feat(chemistries): add 10x-flexv2-gex-3p-config-b preset

[an-altosian]

Summary

Adds a chemistry preset for 10x Genomics GEM-X Flex v2 sequencing Configuration B (R1=28 / R2=90), as documented in 10x's Sequencing Requirements for Single Cell Gene Expression Flex.

Users running Config B today have to override --geometry and --sample-bc-ori by hand (or use the new CLI flag from #199). This preset gives them a one-line --chemistry 10x-flexv2-gex-3p-config-b option that resolves all preset-controlled parameters automatically — geometry, sample BC orientation, cell BC whitelist, sample BC TSV, and per-organism probe sets.

Two fields differ from 10x-flexv2-gex-3p

All other assets (plist_name, remote_url, sample_bc_list.{plist_name,remote_url}, probe_sets.human/mouse, meta) are byte-identical to the Config A preset.

- "geometry": "1{b[16]u[12]x[0-3]f[TTGCTAGGACCG]s[10]x:}2{r:}",
+ "geometry": "1{b[16]u[12]x:}2{r[50]f[CCCATATAAGAAAACCTGAATACGCGGTT]s[10]x:}",
...
-     "sample_bc_ori": "reverse"
+     "sample_bc_ori": "forward"

Why geometry differs: In Config B, R1 ends at 28 bp (cell BC + UMI only — no room for the probe-side anchor TTGCTAGGACCG that Config A uses), and R2 reads 90 bp from the probe end. So the 29 bp constant CCCATATAAGAAAACCTGAATACGCGGTT (the RC of 10x's documented sense-strand AACCGCGTATTCAGGTTTTCTTATATGGG), the 10 bp sample BC, and the 50 bp probe-mappable region all live on R2 in that order.

Why sample_bc_ori differs: R2 reads the opposite strand vs R1's "view" of the library construct, so the canonical sample BC appears in the read in its forward-canonical form (whereas Config A reads the BC's RC on R1's strand).

Validation

End-to-end verified by running simpleaf multiplex-quant --chemistry 10x-flexv2-gex-3p-config-b --organism mouse --probe-set <mouse probe CSV> against an internal Config B Flex library and comparing against the corresponding cellranger ground truth:

• High overall mapping rate against the auto-built probe index, consistent with what other Flex presets produce on Config A data.
• Sample BC demultiplexing recovers exactly the wells declared in the cellranger multi config, with realistic cell counts per well; unused wells stay at the noise floor (≤1 cell / few reads).
• The same input run against the 10x-flexv2-gex-3p (Config A) preset produces noise-floor sample-BC matches only, confirming the sample_bc_ori/geometry split between the two presets is real, not a CLI quirk.

Relation to #199

#199 exposes --sample-bc-ori as a CLI override, which is the general fix for cycle-plan variants. This PR is complementary, not redundant: the preset gives Config B users --chemistry 10x-flexv2-gex-3p-config-b as a one-line alternative to remembering the right geometry + override combo. Both code paths were validated against the same dataset and produce byte-identical results, confirming the preset's declared fields and the CLI override flow into the same downstream pipeline.

Test plan

• JSON parses (validated via python3 -m json.tool).
• Preset entry is sibling-style appended after 10x-flexv2-gex-3p with the same field order — minimal diff (+28 lines), no whitespace churn on unrelated entries.
• End-to-end run with --chemistry 10x-flexv2-gex-3p-config-b against a real Config B Flex library: ~98% mapping rate, all expected sample wells recover with realistic cell counts, unused wells at noise floor.
• Byte-identical results vs the CLI-override path in PR feat(multiplex-quant): expose --sample-bc-ori as a CLI override #199 (rules out preset-vs-CLI divergence).
• simpleaf inspect lists the new preset alongside the existing ones (chemistries.json registry load path works).

Closes #198.