| Age | Commit message (Collapse) | Author |
|
Both upstreams gained the allele labels: gemma-wrapper's geno2mdb.py now
writes the marker record as "name<TAB>allele1<TAB>allele0" and pangemma prints
those in the assoc's allele1/allele0 columns instead of '-'. Stores built
before this still read (the columns fall back to '-'), so no store has to be
rebuilt.
Verified with the packaged pair on the BXD fixture through LOCO: the store
built by bin/geno2mdb.py (Gf) and scanned by pangemma matches GEMMA on the same
BIMBAM on the header, on all 636 rows and on every column, allele1/allele0
included ('X'/'Y' in both). Note for the same workflow: a Gb store is exact
only for integer calls (0/1/2 -> 0/127/254), so a mean-genotype panel like BXD
needs Gf -- with Gb the same comparison differs on 62 beta / 48 se / 32 l_mle /
76 p_lrt of 636 rows.
|
|
pangemma-git builds the genenetwork fork of GEMMA, the one that reads an
lmdb genotype store. Built from its own premake project, as the fork's
guix.scm does (premake5 gmake, then make config=release with the lmdbxx and
lmdb include paths), with gsl, openblas, zlib, guile-3.0 and lmdb as inputs
and lmdbxx, pkg-config and premake5 to build. Tests are off because the
suite needs its example data and ruby.
It installs as `pangemma` and not `gemma`, so a profile can hold both engines
and pick per run -- which is what genenetwork4 does: the same workflow runs
either engine. The point of the separate binary is the store: `-g X.mdb`
reads genotypes from a memory map instead of re-parsing BIMBAM text, measured
on HLC as 141 s of cluster job time for a 23-chromosome LOCO against 958 s
for GEMMA on BIMBAM, with identical results.
gemma-wrapper moves from 156939f to b00c1da (revision 6): that range adds
bin/geno2mdb.py (the store builder that works without the ruby lmdb gem),
its two interop fixes, the cached-K/PLINK fixes, the drop-in version banner
and the one-thread-per-parallel-job cap. The comment above the package now
names each commit and what it buys.
One build fix, found by running the installed wrapper: ruby-build-system wraps
every file in bin/ with a ruby launcher, bin/geno2mdb.py included, so the
installed converter was a Ruby file with a .py name and `python3 $(which
geno2mdb.py)` was a SyntaxError. A phase after 'wrap restores the gem's own
copies, which carry the python shebang. Verified end to end from a scratch
directory with both packages: the packaged converter builds a 7320 x 198 Gb
store from the BXD fixture and the packaged pangemma scans it
(-loco 1 -gk, then -lmm 9 -maf 0.1: 636 analyzed SNPs, a 637-line assoc).
|
|
12c2a37 makes `Spearman' public. It was a private struct, so a caller
holding two value vectors could only compute the rank correlation by
going through `Compute' -- writing the dataset to a file and reading the
result out of a CSV. genenetwork4's /compute/correlation needs the
in-memory path, and its docs/correlations.md says why the p-value must be
the crate's rather than a rewrite.
The source now comes from https://git.genenetwork.org/correlation_rust
(the forge), where this is refs/heads/main; upstream on GitHub still sits
at 473ad82.
Hash note for future bumps: compute it from a *fresh clone*, not a
working tree. `guix hash -x` excludes VCS directories but not untracked
files, so a `output.txt` left behind by the crate's own tests changes the
hash and the build fails with a mismatch that looks like a wrong revision.
|
|
Pins revision 5, which adds --gemma-dir DIR / GEMMA_DIR: the directory that
holds gemma and the helpers the wrapper shells out to (time, tar, parallel,
and the PATH that tar's xz child gets) is named once instead of being
arranged as a PATH. Without it the failure arrives as
"No such file or directory - time (Errno::ENOENT)" deep inside
Open3.popen_run, or "env: 'tar': No such file or directory" from a shell --
both measured in the GN4AI container, where gemma-wrapper now lives in a
profile of its own rather than in the container's package list.
set-gemma-path's substitute pattern was `gemma_command = ENV.*`, which ate
the newline after the match. That was invisible while the line ended at
ENV['GEMMA_COMMAND'], and fatal the moment it gained a tail: the replacement
then glued the next line onto itself and Ruby refused the file
("unexpected local variable or method, expecting end-of-input"). The pattern
is now `gemma_command = ENV\['GEMMA_COMMAND'\]\[^\n]*`, which stops at the
newline whatever the tail becomes.
