns_stack(method=1)` (`nns_reg`) on exactly those regressors. |
| ARMA and VAR: `nns_arma`, `nns_arma_optim`, `nns_var` | implemented | high | Forecasts, `nns_arma_optim` (including its smoothed `nns_reg` paths) and every VAR dimension-reduction path (`cor`, `NNS.dep`, `NNS.caus`, `all`) match R NNS 14.2 to the last bit on the live sweeps, including own-lag weights, `objective="max"`, vector and list `tau`, and interior/trailing missing values. ARMA prediction intervals reproduce R's `set.seed` stream through `random_seed`. VAR's `relevant_variables` are the regressors the cross-validated dimension reduction threshold keeps, and `multivariate` is `nns_stack(method=1)` (`nns_reg`) on exactly those regressors. |
| Bootstrap/Monte Carlo: `nns_meboot`, `nns_mc` | implemented | high | With `random_seed = s` the replicates and summaries are R's after `set.seed(s)`, to the last bit. |
| Stochastic dominance/superiority: `fsd`, `ssd`, `tsd`, `.uni` wrappers, `nns_ss`, `nns_sd_cluster`, `sd_efficient_set` | implemented | high | Contracts, the efficient-set order and the clustering match R 14 to the last bit; `nns_ss` intervals reproduce R's `set.seed` stream through `random_seed`. SD uses exact pure-NumPy prefix-pair kernels plus a degree-1 discrete order-statistic matrix path; R's C++ core remains faster on full finance fixtures. |
| ANOVA: `nns_anova` | implemented | high | Binary, multi-group, pairwise, and degenerate `NaN` conventions are covered. |
Expand Down
52 changes: 39 additions & 13 deletions docs/conventions.md
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,33 @@ NNS Python raises `ValueError`.

Rationale: empty arrays in Python are upstream bugs, and NumPy convention is to warn or fail on empty reductions rather than silently produce a meaningful statistic.

## Input Validation

R NNS 14.2 checks every data argument through one input layer (`R/inputs.R`),
and NNS Python follows it (`nns._inputs`): the same inputs are accepted and the
same `ValueError` messages raised, in R's order. Python inputs stand for R types:
a pandas `Categorical` (or categorical `Series`) is a factor, a `DataFrame` a
data.frame, a 2-D array a matrix, a string array a character vector and a
boolean array a logical vector. So, as in R, a factor is rejected with
``"`x` must be numeric, not a factor."``, a character or logical vector passed to
a partial-moment ratio or co-moment with ``"x should be numeric vector, or data
table"``, and `nns_seas`, `nns_arma`, `nns_meboot`, `nns_mc`, `nns_ss`,
`nns_dep`, `nns_part`, `nns_causation`, `nns_copula`, `nns_reg`, `dy_dx` and
`nns_anova` give R's messages for missing, infinite, mis-shaped and too-short
inputs. A one-column `DataFrame` or 2-D array is read as its column.

`scripts/live_parity/fz_inputs.py` checks this against R. Beyond the empty-input
rule above, the differences it reports as documented are:

- A matrix with several columns is rejected (R flattens it into one vector).
- A one-column `DataFrame` is accepted where R fails converting the data frame
(`"'list' object cannot be coerced to type 'double'"`).
- Infinite values: R's `LPM.VaR`/`UPM.VaR` and `NNS.ANOVA` stop with internal
errors (`"missing value where TRUE/FALSE needed"`), and R 14.2's `NNS.MC`
crashes the R session; NNS Python computes or raises its own message.
- `nns_copula(x, y)` takes the two variables separately, so R's recycling of
unequal columns in `NNS.copula(cbind(x, y))` does not arise.

## Co-Moment Lengths

`co_lpm`, `co_upm`, `d_lpm` and `d_upm` raise `ValueError` unless `x` and `y` are
Expand All @@ -69,12 +96,11 @@ check the same rules.

## PM Matrix Target Defaults

R `PM.matrix` uses column means when `target` is `NULL` or any non-numeric value.
NNS Python accepts `None` and `"mean"` for this behavior, with Rcpp's `colMeans`
arithmetic (a row-by-row double sum divided by the row count). NNS Python also broadcasts a
scalar numeric target across all variables; R requires callers to pass an
explicit vector such as `rep(0, ncol(variable))`. Target vectors whose length
does not match the number of variables raise `ValueError`.
Use `None`, `"mean"` or `"means"` for column means. Python and R sum each column
row by row in double precision and divide by the row count. Numeric vectors and
matrices supply one target per variable column. R reads matrix targets in column
order, and Python does the same. Python also broadcasts a numeric scalar across
columns. In R, pass an explicit vector such as `rep(0, ncol(variable))`.

