check_cell_symmetry: the Le Bail cell / space-group gate (PR C)

[drthyang]

PR C of the LLM-usage plan. The Rietveld skill's Stage 0 gate becomes a computed check instead of a judgement stated in words.

The question

Before any structural refinement, the tool answers two questions:

• Does the cell index every peak? If not, the cell or lattice is wrong, or a phase is missing.
• Does the data respect the space group's absences? If not, the group is too symmetric: it assumes a centring or glide the crystal lacks.

A low Le Bail wR can't answer either, because free intensities will fit a wrong group happily.

How (core/workflow/cellSymmetryCheck.ts)

1. A Le Bail fit refines the cell from peak positions alone, with free intensities (the existing leBailCellPrefit).
2. The pattern is extracted with the allowed reflections only. What remains (the leftover), minus a running-median baseline, is read in σ, with a 5% model-error floor so peak-shape mismatch under strong peaks doesn't count.
3. Absences: each forbidden family is tested at its own position. It's violated at ≥ 5σ, and untestable within 1.5 peak widths of an allowed reflection (or one from a named extra phase).
4. Indexing: a leftover peak (an apex over a whole peak width) that sits near no reflection is unindexed.
5. Peak shape: pseudo-Voigt for constant wavelength, since a Gaussian's missing tails read as shoulders. Gaussian for time of flight, with the width scaled by TOF.
6. Window: only d ≥ 0.7 Å is read by default. Below that, reflections are too dense to judge, and the reflection count grows as 1/d³.
7. Stated limits: the result's limits names what the check can't see: a cell that is too large (a supercell indexes anything), and a group with too few absences.

The MCP tool check_cell_symmetry wraps the check. It takes extraPhases, fitRange, dMin and significance. The skill's Stage 0 now calls it, and its "Dependency (MCP)" note is removed. LIMITATIONS and the generated AGENT_TOOLS table are updated.

Evidence (src/mcp/cellSymmetry.test.ts)

Synthetic data (pseudo-Voigt peaks with Caglioti widths, a sloping background, Poisson noise):

Case	Result
CsCl in Pm-3m (correct)	passes
CsCl declared Im-3m	fails: 100 violated at 87σ, plus 210, 111, 300 and more (all h+k+l odd)
Tetragonal crystal declared cubic	fails: 32 unindexed peaks
Two-phase data	fails with the second phase unnamed; passes once it's passed as extraPhases

Real data:

Case	Result
GaNb₄Se₈ at 28-ID, published F-43m (local data, skipped without data/)	passes, with 143 absences tested
GaNb₄Se₈ declared Fd-3m	the d-glide's forbidden 200 (16σ) and 420 (6σ) are caught
Mn₃Ga POWGEN 600 K (committed data)	MnO (111) is unindexed alone and indexed once MnO is named; P6₃/mmc's absences hold

Full suite: 195 files, 1,483 tests pass. tsc -b and eslint . are clean.

Worth knowing

• Unindexed lines on Mn₃Ga TOF. Even with MnO named, the check reports leftover lines 2–5% in d from any Mn₃Ga reflection, far beyond POWGEN's resolution. Two sit at the vanadium sample can's (110) and (200) positions (d = 2.151 and 1.523 Å), and one near Al (220). They look like real sample-environment lines, not false alarms. The test therefore asserts only the MnO behaviour, not a full pass.
• Tuning changes, each driven by a case above:
  • pseudo-Voigt for constant wavelength (GaNb₄Se₈'s tails had read as three marginal 5σ "peaks");
  • the pattern's full d range for forbidden families (bcc's 100 lies beyond the largest allowed d);
  • an apex over a whole peak width (noise on an unfitted extra phase's flanks).

🤖 Generated with Claude Code