Built with its own `rake test` check phase green -- the LOCO pipeline and
the cache-hit assertions that gate this package.
|
|
python-htmlgen and python-asserts both build with python-build-system,
which used to pull setuptools in through its build phase; it no longer
does, so each package now names python-setuptools in native-inputs
rather than failing on a missing setup.py runner. Both still skip
their tests.
|
|
upstream gemma's check phase fails on a cold build, so the profile takes
gemma-exec-fix, which is upstream gemma with the check phase fixed. Its
package name is still 'gemma', which is why the path substitution below
still reads (assoc-ref inputs "gemma"); the comment says so, because
the mismatch looks like a bug otherwise.
javascript-purescript-genome-browser is disabled: its pinned commit
(93d45a5) is no longer served by GitHub, and the prebuilt js/ tree it
copied exists in no reachable commit (current master ships only
dist/{index.html,style.css}), so the source cannot be fetched. The
genome browser page then has no client-side JS.
|
|
d3js-7 fetched a prebuilt d3.min.js.gz mirrored on
files.genenetwork.org/software, which 404s now, and D3 ships no prebuilt
bundle in its GitHub release. Take the npm tarball
(registry.npmjs.org/d3/-/d3-<version>.tgz) and copy package/dist/d3.min.js
out of it; tar needs gzip on PATH to decompress, so both are inputs.
The genome browser's purescript bundle is a separate matter (see the
genenetwork package).
|
|
rpy2 3.6.2 and older import R_nchar, a C API entry point R 4.6 removed,
so _rinterface_cffi_api fails to load and gn-guile's rpy2 users cannot
start. 3.6.7 handles the removal.
Both source distributions (rpy2-rinterface and rpy2-robjects) live in
the one rpy2 git repository, so a shared origin fetches it once and each
package unpacks with a chdir into its own subdirectory. With that, the
split between inputs and native-inputs also needed fixing: icu4c,
libdeflate, zlib and r-survival are runtime inputs, python-cffi,
python-pytest and python-setuptools are build-time.
|
|
gemma-wrapper goes from 9f4ffff (revision 2) to 5eb53fa (revision 4), two
fixes that were hiding GEMMA's own errors and blocking the variance
components:
- a failing child's own ERROR/**** lines are printed before the Ruby
backtrace, and --loco without --input says what is missing instead of
dereferencing a nil JSON inside Dir.mktmpdir;
- each run's .log.txt is copied beside the archive. The logs were only
inside the .tar.xz, so reading GEMMA's 'pve estimate' meant unpacking a
multi-megabyte archive to see four lines. /compute/gemma in
genenetwork4 now passes -pve and reports pve/vg/ve per chromosome.
The hash is in nix-base32 (guix hash -f nix-base32): --format=base32
emits RFC-4648 base32, which Guix rejects, and because that rejection
happens at module-load time it reports 'gemma-wrapper: unbound variable'
rather than a bad hash.
gemma-gn2 needed Catch2 v2's single-header <catch2/catch.hpp>; Guix now
provides Catch2 v3, so 'make check' cannot be built here and the tests
are off, with the catch2 input commented out rather than left dangling.
|
|
python-scikit-allel was removed from Guix upstream.
|
|
|
|
JBrowse 2's web UI is a single-page app distributed as a prebuilt zip on
each GitHub release -- the artifact `jbrowse create' unpacks into a
user's install. Packaging it from source would mean vendoring ~1700 npm
packages for one webpack run of the pnpm monorepo (packages/, plugins/,
products/), and node-build-system has no pnpm support, so this packages
the released artifact instead: index.html, the static/ bundles and the
volvox demo data under share/jbrowse-web.
The app is configured at deployment time by placing a config.json next
to index.html; test_data/config.json is a remote-hg19 demo, while
test_data/volvox/config.json is fully offline.
|
|
Guix's stock nginx is built without the real_ip module, so
set_real_ip_from, real_ip_header and real_ip_recursive are rejected as
unknown directives. Deployments behind a reverse proxy need them:
rate limits keyed on $binary_remote_addr otherwise see every request as
coming from the proxy, so per-client budgets collapse into one bucket
shared by all clients (which is what produced the 429 storm on
git.genenetwork.org).