## Mode Default Divergence From R

Expand All @@ -94,6 +120,8 @@ multiplication, fractional degrees by libm `pow`, double sums accumulated
observation by observation and scaled by `1 / n`. `pm_matrix` fills the upper
triangle and mirrors it (CUPM, CLPM and the covariance plainly, DUPM and DLPM
crossed), normalising each cell by `CUPM + DUPM + DLPM + CLPM` in that order.
A normalised n-dimensional moment or matrix cell with an exactly zero total takes
its degree-zero value. In a joint survival calculation, ties are excluded.
`nns_moments` is R 14's `NNS.moments`: deviations from R's `mean()`,
`mean(z^3)` and `mean(z^4)` with R's `^`, and the variance raised by `^1.5`.

Expand Down Expand Up @@ -139,10 +167,9 @@ unrecognised `type` warns (and `fsd` then runs its continuous branch, while
returns the column names), and a vector is one column, as with R's `as.matrix`.
Inputs are checked in R's order and with R's messages: missing values raise
"You have some missing values, please address.", empty samples "Each sample
needs at least one observation.". Like R, the `.uni` tests, `sd_efficient_set`
and `nns_sd_cluster` accept infinite values and compare them as R's C++ test
does, while `fsd`, `ssd` and `tsd` reject them (R first evaluates the plotted
partial-moment curves, which check `x` and then the threshold grid).
needs at least one observation.". Only discrete first-order comparisons accept infinite values in
the `.uni` tests, efficient sets and clustering. The top-level `fsd`, `ssd` and
`tsd` check the partial-moment curves first and reject infinite values.
`nns_sd_cluster` validates its data in the first efficient-set search, as R
does, so a matrix with no more than `min_cluster` columns comes back as one
cluster unchecked, and `min_cluster` is compared as a real number (`1.5` merges
Expand Down Expand Up @@ -657,9 +684,8 @@ last bit: each column is scaled by `1 / range` (the range including the
target), the distance accumulates `diff^2 + |diff|` column by column, and the
weights are R's nmath densities summed in R's order.

* `nns_distance` is `NNS_distance_cpp`: the `k` nearest rows in the order
libstdc++'s `std::partial_sort` (or `std::sort` when every row is used)
leaves them, which decides where tied distances fall, then the eight weights
* `nns_distance` is `NNS_distance_cpp`: the `k` nearest rows, equal distances
ordered by RPM row as R 14.2's kernels order them, then the eight weights
(uniform, `dt`, inverse distance, `dexp` with scale `k`, reversed
`|dlnorm(rank, 0, sd(1:k), log = TRUE)|`, `rank^-2`, `dnorm` and an RBF),
normalised and combined.
Expand Down
81 changes: 43 additions & 38 deletions docs/parity.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,14 +2,13 @@

## Target and status

R NNS 14.0 is the parity target: the source at
`OVVO-Financial/NNS@9c8bd21a77b377cb49d43fececc662d62b00a94e`
(`NNS_14.0.tar.gz`, SHA256
`754dad65629b95bf3ddca9bf9018ef197f5c346dd12dcfb49e505835920bc5cb`; the same
archive is still committed at the R repository's later `de5d4fe`). Every shard
in `tests/_r_cache/` is stamped `nns_version` `14.0`. NNS-core remains the
native C++ foundation; R 14 kernels it does not carry enter as hashed overlays
recorded in `sync/nns_source.json`.
R NNS 14.2 is the parity target: `NNS_14.2.tar.gz` as committed at
`OVVO-Financial/NNS@f0387ed857de3d3bf047fa581b854dc5e3d0c3a6` (SHA256
`a4e04b26b3ea01aeae1d6676448194b275bdd3f2aacbafd78b82ba5c2a649201`). Its `R/` and
`src/` files equal the Git tree at that commit apart from CRLF line endings and
final newlines. Every shard in `tests/_r_cache/` is stamped `nns_version`
`14.2`. NNS-core remains the native C++ foundation; R 14 kernels it does not
carry enter as hashed overlays recorded in `sync/nns_source.json`.

