[Paullo9]

Requested by Paul · project thread

Part of #9

Test only; no number moves. This pins two findings from the demo run as a new test file, DemoEDXAutoIDPinTests. Data: the public AlMgSi demo pools (edx-demo-pools.json), Poisson counts. The truth is planted on synthetic data: Al, Mg, Si, O, stray Cu K and sum peaks, and no Ta. Everything stays validation: "none".

False Ta on an unclaimed Si K-alpha

• ElementProposer().proposeOnRefinedAxis proposes no Ta on any of the five pools (L5, L6, L7, recipe A, matrix map), and Ta is never a sum-peak question. The 4.3 L_D Ta proposal from the demo run does not happen on the kit fixture at main e91a75a. The same holds on the feat: self-pair half bound on sum-peak questions and the opt-in overlap check #54 head, with the overlap check on or off.
• Recipe A: Ta M-alpha (3.3 L_D) and Si K-alpha (4.4 L_D) tie for the 1.727 keV peak, and each reason names the other. Both stay possible only.
• Known issue, pinned with withKnownIssue: because of that tie, the planted Si is not proposed on recipe A. The Ta trial is not withheld by the confirmation rule. The fix is in the loop (ElementProposer.swift), and three open PRs already touch that file, so it waits.

Mg and O missed in the matrix pools (planted near 1 L_D)

• At the default bars neither element is a candidate on L5, recipe A or the matrix map.
• The cheap experiment: drop the residual bar from 5 sigma to 3 sigma and the low-energy factor from 1.5 to 1. The matrix map then nominates Mg (1.03 L_D) and O (2.05 L_D), and both fail the gain bar ("improves the fit by 12" and "by 47", where a new element needs 100).
• So the miss comes from the published loop bars, not from a fit bias. On the noise-free expected spectrum x100, every pool proposes Mg and O. The hypothesis would have been refuted if the two elements had stayed absent at the lower bar.

Test evidence

• swift test -c release --parallel --num-workers 1 -j 1 --skip SpectroKitParityTests: green, 301 tests with 4 known issues (one of them new here), plus the other targets.
• This pins today's loop. If feat: self-pair half bound on sum-peak questions and the opt-in overlap check #54, feat: auto ID explains candidates whose peak went to a detected sum peak #61 or the line-weight PR changes these outcomes when it lands, update the pin in that PR.

Review: none. Test only, no numerics or public API change.

🤖 Generated with Claude Code