The package only adds the configure flag to nginx; everything else is
unchanged. Used by gn-machines' cgit container.
|
|
The stress view said only "All related sentences (Gemini classification
not available or no specific stress types found)" for a missing key, a
failed API call and a model that returned no labels alike -- so a
deployment where the key is present but the model is not callable looked
like "nothing to classify", with no hint that anything was wrong.
c8321ee separates the three cases on the page and in the log (no-key /
error with the exception, model and attempt count / no labels with the
raw response), names the model in one constant, quotes non-JSON replies,
and logs every retry plus the final traceback.
Hash verified against a fresh clone of git.genenetwork.org/genecup
(guix hash -rx), not taken on trust.
Together with the gn-machines boot check that calls the model, this is
what turns "classification not available" into a diagnosable failure.
|
|
|
|
9f4ffff (pushed upstream on top of 3a9286c) fixes the cache-dir
archives so repeated kinship/GWA runs actually hit cache: the
archives never included *.log.txt, which the cache-hit check requires
alongside the data file, and the GWA branch had no unconditional
archive-unpack step without --lmdb (unlike the kinship branch), so a
valid cached archive could never be found.
Verified by building the package: the upstream Rakefile's own `rake
test` (run by this package's check phase) now reports real CACHE HIT!
lines for both the LOCO kinship and GWA steps instead of recomputing
GEMMA every time.
|
|
|
|
Add julia-1.10: inherit the 1.8.5 recipe, bump source, use the tree's
LLVM 15 and libgit2 1.6, seed the three new deps tarballs (Julia LLVM
fork, JuliaSyntax, ittapi) into deps/srccache, pass
CMAKE_POLICY_VERSION_MINIMUM to ittapi, drop the CPLUS_INCLUDE_PATH
leak that shadows C headers, and skip the CUDA-only SuiteSparse libs.
Build now gets through LLVM and most of julia-base; it fails compiling
Julia's own src/pipeline.cpp against system LLVM, because Julia 1.10
needs its patched LLVM fork (copyable PassManager), whereas the recipe
is built to use system LLVM. Reworking that (build and use the fork's
LLVM, adapt the shared-objects-paths phase) is the remaining work.
|
|
julia-bin: patchelf every executable (bin/ and libexec/), not just
bin/julia -- the bundled 7z (used by Pkg to unpack registries) otherwise
has no interpreter and exec fails.
pluto-notebook: put curl and git on the wrapper's PATH so Pkg can
download the registry and packages on first run.
|
|
Replace pluto-notebook with a modern server: it uses julia-bin (Julia
1.10, which bundles its own libcurl) and installs the current Pluto
(1.0.4) / PlutoSliderServer (1.9.0) stack into a Pkg environment on
first run (the depot is expected to persist). The notebooks get their
embedded Pkg environments stripped and an inert Pkg.activate reference
added, which disables Pluto's package manager so they run against the
server's environment.
This drops the Julia 1.8 libcurl crash and the Pluto 0.19 export/watch
problems, so the workarounds are no longer needed.
|
|
Building Julia from source is a large, heavily-patched effort and the
version packaged in Guix is pinned to 1.8 (whose system-libcurl build
trips a Downloads.Curl crash in PlutoSliderServer). Add julia-bin,
which repackages the official 1.10.9 binary tarball: every ELF file is
patchelf'd to the store's glibc and gcc runtime libraries, and libunwind
and zlib are added to the run path.
The upstream tarball bundles its own libcurl, so Julia's Downloads will
not hit the system-libcurl bug. The existing notebook packages (Pluto
0.19.9 etc.) still need to be bumped to load on 1.10; see the header
comment.
(julia-bump branch.)
|
|
runsliderserver now keeps the static export in sync with the notebook
directory (a background loop re-runs the server-less exporter, whose
cache means only changed notebooks are re-run) and watches the
directory, restarting the slider server when a notebook changes. So a
user managing notebooks in their own (mounted) directory can edit them
and see the change picked up without restarting the container or the
service.
