[Paullo9]

Requested by Paul · project thread

Part of #9

Stacked on issue-8-pooled-sum-peaks (#59). This PR's base is that branch; retarget to main once #59 merges. It is stacked because ElementProposer.swift is already in two open PRs (#59 and the issue-9-autoid-leftovers branch). This is a third PR on that file, so expect a merge after those land. The touched regions differ: the loser loop, the candidate tiers and the notes.

What changes

Auto ID (proposeOnRefinedAxis, nothing listed) already gave the Al+Al sum peak (2.973 keV) its own column. The element candidates that lost that peak to it stayed in the possible tier, because only proposed candidates got the sum-peak question. On the Al-rich demo pools these were Ar K-alpha (15 to 27 L_D), Ag L-alpha (6 to 11 L_D) and Pd L-alpha.

Now a candidate that clearly lost its peak to a sum column is explained:

• ElementCandidate.explainedBySumPeak holds the sum label, and ProposalResult.explainedBySumPeaks is the new tier.
• "Clearly" means three things:
  • its peak gain is outside the ambiguity margin (30) of the sum's;
  • its total gain is no larger than the sum's;
  • the sum column stays in the final fit at or above its own L_D.
• Explained candidates leave possible. They are not proposed, not a sum-peak question and not withheld.
• A note names them, for example "Explained as pile-up: Ag, Ar, Pd lost the peak they were tested on to the Al+Al sum column".
• ProposalRules: an explained candidate no longer counts as a K peak in the beside-K rule (the peak is pile-up).

What it cannot do: one spectrum cannot separate a real line from the sum when the two are within the detector resolution.

• A planted Ar K-alpha of 1000 counts (about half the sum, 16 eV away) is still explained. Its peak gain is 105591, against 105800 for the sum.
• It keeps its sum-peak conflict and the two-beam-current remedy, so the question stays visible.
• At 4000 Ar counts it is a sum-peak question in the proposed tier.

Additive public API: one new init parameter (default nil), one new property and one new computed tier. Apps that list possible should also show explainedBySumPeaks; otherwise those candidates drop from the list. The app change is not in this PR.

Test evidence

The truth is planted on synthetic data, not a measurement: the public AlMgSi demo pools (mac4DSTEM demo-edx, sim_edx.py v2.0, edx-demo-pools.json from #59). No Ar, Ag or Pd is planted.

• Red first, before this change, on the Poisson counts of L5, M_A and M_matrix: Ar, Ag and Pd were in possible, with reason "The peak at 2.97 keV went to Al+Al sum".

• After:

  pool	proposed	possible	explained by Al+Al
  L5	Al, Si	none	Ag, Ar, Pd
  M_A	Al	Si, Ta	Ag, Ar, Pd
  M_matrix	Al, Si	none	Ag, Ar, Pd

  On L7 (pure Mg5Si6, no Al) there is no Al+Al column and nothing is explained by it.

• New test file: DemoEDXAutoIDTests (the three pools, a real Ar planted at 1000 and 4000 counts, and L7).

• Golden baselines: one line per dumped candidate (the new field) and three notes lines in eds-proposer-sim. No number changed. These are in their own commit.

  • In that scenario nine unplanted candidates are now explained by the Al+Mg, Al+Si and Si+Si sums: Rh, Pd, Cl, Ag, Ar, In, Sn, Cd and K.

• Fast gate: swift test -c release --parallel --num-workers 1 -j 1 --skip SpectroKitParityTests is green: 287 SpectroEDX tests, 2 known issues (unchanged).

• Parity: SpectroKitParityTests ran once, because the golden baselines moved. Green, 83 tests.

• One numerics review before the commit. It found no blocking bug. Fixed from it:

  • only a clear loss explains (the ambiguity margin and the total-gain guard);
  • the beside-K rule ignores explained candidates;
  • a planted real Ar test was added.

What remains for #9

• The overlap check (Ta M-alpha on the unclaimed Si K-alpha on M_A), which is in the issue-9-autoid-leftovers branch.
• Mn on the Al alloys, and the line tail beside a dominant K line (see Detector Al K-alpha tail calibration for Mg in a dilute Al matrix #60).
• UnlistedLineCheck still takes every candidate at or above L_D, explained or not. It runs on the quantification's own sum-peak decision (feat!: the pooled fit models sum peaks when the strongest line's self pair is significant #59), where the sum is a fixed explained part, so Ar and Ag are not candidates there. Not changed here.
• The apps' Auto ID panels need to show the explained tier.
• UI journey pending (kit change, no UI here).

🤖 Generated with Claude Code