Repository navigation
feat(verify-pyprocar): validated references and fresh status tables - #297
Conversation
fc701e9 to
c3dee63
Compare
|
VERDICT: FAIL Reviewed SHA: c3dee63 (base origin/dev bf65596, which the head already contains, so head merged with dev is the head itself) Blocking finding (one, small)
In a linked worktree, Non-blocking findings
Gates at head (exit codes)
#287 references (live)
#288 read-only fixtures (live)I made a private reflink copy of every fixture (codes, examples, io, issues) in my worktree and pointed the base worktree's
#290 status tablesThe driver results at head (dev bf65596 plus skill-only changes; the tables say d6d4aaa, one library commit earlier) match CLAUDE.md rowsBoth rows are accurate once the operator lock has run. Until then, existing fixtures stay writable (doctor shows 2223), as the PR's Scope section says. |
0e3d2f0 to
046111a
Compare
|
VERDICT: PASS+NOTES Reviewed SHA: 046111a (its merge base is origin/dev 3ae5842, so the head already contains trunk) Both round-1 blockers are fixed. Gates at head (exit codes)
fetch guard, attacked liveI built scratch repos inside my worktree in three layouts:
Read-only fixtures (#288)I made a private reflink copy of
References (#287)
Acceptance
Notes (non-blocking)
|
|
VERDICT: FAIL Reviewed SHA: 046111a. origin/dev 3ae5842 is the merge base, so the head is already merged with trunk. Blocking finding
What I ran (each gated on exit code)
Acceptance
Notes (non-blocking)
|
…rbit references Each reference in scripts/lib/references imports no pyprocar, has an analytic validate() with a negative control that raises ValidationError also under python -O, and a ref_*.py driver whose first step runs it. references/check.py runs the three validations without a fixture.
…d list the references Every feature driver was rerun at d6d4aaa, and features/README.md and the per-feature status lines now match the runs. The fermi2d driver attaches the projection with fs.set_values and checks that the slice is colored.
fe881ed to
9f274e1
Compare
9f274e1 to
f7044d5
Compare
|
VERDICT: PASS+NOTES Reviewed SHA: f7044d5. Its merge base is origin/dev 5884b80 (which includes #299), so the head already contains trunk. I found no defect in scope. The round-2 UX1 survivor now fails Gates at head (exit codes)
No #288 left
References on dev's library (5884b80)
The ref_spin_ibz change is a real check
Acceptance
Notes (non-blocking)
|
Why
Two verify-pyprocar backlog items. The independent references behind the #280 to #282 verdicts lived only in session scratch space (#287). The feature status tables still said the legacy functions crash (#290). The read-only fixture work (#288) moved to #303 after two review rounds, so this PR stands alone on dev.
What changed
scripts/lib/references/holds three references with no pyprocar import:unfold.py(Popescu-Zunger weights from raw PROCAR phases),spin_ibz.py(spglib-reduced non-collinear mesh plus the axial spin transform) andorbits.py(16-cell tiled marching-cubes cut orbits). Each has an analyticvalidate()with a negative control, aref_*.pydriver whose first step runs it, and an entry in the feature files.references/check.pyruns all three validations without a fixture.require, which raisesValidationErrorunderpython -Otoo. The asserts left in the drivers only narrow a type, and the next line fails onNoneanyway. The unfolding validation includes the non-normal shear[[1,0,0],[1,2,0],[0,0,1]], whose translations differ between the row and column conventions.features/README.mdplus the per-feature status lines now match the runs. The fermi2d driver attaches the projection withfs.set_valuesand checks that the slice is colored.ibz2fbzpermutes atoms and rotates orbitals on each image, soref_spin_ibzchecks that channel 0 keeps its total per band (2e-15) instead of an exact copy. The spin and orbit numbers in the feature files are remeasured at dev 5884b80.Scope
Blast Radius
Skill scripts and docs only. No library or test code changes, and nothing touches
data/.Verification
verify.sh exec python .claude/skills/verify-pyprocar/scripts/lib/references/check.pyexits 0, with and without-O. A first-zone test on fractional G failsorbits.validate()(fcc_zone_outside_unit_cube: 0 orbits, not 1). Translations built asbox @ inv(M).Tnow failunfold.validate()on the non-normal shear (the k-phase control stops failing, 0.0296 < 0.1), with and without-O.ref_unfold,ref_spin_ibz,ref_orbits,fermi3dandbs2dexit 0;bandsandfermi2dexit 1 at the documented steps, and the fermi3d and bs2d numbers match the feature files. Unfolding weights matchebs.unfoldto 8.9e-16 over all 150 x 80 values. The unfolded BiSb spin equals the reference to 2.8e-6 on generic orbits and the full mesh to a 0.0032 mean. All 25 van Alphen cuts match the 4 printed decimals. Before the fix(ebs): make ibz2fbz unfold magnetic groups, phases, atoms and orbitals correctly #299 fix,ref_spin_ibzfailed on dev withchannel 0 is copied unchanged from the IBZ source.ruff_new_violations.pychecks, the F821/NPY201 check, andred_green.py(no test functions added or changed).Closes #287
Closes #290
Refs #288 (moved to #303)