(The server's own directory watch cannot be used: it depends on
betterfilewatching, which is not packaged, so its callback is never
invoked and changes are never noticed.)
|
|
The one-shot export runs with the slider server disabled, so
PlutoSliderServer left slider_server_url undefined in the generated
HTML; the pages then could not reach the backend and the notebooks
could not be run interactively. Set Export_slider_server_url="." so the
pages post their state requests to the server (same origin).
|
|
The slider server aborted (SIGABRT) shortly after all notebooks were
ready, for two Julia 1.8 / PlutoSliderServer reasons:
* the default soft file-descriptor limit (1024) is too small for ~10
live notebook workspaces -- raise it in the runsliderserver wrapper;
* a periodic outbound HTTPS request (~20s after all notebooks are
ready) trips Julia's libcurl bug ("curlmultiassign: 1" from
Downloads.Curl) -- for the server, point the proxy variables at a
closed port so such requests fail at connect instead (the notebooks
run offline anyway).
Also extend the PlutoSliderServer patch so the server can serve a
pre-generated export directory even with export disabled, and so it
runs all notebooks (not only @bind ones) when PLUTO_SERVE_PREGENERATED
is set. With these, the server stays up (serving all notebooks) once
the one-shot export has completed.
|
|
PlutoSliderServer makes Julia 1.8 abort (SIGABRT) when it generates a
notebook's static export while the slider server is up. Patch it (via
PLUTO_SERVE_PREGENERATED) to skip the live export step, and rewrite
pluto-notebook's runsliderserver wrapper as a bash script that:
1. generates the static export once with a server-less, one-shot
run_directory call (retrying, because a cold depot can also make
Julia abort while precompiling), and
2. supervises the slider server, which serves that persistent export
directory and runs the notebooks, restarting it if it exits.
The notebooks and their pages now load and are served without network.
|
|
Each bundled notebook embeds its own Project.toml/Manifest.toml cells,
which Pluto tries to resolve (and download) at run time. Offline that
fails, and Pluto's package-sync additionally crashes on a libcurl error
("curlmultiassign"), taking the server down mid-run. Strip the embedded
environment cells so each notebook runs against the shared
JULIA_LOAD_PATH environment the container provides.
Combined with the aggregated load path and JULIA_PKG_OFFLINE, this makes
the packaged-environment notebooks (Plots/Optim/Distributions/...) run
and export; the power notebook test passes.
|
|
The wrapper listed every package's share/julia/loadpath separately, so
JULIA_LOAD_PATH grew to 332 entries / ~30 KB. Pluto spawns its
workspace worker with that environment, and the exec then failed with
E2BIG ("Argument list too long"), so no notebook could run. Build one
directory of symlinks to the packages and put that single entry on
JULIA_LOAD_PATH instead.
|
|
Add 14 pure-Julia packages the Makie/CairoMakie closure needs that Guix
did not have: ArgTools, ColorBrewer, ExproniconLite, Grisu, Match,
PolygonOps, SignedDistanceFields, IntervalSets, PkgVersion, Packing,
GridLayoutBase, Netpbm, QOI and Animations. All build.
SignedDistanceFields is old-style (no Project.toml) and gets a generated
one with its Statistics dependency. The JLL-backed packages and
Makie/CairoMakie themselves are still to come.
|
|
Add julia-latexify (0.15.18) and julia-plots (1.25.12). Everything else
Plots needs was already packaged (RecipesPipeline, UnicodeFun,
Formatting, GR, FFMPEG, GeometryBasics, ...), so only Latexify was
missing. Add Plots to the shared environment.
Verified in the container: Plots, Optim, Distributions, Roots, QuadGK,
PrettyTables, LaTeXStrings and PlutoUI all import from the shared load
path.
|
|
Add julia-optim and its unpackaged dependencies, staying within Julia
1.8 (the version Guix ships): ArrayInterfaceCore + ArrayInterface 6.x
(Static 0.8, matching Guix's), FiniteDiff 2.17, NLSolversBase 7.8.3 and
LineSearches 7.2.0 (Optim 1.6.2 predates its MathOptInterface
dependency). ArrayInterfaceCore is extracted from the
lib/ArrayInterfaceCore subdir of the ArrayInterface monorepo.
