python 3.15.0rc2: align python-feedstock with conda-forge

[skupr-anaconda]

Destination channel: ad-testing/label/py315

Links

• Jira
• Upstream repository
• conda-forge python-feedstock main @ 3.15.0rc2 (#885, #928)
• Debug spike (not this PR): #258 / PKG-2624

Explanation of changes:

Align the 3.15.0rc2 recipe with conda-forge main, reconstructed on origin/master after #257. Conflict merge 2a733207 (stray {% endif %}) was discarded.

Outputs / install split

• New libpython (install_shared.sh) and libpython-static (install_static.sh). Deleted unused build_static.sh / build_static.bat.
• python does not install libpython*.so/.dylib; tests assert those files are absent. Shared libpython is the libpython output.
• CF Add a libpython3.dylib to support ABI3 applications conda-forge/python-feedstock#885: osx libpython3.dylib stable-ABI reexport in install_shared.sh (blob 5dafbd13). Tests: libpython3${SHLIB_EXT} on unix release; libpython3.dylib on osx.
• CF add run-export to libpython conda-forge/python-feedstock#928: libpython unix weak run_exports pin_subpackage("libpython", min_pin="x.x", max_pin="x").
• cpython noarch metapackage; python-freethreading / python-gil / python-jit kept. python_site_packages_path set. ignore_run_exports_from cxx on libpython outputs (unix). libpython-static force_use_keys: freethreading, build_type.

CBC / variants (recipe/conda_build_config.yaml)

• gil_type → freethreading: yes/no and PY_FREETHREADING. FT is built on all platforms (s390x / osx-64 skip removed).
• build_type: release only. Unix debug is not in this PR — exclusive to #258 / PKG-2624.
• Still clobber aggregate python/numpy/is_freethreading/is_python_min; tk 9.0; osx clang 21.

recipe/meta.yaml

• build_number 1.
• source: is url + sha256 + patches only (no from_source_control / git_url wrapper).
• Drop gil_type, PY_GIL_DISABLED, bootstrap, linkage_nature.
• Win FT patches 0026 and 0030 wrapped in {% if freethreading == "yes" %}.
• Patch list matches master numbering: 0028 conda-site.pth, 0029 zlib-ng, 0030 xxlimited. 0023-branding.patch always applied.
• CMake tests use Python_FIND_ABI (cmake_test_args). Win import-lib tests cover both lib\ and libs\ (CFEP-27 / CF keep import libraries in libs to help meson and other build systems conda-forge/python-feedstock#918).
• License stays PSF-2.0. Dummy channel_targets tuple for the linter; 'conda-forge' in channel_targets is false so CF skip_compile_pyc / branding / downstreams jinja does not fire.

Build scripts

• build_base.sh: shared vs static split (SYSCONFIG / build-details.json from shared build, CF LDLIBRARY is set to libpython3.9a although no static lib is included in mambaforge for Linux conda-forge/python-feedstock#565). llvm-tools-21 via $HOST-llvm-ar → llvm-ar-$LLVM_VER. Dropped dead --with-tcltk-includes / --with-tcltk-libs (unrecognized in 3.15.0rc2; _tkinter via pkg-config). Dropped CentOS OLD_HOST / conda_cos6 / conda_cos7 sysconfig clones; system-gcc rewrite uses $HOST. JIT still off for debug, FT, and non-*-64.
• build_base.bat: PY_GIL_DISABLED → PY_FREETHREADING. Win tk still TclVersion=%tk% / TkVersion=%tk% (tk 9). CF cross MSBuild block left commented (AR native workers).

Tests / other

• Prefix-replacement: CF case $PKG_NAME (static or shared), not rm *.a then relink.
• run_test_jit.py: skip (exit 0) when this interpreter was built without JIT; fail if JIT exists but is not enabled. python-jit still skipped for FT.
• .gitignore: drop .vscode/.

Frozen AR-delta vs conda-forge (intentionally not copied)

• tk 9.0 from pkgs/main (not CF 8.6 / %THREAD% tkinter lib).
• llvm-tools-21 (not unversioned llvm-tools).
• No _python_rc; no channel_sources: python_rc / python_debug zip.
• Pinned aggregate host deps; msys2-gcc-libs; no binutils_impl / patch / m2-patch.
• downstreams commented; cmake-no-system / ninja-base / make.
• Keep 0023 branding, 0029 zlib-ng, 0030 xxlimited; linux sysroot missing_dso whitelist.

