Fix pathological alignment of dead modules for CVH

[davidwalter2]

Motivation

The data/MC comparison of the CVH refit efficiency (Z→μμ, pT 25–65 GeV) shows two localized regions where data is less efficient than MC, both absent in MC because MC uses the ideal geometry. Both trace back to the same origin: modules that were off during data-taking collect no hits, so their alignment DOF are unconstrained and the persisted constants can be arbitrary. The garbage constants then break the Geant4e propagation of the CVH refit for tracks crossing those modules — the muons themselves are perfectly reconstructed by standard tracking (full hit complement, high purity); only the refit aborts.

	Hotspot 1	Hotspot 2
location	η ∈ [0.1, 0.65], φ ∈ [0.80, 1.00]	η ≈ −1.6, φ ≈ −1.05
effect	up to 41% efficiency loss	3.2% refit-failure rate (vs 0.13% baseline)
module	TIB glued 369141860, mono face 369141862	TID glued 402666796, r-φ face 402666798
pathology	face tilted 0.75 rad	face translated 1.3 cm (5.4 mm out of its own module plane)

Both pathological modules turn out to be faces of glued pairs whose partner is sane, which is what the final design exploits.

What this PR does

Detection

Two independent flags per module, both untracked parameters with a default, 1e9 = off:

• orientation — aligned-vs-ideal tilt above gluedGarbageTiltThreshold (default 50 mrad; genuine corrections are < 1 mrad, the pathological one is 0.75 rad).
• translation — deviation from the local consensus, i.e. the median aligned-vs-ideal displacement of same-layer neighbours (|Δz| < 15 cm, |Δφ| < 0.6), above moduleGarbageShiftThreshold (default 4 mm).

A plain aligned-vs-ideal position cut cannot separate garbage translations from legitimate coherent large-structure movements (up to ~1 cm), hence the consensus. Relative to it, genuine module-level misalignment is sub-mm over the whole tracker (p50 = 0.29 mm, p99 = 1.3 mm, p99.9 = 2.0 mm of 16588 det units), and exactly six modules exceed 4 mm in the 2016 UL alignment: 402666798 (the only one inside the muon acceptance) plus five dead FPix edge modules at |η| ≈ 2.7–2.8. Exactly one composite exceeds the tilt threshold: 369141860.

A glued-pair-local criterion was considered as a simpler discriminator and does not work: over the 3288 glued pairs, |disp_mono − disp_stereo| has p99.9 = 1.6 mm and a largest genuine value of 2.1 mm, against 5.5 mm for the garbage face — no clean separation. The consensus scan is needed anyway, since 5 of the 6 translation outliers are pixel endcap modules with no glued partner at all.

Repair

The two faces of a glued module are a single rigid body, so a sane face determines both the bad face and the composite plane exactly. Both guards therefore share one repair path: when exactly one face of a glued module is flagged (by either criterion), the bad face and the composite are rebuilt as

anchor_aligned ∘ anchor_ideal⁻¹ ∘ (ideal surface)

with the sane face as anchor. Composite↔face relative position, orientation and the Rglued Jacobian then take their ideal values by construction, which is what the downstream plane reconstruction and the alignment parameter mapping assume.

Ideal surface + consensus displacement (with ideal orientation) survives only as the fallback for modules with no local reference: non-glued modules, and the so-far-unobserved case of both faces bad.

Hit ordering

The mono/stereo inner/outer decision for a split matched invalid hit read the raw alignment constants. Since 402666798 sits 5.4 mm out of its module plane — more than the physical face separation — that dot product came out with the wrong sign, and the two dead-channel placeholder hits were emitted back-to-front relative to the surfaces the fit then uses, which the forward-only propagation cannot traverse. Both producers now take those positions from surfacemapD_, the surfaces the fit actually uses.

Applied in both the single-track and two-track producers. Both guards are automatically inert on MC (useIdealGeometry → zero displacements).

Validation

Run2016G Charmonium ALCARECO, 20k events, aligned geometry, single-track producer:

	propagation aborts	on 369141862	on 402666798
guards off	41	31	9
guards on	1	0	0

The single remaining abort is on FPix 352467464, which neither guard flags — a separate, pre-existing case.

Geometry-level closure, composite vs its sane face, in the sane face's frame:

	402666796	369141860
uncorrected	2.5 mm out-of-plane	0.8 mm, + 0.75 rad face rotation
ideal + consensus, ideal orientation	0.57 mm, 2.7 mrad	0.41 mm, 1.6 mrad
anchored on the sane face	0, exact	0, exact

Run2016G SingleMuon, 6 files, 40k events (earlier iteration of the guards, hotspot-2 focused):

• Hotspot-2 refit failures 17 → 7 = the level of the mirror-η control cell (4); all 10 failures on 402666798 gone
• +10 recovered tracks in the cell, healthy (16 valid hits, pT ≈ 42 GeV)
• All other tracks: 50 812 matched, 99.78% bit-identical momentum; the remainder cross repaired modules, |Δ(q/p)| ≲ 10⁻⁵ at p99.9
• The re-included "hit" is the dead-channel placeholder — the hit-type census confirms the r-φ face was recorded inactive by the conditions data (its stereo partner is alive with normal hit rate), so no measurement ever existed

Notes

• With these two guards, and no other module deviating by more than 3.3 mm inside the muon acceptance, hotspots 1+2 are the only alignment-induced CVH efficiency holes in the 2016 data.
• The effect on the analysis appears only at the next reprocessing; for the current dataset the measured data→MC efficiency SF stands.
• In the CMSSW 15 workflow with the global material model, skipHitlessSurfaces removes the hotspot-2 class by construction (dead-module placeholders are never propagation targets); the guards remain complementary there for garbage constants on modules with valid hits.
• The corrupted constants themselves (in particular 402666798, −1.3 cm in z) should be reported to tracker alignment.

Review history

Josh asked whether the translation guard fires for glued modules at all, noted that if it does then rebuilding the composite at ideal position + consensus displacement is not right, and asked that the composite be tied to the good side instead — since the composite must keep defined position/orientation relations to the two sides or other alignment logic breaks. The third commit answers that: it does fire for one glued module, the composite is now anchored on the sane face by the same rule as the tilt guard, and the two separate composite rebuilds collapse into one. Fixing the split-hit ordering at its source then made reorderGarbageShiftHits provably the identity (9 calls over 20k events, all on 402666798, no permutation in any of them), so it was removed together with garbageShiftModules_, the garbageShiftReorderHits policy and the drop branch — the per-track hit loop no longer touches a hash set. Net for that commit: 182 insertions, 218 deletions.

🤖 Generated with Claude Code