Also fix the shared environment: the wrapper only put the direct
notebook packages on JULIA_LOAD_PATH, so their dependencies were
missing (Distributions failed to precompile). It now prepends the load
path of the whole dependency closure.
Verified in the container: Optim, Distributions, QuadGK, PrettyTables,
LaTeXStrings, Roots, PlutoUI and Pluto all import.
|
|
pluto-notebook now serves the repo's notebooks/ directory (the legacy
sens/visuals notebooks) with the smoke-test notebook installed there,
so PlutoSliderServer lists all of them, instead of a smoke-test-only
notebooks-default/ directory.
|
|
Add pluto-notebook, superseding julia-visuals. It serves a directory
of Pluto notebooks with PlutoSliderServer, and provides a
runsliderserver that honours PLUTO_NOTEBOOKS_DIR and PLUTO_PORT (the
keys the julia-notebook service sets). A smoke-test notebook is
installed under notebooks-default/, which is what the wrapper serves by
default; the repo's own notebooks, each pulling a heavy environment,
stay in notebooks/.
Add the notebook dependencies that are packaged (PlutoUI,
Distributions, Roots, LaTeXStrings, QuadGK, PrettyTables) as inputs and
put each input's Julia load path on the wrapper's JULIA_LOAD_PATH, so
the notebooks share one Julia environment instead of each downloading
its own. Plots, CairoMakie and Optim are not packaged yet.
|
|
(gn packages gemma) imported both (gnu packages base) (the `which`
package, needed in gemma-gn2's native-inputs) and (guix build utils)
(a `which` procedure for locating executables). The later import won,
so gemma-gn2's `("which" ,which)` native-input silently became a
procedure, not a package, and guix rejected it outright:
guix shell: error: gn/packages/gemma.scm:45:2: package
`gemma-gn2@0.98.5-8cd4cdb' has an invalid input:
("which" #<procedure which (program)>)
Hide `which` from the (guix build utils) import so the package
resolves to the intended one. Verified: gemma-wrapper now builds and
runs under `guix shell -L . -D -f <consumer>/guix.scm`.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
Add python-bcrypt-5 (5.0.0), built with pyproject-build-system +
setuptools-rust + the cargo-build-system crate-vendoring phases, and
use it for genecup-gemini in place of the C-based python-bcrypt 3.2.2.
Guix's (gnu packages rust-crates) carries every crate in bcrypt's
Cargo.lock except the bcrypt crate itself -- which does not exist in
Guix -- and has no cargo-inputs entry for a Rust-based bcrypt. The
missing crate source and the lookup-cargo-inputs table entry live in
the channel's (gn packages rust-crates) module, next to gfaffix's
crates; every other crate in the graph is referenced from
(gnu packages rust-crates). Cargo.lock is dropped from the source so
cargo regenerates it against Guix's vendored directory sources.
Verified with a hashpw/checkpw round-trip and the end-to-end container
smoke test (11/11). Upstream's pytest suite is not wired into the
build (it needs the extension importable at check time), hence
#:tests? #f.
|
|
- python-google-genai 1.68.0 -> 2.28.0
- edirect 25.2.20260328 -> 26.3.20260928: drop the local
xml-bounds-check patch (upstream now performs the same checks) and
relax the go.mod directive to the available toolchain (go 1.26.6)
- add python-pytz-2026 (2026.5) and use it for genecup-gemini
nltk stays at 3.9.1: 3.10's new pathsec hardened-open refuses Guix's
hardlinked store files (st_nlink=2), which breaks loading the
nltk-punkt tokenizer data. bcrypt stays at 3.2.2: 5.0 is a
Rust/maturin build.
Verified with the edirect network tests (esearch/hybrid) and the
end-to-end container smoke test (11/11).
|
|
gn-guile cf32f16 ships the stripped branding webserver in its own
tree (web/webserver-aging.scm + bin/gn-guile-webserver), fixes the
doctype rendering and computes the copyright year at run time. To
make the full package build again at the new commit:
- propagate guile-config (web/config.scm uses (config), added
upstream after the previous pin)
- install bin/gn-guile from the repository (gn-guile.sh was removed
upstream) and replace wrap-script/wrap-program with a hand-written
wrapper: this sandboxed daemon's store overlay rejects the
hard-link that both wrappers create (ENOENT)
Hash from the daemon's git-fetch checkout. Verified: guix build
gn-guile; the gn-guile container serves the aging brand pages with
<!DOCTYPE html>, footer "Copyright © 2005-2026", css/images, all API
routes disabled, SSH on 2225.
|
|
- gemma: add gemma-exec-fix, a variant of upstream gemma with
#:tests? #f. Upstream's check phase (./test_suite.sh) cannot run
on build hosts where /tmp is mounted noexec ("Permission denied")
and cuirass.genenetwork.org has no gemma substitute, so every cold
build of gn-guile died on it.
