Epic: STAAR/MultiSTAAR 1:1 parity with STAARpipeline + tutorial end-to-end

[vineetver]

Umbrella tracker for parity with xihaoli/STAARpipeline and runnability of STAARpipeline-Tutorial end-to-end via favor-cli. Closes when every subtask closes.

Tutorial blockers (must-fix to run tutorial end-to-end)

• GDS (SeqArray) genotype format support #69 aGDS input (SeqArray)
• generalize ingest reader: VariantReader trait for VCF/BCF/BGEN/GDS #87 VariantReader trait for VCF/BCF/BGEN/GDS
• Wire individual (single-variant) analysis producer #108 individual analysis producer wired
• Rdata I/O: phenotype input and per-mask output for STAARpipelineSummary #109 Rdata I/O for phenotype and per-shard outputs
• LD pruning step for known-loci selection (pre-conditional) #103 LD pruning pre-step
• Conditional STAAR single-study: _cond and _cond_spa paths #102 single-study _cond and _cond_spa
• Sparse ancestry-adjusted GRM builder (FastSparseGRM / Lin-Dey) #99 FastSparseGRM with EnsureGrm auto-invoke

Correctness parity

• Coding masks: add pLoF+disruptive_missense, align PTV+DS with STAARpipeline #77 disruptive_missense and ptv_ds predicates (MetaSVM)
• Add MetaSVM and structural annotation channels to annotation contract #107 MetaSVM and structural annotation channels
• Annotation channel catalog: lock 11 STAAR weight channels 1:1 with STAARpipeline #104 STAAR weight channel catalog lock
• Ground truth tests: validate SPA binary and AI-STAAR against R output #78 ground-truth SPA and AI-STAAR vs R

Analysis parity

• SCANG: implement Monte Carlo empirical threshold for FWER control #79 SCANG Monte Carlo empirical threshold
• staar2scang_nullmodel bridge: times and Omega on null model #110 staar2scang_nullmodel bridge (times, Omega)
• Multi-trait STAAR: add kinship/GLMM support for related samples #80 MultiSTAAR kinship
• MultiSTAAR binary (logistic) null without kinship #111 MultiSTAAR binary null
• GENESIS null bridge: accept externally fit null models #100 GENESIS null bridge
• Longitudinal (random.slope) and heteroscedastic (groups) null fits #101 longitudinal and heteroscedastic null
• AI_Individual_Analysis: ancestry-informed single-variant test #112 AI_Individual_Analysis single-variant
• coding_incl_ptv mask: coding union plus PTV #113 coding_incl_ptv mask union

MetaSTAAR

• Heterogeneous conditional MetaSTAAR #72 heterogeneous conditional meta

Supporting (not parity but unblocks iteration)

• Cache kinship-aware null models to disk #81 cache kinship null models to disk

Already present (confirmed)

GLM / Logistic / REML (dense and sparse) / PQL nulls; SPA path; AI-STAAR seeded ensemble weights; all 11 STAAR annotation weight channels; all 8 noncoding masks; sliding window 2 kb default; SCANG windows with configurable Lmin/Lmax; MetaSTAAR sumstats with canonical lex-min alleles; run manifest and resumability; score cache.