Parity is bounded by the committed tests and cache:

Expand All @@ -22,7 +21,7 @@ Parity is bounded by the committed tests and cache:
(`tests/parity/test_original_*`).

Any cache miss under `PYNNS_R_CACHE_ONLY=1` is a parity-data gap until the cache
is regenerated with Rscript and the installed R NNS 14.0 oracle.
is regenerated with Rscript and the installed R NNS 14.2 oracle.

Plot artifact policy is unchanged: plots and `Rplots.pdf` artifacts are not
parity outputs in pytest; returned values are. See
Expand All @@ -35,15 +34,15 @@ reproduces R's arithmetic: long-double `sum`, `mean`, `var`, `prod` and
`rowSums`; Rcpp's double accumulations and R's C++ operation order; R's `^`;
libm `exp`, `log` and `pow` in place of NumPy's vectorised versions; R's nmath
densities and quantiles; `optimize`; the partial sort behind
`mean(trim = )`; the order libstdc++'s `std::sort` and `std::partial_sort`
give tied keys; and R's smoothing spline. It does so with the native extension
`mean(trim = )`; and R's smoothing spline. It does so with the native extension
and with the NumPy fallback (`NNS_DISABLE_NATIVE=1`). Where R draws random
numbers, `random_seed = s` (`seed` for `nns_causation`) reproduces R's stream
after `set.seed(s)`.

Live differential sweeps compare the port with installed R at exact equality
(`rtol = atol = 0`). At the time of writing every case below matches, with and
without the native extension:
Live differential sweeps (`scripts/live_parity/`, see its README) compare the
port with installed R at exact equality (`rtol = atol = 0`). At the time of
writing every case below matches, with and without the native extension
(`python scripts/live_parity/run.py --mode both`):

| Area | Cases |
| --- | ---: |
Expand All @@ -52,17 +51,22 @@ without the native extension:
| `NNS.distance`, `NNS.distance.bulk` | 928 |
| `NNS.part` | 920 |
| `NNS.reg` (univariate, multivariate, classification, dimension reduction, smoothing) | 480 |
| R's smoothing spline under `NNS.reg(smooth = TRUE)` | 264 |
| `NNS.seas`, `NNS.ARMA`, `NNS.ARMA.optim`, `NNS.diff`, `dy.dx`, `dy.d_` | 345 |
| `NNS.VAR` | 40 |
| `NNS.stack`, `NNS.boost` | 56 |
| Classification `NNS.stack` built to tie (coarse predictors, pooled folds) | 72 |
| `NNS.ANOVA` | 224 |
| Stochastic dominance (`NNS.FSD`/`SSD`/`TSD`, `.uni`, `NNS.SD.cluster`, `NNS.SD.efficient.set`) | 2,583 |
| Stochastic dominance (`NNS.FSD`/`SSD`/`TSD`, `.uni`, `NNS.SD.cluster`, `NNS.SD.efficient.set`), edge cases included | 2,583 |
| `NNS.meboot`, `NNS.MC`, `NNS.SS` and ARMA intervals under `set.seed` | 502 |
| Permutation p-values (`NNS.seas`, `NNS.dep`, `NNS.caus`) | 198 |
| Permutation p-values (`NNS.seas`, `NNS.dep`, `NNS.caus`) and `NNS.dep` permutation matches | 348 |
| Multivariate `NNS.reg` predictions on 900-1,500 rows (every distance, `n.best`, tied coordinates) | 84 |
| Paths compiled in the 14.0 update: integer-degree `LPM.VaR`/`UPM.VaR`, `NNS.copula`, `NNS.dep` matrices, classification predictions, mixed `dy.d_` | 110 |
| Compiled paths: integer-degree `LPM.VaR`/`UPM.VaR`, `NNS.copula`, `NNS.dep` matrices, classification predictions, mixed `dy.d_` | 110 |
| Input validation against R 14.2's input layer (accepted inputs and messages; see [conventions](conventions.md#input-validation)) | 546 |

The `*_r14.py` parity tests pin these results with exact cache entries.
The `*_r14.py` parity tests pin these results with exact cache entries. The
cache builder runs every sweep in both modes after regenerating the cache for a
new R release and reports the table in its pull request.