- guile.scm: gn-guile now propagates gemma-exec-fix instead of gemma;
guile3-dbi and guile3-dbd-mysql get a fix-exec-bits phase
(chmod autogen.sh) -- the unpack of a directory source drops the
exec bit, which broke their bootstrap phase the same way.
|
|
Running a real search through the container (GRIA1 x function
ontology, anonymous flow) exposed four environment problems, all
fixed in the package:
- edirect's nquire shells out to curl/wget and join-into-groups-of
shells out to xargs -- add curl, wget and findutils (xargs) to the
genecup wrapper PATH ("nquire is unable to find either curl or
wget", "xargs: command not found")
- curl exited 60/77 (CA verification) in the minimal container: the
concatenated ca-certificates.crt bundle is generated per-profile
from nss-certs, so the wrapper exports SSL_CERT_FILE and
CURL_CA_BUNDLE pointing at the system profile's bundle
- xfetch/rchive require EDIRECT_LOCAL_ARCHIVE in the environment
("Configuration does not specify the required ARCHIVE path"); the
wrapper exports it pointing at <out>/share/genecup-archive, whose
pubmed symlink exposes the bundled minipubmed archive in the
<archive>/pubmed/Archive/<NN> layout xfetch expects (minipubmed
itself is laid out flat); hybrid_fetch_abstracts falls back to
NCBI efetch for anything the local archive does not have
- the more_functions.py startup check now accepts the flat
minipubmed layout as well as the full edirect layout (it demanded
<archive>/pubmed/Archive/00 exactly)
test/genecup_smoke.py in gn-machines runs the full flow: front page,
about, search kickoff, the /search SSE stream, cytoscape graph page,
table view, sentences page and static assets -- 11/11 against the
container (with anubis? #f during testing, #t restored after).
|
|
Port of the genecup guix.scm package definition into (gn packages
genecup): genecup-gemini 1.9 pinned to commit 205d331 of
https://git.genenetwork.org/genecup (git-fetch, like the other
GeneNetwork packages -- the guix.scm local-source variant stays for
development), plus its support packages nltk-punkt, minipubmed,
edirect-25 (with the xml-bounds-check patch moved into
gn/packages/patches/) and python-google-genai.
One packaging fix over guix.scm: the archive-existence check in
more_functions.py only knows the full edirect layout
(<archive>/pubmed/Archive/00); the bundled minipubmed test archive
uses the flat <archive>/Archive/<NN> layout and the server refused to
start without a full local archive. The package accepts both layouts
so the container boots and searches go over the network when no
archive is shared in.
|
|
gn-bnw c7c89dbb validates the GN2 payload before prep runs (header-only
payloads, ragged rows, single variable/sample, all-NA columns and rows,
non-numeric cells), normalizes CRLF and redundant row separators, renames
duplicate variable names with a note, defuses the runtime_check.php
"int + string" fatal, and adds plain-English hints to the structure
learning failure page. Hash from the daemon's git-fetch checkout.
Verified: guix build bnw (lfc4jpj8) contains all fixes; gn-bnw HTTP
suite 39/39 on the container.
|
|
gn-bnw f9d5c411 makes sourcecodes/run.sh locate the C binaries
(network_score, structure, k-best/) next to itself in the package
store item instead of cd-ing to /var/lib/genenet/bnw and depending on
the activation copy -- which was incomplete on production, killing
every structure-learning run with "./network_score: No such file or
directory". Hash from the daemon's git-fetch checkout.
Verified: guix build bnw (fk19kbbli) contains the self-locating
run.sh, version.php at commit f9d5c411, both php fixes and the C
binaries; bnw container suite 19/19.
|
|
version.php reports name, version, the gn-bnw commit this package was
built from and the origin URL as JSON (served at /version.php on the
site). This makes a mis-deploy visible at a glance: the earlier
production debugging session had to infer the running version from
error messages.