Notes:

• Upload dest is ad-testing/label/py315 only. Do not merge to pkgs/main.
• Human review/merge only when PBP is green.
• Linter parsing_jinja_failure is skipped (freethreading / dummy channel_targets). Graph generate must succeed; the linter skip may stay red.
• Debug builds are off here. Work them on #258; they will not go to pkgs/main.
• Linux libpython3.so is empty upstream (cpython#104612); we follow CF and only test libpython3${SHLIB_EXT} on unix release.

[skupr-anaconda]

Interesting, there is a new vendored d3-flame-graph in v3.15.0rc2: CPython stdlib profiling/sampling/_vendor/. . The upstream source is this https://github.com/spiermar/d3-flame-graph

[skupr-anaconda]

@ifitchet @cbouss

Artifact compare vs conda-forge python 3.15.0rc2

Staging: https://staging.continuum.io/pbp/fs/python-feedstock/pr256/cffd5c1
Pairs: GIL _101_cp315, FT _1_cp315t on linux-64, linux-aarch64, osx-arm64, win-64, noarch.

osx-64 is out of scope. win-arm64 is not in this graph (recipe is ARM64-ready; PBP has no win-arm64 worker yet).

Same

• Same outputs, same extension-module names.
• Unix python: ~1164/2814 files hash-identical (stdlib .py, headers, data).
• Win python: 1953/2013 GIL, 1946/2009 FT hash-identical.
• Remaining Unix “diffs” are almost all same-size .pyc (magic 520e0d0a, timestamp only).
• Empty on both sides: win libpython / libpython-static, noarch cpython / python-gil / python-freethreading.
• python-jit payloads match.

Not the same (review these)

What	Staging	conda-forge	Action
Linux linker-compat dir	compiler_compat/{ld,README}	share/python_compiler_compat/	It's historical. Optional: move to CF path
GIL conda-site.pth	missing	present	Keep missing. Do not restore CFEP-65 lib/python/site-packages (numba failure)
FT conda-site.pth	1 line	2 lines	Keep 1 line
platform.py / sys.version	Anaconda branding	conda-forge branding	Keep 0023-branding.patch
_tkinter	libtcl9tk9.0	libtk8.6	Keep tk 9
osx-arm64 libpython*.dylib	11.3 MB + __llvm_prf_*	6.6 MB, no PGO sections	Look — leftover PGO on the shared dylib

Shebangs / _sysconfig* / conda-meta differ by prefix (croot/ vs conda-bld/). Cannot match across CIs.

Why binaries differ (do not “fix” to CF)

• tk 9 vs 8.6, glibc 2.28 vs 2.17, libgcc 14 vs 15, libffi 3.4 vs 3.7, sqlite vs libsqlite, xz vs liblzma, no _python_rc, license PSF-2.0 vs Python-2.0.
• linux-64 CFLAGS: -march=x86-64-v2 (aggregate) vs CF -march=nocona. .comment: gcc 14.3 vs 15.3.
• linux pyconfig.h: extra HAVE_* (statx, getrandom, copy_file_range, …) from el8 vs cos7. osx: only HAVE_READLINE_READLINE_H.
• linux python3.15 ~40.4 vs 40.3 MB (DWARF, not a file-list bug). Win .pyd ~8 KB larger (MSVC + tk 9).

osx-arm64 libpython (the one real size bug)

python3.15 is fine (~6.57 MB both). libpython3.15.dylib is not:

• Staging __TEXT,__text 6.46 MB vs CF 3.86 MB.
• Staging has __llvm_prf_cnts/data/names/vnds (~2.0 MB). CF has none.

Cause: both configures get --enable-optimizations, but only the static tree runs make profile-opt. Shared make leaves PGO instrumentation in the dylib that install_shared.sh ships. Same story for libpython3.15.nolto.a (21.5 vs 14.6 MB).

Fix (follow-up, not required for file-list parity): do not pass --enable-optimizations into the shared configure, or run/strip PGO on that tree before cp libpython*.dylib.

Bottom line

File names: aligned except linux compiler_compat/ and GIL conda-site.pth.
File bytes: will never match CF while we keep tk 9, glibc 2.28, x86-64-v2, branding, and PBP prefixes. That is the frozen AR-delta, not a regression.

