fix(qe): reject unresolved cutoffs and close silent under-convergence paths

[k-yoshimi]

Summary

Make the QE input generator handle unresolvable cutoffs (ecutwfc/ecutrho) explicitly instead of silently writing 0.0 into the generated input.

The first commits (ba4b6c2, 70bbec9, f0595f0) implement the rejection of unresolved cutoffs; the follow-up commits (0bfeeef, b5771ae) fix the gaps found by two code-review rounds tracked in #13.

Closes #13

Changes

Reject unresolved cutoffs (ba4b6c2 ...)

• _find_elem_cutoff raises InputValidationError instead of falling back to 0.0 when neither the cutoff CSV nor the UPF header resolves a cutoff.
• Only the fields the user did not supply via content: are resolved (need_wfc/need_rho).

First review round (0bfeeef, #13)

• Normalize blank CSV cells (pandas NaN) and literal-0.0 UPF header attributes (common in ld1.x-generated files) to "not specified" (_normalize_cutoff). Previously they slipped past the is None checks and ecutrho = NaN / ecutwfc = 0.0 ended up in the generated input.
• When the CSV provides only one field, fill just the missing one from the UPF header instead of overwriting both.
• Fill ecutwfc also when the template omits the key entirely: it is mandatory for pw.x, and the input used to be written without it.
• Omit ecutrho when it cannot be resolved and let pw.x apply its own default (4*ecutwfc). It is optional in QE and normally absent for norm-conserving sets, so only ecutwfc remains a hard error.
• Point the error remedy at the real YAML path content.namelist.system.ecutwfc (content.system.* would be treated as a card).

Second review round (b5771ae, #13 comment)

• Floor an explicitly written ecutrho at 4 * the ecutwfc actually emitted (the resolved maximum, or the user-supplied value), so it never falls below the default pw.x would have chosen; rounding-level differences are tolerated so CSVs storing a rounded 4*ecutwfc keep their value.
• When the template carries no cutoff keys at all, resolve ecutrho too: filling only ecutwfc left ultrasoft sets running on the 4*ecutwfc default instead of the suggested ~8-12x.
• Fill only header-missing fields from the pp_info "Suggested" lines; an unset "0.0 Ry" line could reset a valid header wfc_cutoff to None and turn a resolvable pseudopotential into a hard error.
• Warn on unparseable cutoff values and on elements excluded from the ecutrho competition (both were silent).
• Reject pw.x tasks without a &system namelist instead of writing an input pw.x cannot run.
• Consult the UPF file only when a needed field is unresolved, and log an open failure as a warning: successful runs no longer emit ERROR noise.
• Align utils/pp_cutoff.py with the runtime convention: normalize 0.0/unparseable header values and emit a blank CSV cell (instead of dropping the row) when only rho_cutoff is missing.

Verification

• pytest tests/ → 261 passed (all fixes were implemented test-first).
• End-to-end CLI runs on sample/cif2x/qe/NaCl:
  • default case: generated cutoffs match scf.in_ref
  • blank ecutrho CSV cell: the ecutrho line is omitted, ecutwfc keeps the CSV value
  • missing cutoff row: clean one-line error with the correct remedy, exit 1, no file written
  • template without ecutwfc:: filled from the CSV; with ecutwfc: 100 supplied, ecutrho = 400 (the 4*ecutwfc floor)
  • task without &system: rejected with exit 1

Note: the QE test modules skip silently when qe_tools is unavailable; running them locally requires qe-tools<2 (2.x removed qe_tools.parsers.qeinputparser; pyproject pins ^1.1).

🤖 Generated with Claude Code

[k-yoshimi]

第2回レビュー（issue #13 のコメント参照）で確認された残課題を b5771ae で修正しました:

• 明示出力する ecutrho に「実際に書かれる ecutwfc の 4 倍」のフロアを導入（QE デフォルトを下回る値を明示出力しない。CSV の丸め差は許容）
• カットオフキーの無いテンプレートでは ecutrho も解決（US 擬ポテンシャルの推奨値を反映）
• pp_info の Suggested 行はヘッダーで欠けたフィールドのみ補完（0.0 行による上書き防止）
• パース不能値・競合から除外された元素に warning を追加
• &system 無しの pw タスクを InputValidationError で拒否
• UPF ファイル参照を「必要なフィールドが未解決のときのみ」に限定し、成功時の ERROR ログを解消
• utils/pp_cutoff.py をランタイム規約（空欄セル=未指定）に整合

テスト: 261 passed。NaCl サンプルで実機確認済み（参照一致・ecutwfc=100 明示時に ecutrho=400・キー無しテンプレートで両方補完・&system 無しで exit 1）。

🤖 Generated with Claude Code

[k-yoshimi]

Third review pass, this time with Codex as an independent strict reviewer (different angle from the two multi-agent correctness reviews). Three rounds to convergence; fixes landed in 329a7e7:

• Round 1 must_fix: "system" not in content.namelist raised TypeError when namelist is None (the Content() default). The guard now handles falsy namelists and raises the intended InputValidationError.
• Round 2 must_fix: a system: section that parses to None (bare key in YAML) passed the membership check and crashed the update helpers. The guard now requires a mapping.
• Round 1+2 should_fix: a user-supplied non-numeric ecutwfc reached 4.0 * ecutwfc and raised TypeError; numeric strings stayed strings and could render as quoted Fortran values. Non-numeric values are now rejected with a clear error, numeric strings are accepted and written back as floats.
• Also added a pp_cutoff regression test for an absent rho_cutoff attribute (blank CSV cell), confirming the intended output-contract change flagged by the reviewer.

Round 3 reported no remaining must_fix. Tests: 266 passed.