Scope: the oracle is R 4.5.0 on x86-64 Linux (glibc libm, the reference BLAS,
no FMA contraction), which is where the port was verified. Another platform's
Expand All @@ -89,30 +93,30 @@ mypy
python -m build # packaging check
```

## Installing the R NNS 14.0 oracle
## Installing the R NNS 14.2 oracle

Install the pinned R 14.0 source package, never the CRAN release, into R 4.5.0
Install the pinned R 14.2 source package, never the CRAN release, into R 4.5.0
(see the scope note under Bit-level parity):

```bash
curl --fail --location -o NNS_14.0.tar.gz \
https://raw.githubusercontent.com/OVVO-Financial/NNS/9c8bd21a77b377cb49d43fececc662d62b00a94e/NNS_14.0.tar.gz
curl --fail --location -o NNS_14.2.tar.gz \
https://raw.githubusercontent.com/OVVO-Financial/NNS/f0387ed857de3d3bf047fa581b854dc5e3d0c3a6/NNS_14.2.tar.gz
```

```bash
echo '754dad65629b95bf3ddca9bf9018ef197f5c346dd12dcfb49e505835920bc5cb NNS_14.0.tar.gz' | sha256sum -c -
echo 'a4e04b26b3ea01aeae1d6676448194b275bdd3f2aacbafd78b82ba5c2a649201 NNS_14.2.tar.gz' | sha256sum -c -
```

```bash
python scripts/install_local_r_nns.py --source NNS_14.0.tar.gz --expected-version 14.0
python scripts/install_local_r_nns.py --source NNS_14.2.tar.gz --expected-version 14.2
```

`tools/NNS` and `tools/NNS_13.0.tar.gz` hold the earlier R 13.0 source for
reference; they are not the parity target.

## Regenerating the parity cache

After confirming `packageVersion("NNS") == "14.0"`, regenerate the committed
After confirming `packageVersion("NNS") == "14.2"`, regenerate the committed
cache with cache-only/offline toggles unset:

```bash
Expand All @@ -137,28 +141,29 @@ PYNNS_R_CACHE_ONLY=1 python -m pytest -q -n 0 tests/parity

A `RuntimeError: R cache miss ...` means the cache is incomplete (regenerate the
missing live R entries); an `AssertionError`/numeric mismatch means Python
behavior differs from R NNS 14.0 and the Python implementation must be fixed
behavior differs from R NNS 14.2 and the Python implementation must be fixed
without loosening tolerances.

Where an R toolchain is unavailable (for example sandboxed CI or
proxy-restricted runners that cannot install R), the cache cannot be regenerated
live; rerun the local-source install and `scripts/regenerate_r_cache.py` on a
host with R.

## Cache provenance: R 14.0
## Cache provenance: R 14.2

The committed shards in `tests/_r_cache/` were generated from the pinned R NNS
14.0 source above and carry `nns_version` `14.0`. To reproduce, install that
package and point `R_LIBS` at its library when regenerating
(`R_LIBS=<lib> python scripts/regenerate_r_cache.py ...`). The earlier shards,
generated from a patched R 13.3, have all been regenerated from R 14.0.

The R 14.0 update of 13 November 2026 (`9c8bd21`) moved interpreted loops into
compiled code. Regenerating every shard from it under R 4.5.0 changed no value
that the earlier 14.0 build (`7b07a83`) returns: the 37 entries that differ
from the previous cache reproduce identically on both builds, so they had been
written by another R installation (an R that evaluates `x^3` as `x * x * x`,
for one). Generate the cache with R 4.5.0, as the scope note above requires.
14.2 package above under R 4.5.0 and carry `nns_version` `14.2`: every entry was
recomputed with `scripts/regenerate_r_cache.py --fresh`. The cache builder's own
regeneration for the same commit (`automation/r-cache-nns-14.2`, `bb05d3d`)
agrees with it on every entry the two share. To reproduce, install the package
and point `R_LIBS` at its library when regenerating
(`R_LIBS=<lib> python scripts/regenerate_r_cache.py ...`).

Earlier, the R 14.0 update of 13 November 2026 (`9c8bd21`) moved interpreted
loops into compiled code without changing a value that the first 14.0 build
(`7b07a83`) returned; 37 entries that differed then had been written by another
R installation (an R that evaluates `x^3` as `x * x * x`, for one). Generate the
cache with R 4.5.0, as the scope note above requires.

## Automated parity check (with deferred autofix agent)

Expand Down
Loading
Loading