Optional in this PR: relocate compiler_compat. Do not chase CF bytes. Worth a small follow-up for the osx shared-lib PGO leftover.

[skupr-anaconda]

Artifact compare vs pkgs/main python 3.14.7

Staging: https://staging.continuum.io/pbp/fs/python-feedstock/pr256/cffd5c1 (3.15.0rc2)
Main: https://repo.anaconda.com/pkgs/main (3.14.7) · recipe main-3.14
Pairs: GIL _101, FT _1 on linux-64, osx-arm64, win-64. Paths version-normalized (3.14/3.15 → X) before set-diff.

osx-64 is out of scope. win-arm64 is on 3.14.7 main, not on this graph (recipe is ARM64-ready).

Same

• Unix compiler_compat/{ld,README} (both AR historical path).
• libpython-static: same 3 files (libpython3.X.a, .nolto.a, osx config .a).
• Host ABI family: tk 9.0.4, glibc 2.28, libffi 3.4, libgcc 14, openssl 3.5.x.
• Same extension names except the 3.15 stdlib adds below.
• Win: same ~37 .pyd names (after the CFEP-27 path rewrite).

Not the same (review these)

What	Staging 3.15	pkgs/main 3.14.7	Action
Shared libpython	new output; python has no .so/.dylib	so/dylib inside python; no libpython pkg	Keep CF-align split
GIL conda-site.pth	missing	present	Keep missing (numba / CFEP-65)
bin/python3.1	absent	present	3.14 conda#10969 workaround; optional
_sysconfigdata_*_conda_cos6_*	absent	present	Keep dropped CentOS clones
sre_{compile,parse,constants}.py	absent	present	Expected — removed upstream in 3.15
_asyncio*.so/.pyd	absent	present	Expected — folded into interpreter in 3.15
Lib/profiling/sampling/ (~81 files)	present	absent	Expected — new in 3.15
_math_integer, xxlimited_3_13	present	absent	Expected — new 3.15 modules
Win layout	lib/python/ + lib-dynload/	Lib/ + DLLs/	Keep CFEP-27
cpython noarch	present	absent	Keep CF-align metapackage
win-arm64	no artifacts	GIL + FT on main	PBP pool, not recipe

Shebangs / _sysconfig* / conda-meta differ by prefix and by version. Cannot match across CIs or minors.

Why binaries differ (do not “fix” to 3.14)

• Different CPython (3.15.0rc2 ≠ 3.14.7). Interpreter / .so bytes will never match.
• Pin floors only: libexpat 2.8.5 vs 2.8.2, sqlite 3.53.4 vs 3.53.2, ncurses 6.6 vs 6.5, openssl 3.5.8 vs 3.5.7. Same family as 3.14.7, not CF 8.6/glibc 2.17.
• C-API headers renamed (lock.h → pylock.h, typeslots.h → slots.h).
• pkgconfig filenames carry the version (python-3.15*.pc vs python-3.14*.pc).

win-arm64 (recipe yes, this graph no)

build_base.bat already sets PLATFORM=ARM64 / BUILD_PATH=arm64 (arm64t for FT) — the 3.14 #240 port. No extra LLVM override (3.15 upstream is 21).

PBP /check-queue lists win-amd64 only. Perseverance-skills: win-arm64 “(soon)”. 3.14.7 on main proves host deps work (ucrt, vc14, tk 9, openssl 3.5). This SHA never scheduled that subdir.

Bottom line

File names (Unix, version-normalized): aligned except CF-align split (libpython out of python), dropped cos6 sysconfig / python3.1, and 3.15 stdlib adds (profiling/, _math_integer).
File bytes: will never match 3.14.7 — different interpreter. That is expected, not a #256 regression.

Win is a layout rewrite (CFEP-27), not missing modules. Optional later: win-arm64 once PBP has a worker. Do not chase 3.14 paths or CF bytes.

[skupr-anaconda]

Review #259 first (C16+C17 libpython split vs master).

This PR (#256) is now based on PKG-15601-libpython-split so Files should be remainder-only (FT CBC / cpython / CF). Do not merge this before #259. After #259 merges, retarget this PR back to master.

Dest remains ad-testing/label/py315. Do not publish 3.15 to pkgs/main (PKG-15601 PBC).