The origin is hoisted into a bnw-origin variable and the commit into
%bnw-commit so both the origin and the endpoint share one source of
truth (and future pin updates are a one-line change).
|
|
gn-bnw 605900dd sanitizes the GN2 handshake name row: whitespace
inside trait display names ("Age at death") becomes an underscore
before the separator conversion. Unsanitized spaced names corrupted
the input header (the octave parsers split on whitespace) and
structure learning died with "A(0,_): subscripts must be ... integers"
in mk_adj_mat.
Hash from the daemon's git-fetch checkout (authoritative). Note: the
commit hash on the remote differs from the local one after the first
8 characters (605900ddc699...) -- the remote is authoritative.
Verified: guix build bnw (71q2g6ay) carries the sanitize fix; bnw
container suite 19/19.
|
|
gn-bnw e18ba415 wraps the run.sh/run_octave pipeline in try/catch and
renders "Structure learning FAILED: ..." with a back-link instead of
a blank page (the uncaught RuntimeException was the cause of the
blank network-diagram flow on production with real datasets). Hash
from the daemon's git-fetch checkout (authoritative; guix hash -S git
and guix download differ from guix's own checkout serialization).
Verified: guix build bnw (xa80znim) carries both php fixes; bnw
container suite 19/19.
|
|
aac5a06 updated the commit pin but kept the hash of the previous
checkout (3616c964), so the fetch failed with a hash mismatch. Use
the hash the daemon computes from the real git-fetch checkout
(guix hash -S git differs slightly from guix's own checkout
serialization; the daemon-provided value is authoritative).
Verified: guix build bnw (51733v44) contains the forwarding status
page; bnw container suite 18/18.
|
|
gn-bnw 1de3cbe0 makes execute_bn_gom.php emit a status page with
flush() before the C/octave runs ("Forwarding to the network diagram
when done..." plus a click-here fallback link), instead of outputting
nothing and leaving the browser blank for the whole run.
|
|
Remove the local-source test hook again: the package fetches from the
git origin on git.genenetwork.org, which is now pinned at gn-bnw
3616c964 ("bn_genenet: accept the GN2 template spelling My_Genenet",
needed for the GeneNetwork send-to-BNW flow). Hash from the git-fetch
checkout itself (guix hash -S git differs slightly from the real
checkout serialization; the daemon-provided hash is authoritative).
Verified: guix build bnw; bnw container suite 16/16 against the
rebuilt container (including the GN2 legacy-path handshake).
|
|
When BNW_LOCAL_SOURCE points at a gn-bnw checkout, build the package
from that tree instead of the upstream git origin. gn-bnw's Makefile
exports the variable for `make test` (documented there) but the hook
was never implemented in the package, so `make test` silently built
the pinned upstream commit and source changes were not what got
tested.
|
|
- Pin gn-bnw at 00e5b05f5d: upstream dropped the vendored javascript
libraries and prebuilt binaries.
- Drop the origin snippet: the JS deletions now run in the
clean-artifacts build phase, which also covers local-file checkouts
(guarded with file-exists?).
- Fix unbalanced parens: one ')' missing to close the source field, so
build-system/arguments/inputs nested under source and a stray ')' at
EOF rebalanced the file; module load failed with "invalid field
specifier" and bnw resolved as unbound. Also drop the trailing paren.
Verified: guix build bnw; guix system container --network --dry-run;
bnw-container boots via run-guix-container with sshd, nginx and
php-fpm up and BNW served on port 8881.
|
|
Uncomment the vendored-javascript deletion list: the bundled
FileSaver, canvas-toBlob, cytoscape, d3, dagre, font-awesome, jquery
and lodash copies are replaced by guix packages via the
install-javascript-libraries phase (the original accordion.js,
create_table.js and table.css are kept). Verified with the full
container test suite, including the network rendering tests.
Drop the .so/.o/DLL/binary deletions from the snippet: those files
have been removed from the gn-bnw repository itself (build.sh
recompiles network_score, structure and the k-best binaries from
source and links libRmath from rmath-standalone).
|