docs: tutorial notebook `03_score_vcf.ipynb`

[AbdelStark]

Context

Batched throughput demo.

Scope

• Score a toy VCF.
• Demonstrate batched throughput.
• Aggregate receipts.

Out of Scope

Design Reference

• examples/README.md

Acceptance Criteria

• Notebook runs end-to-end on a fixture VCF

Parent tracking issue: #20

[AbdelStark]

Blocked (session-scoped): tutorial depends on APIs that have not yet shipped.

Tutorial notebooks demonstrate the end-to-end user journey through geno_lewm.{encoder,score_variant,plan,read_receipt,...}. Those APIs themselves depend on the ML stack that this session could not bring up (see comments on the encoder / predictor / training / surprise / planning issues). Writing a tutorial against missing APIs would either:

• Hard-code expected outputs against code that does not exist (false demo), or
• Ship a notebook that does not execute (broken docs).

Notebook(s) that ARE buildable today (might be picked up in a follow-up):

• 07_verify_receipt.ipynb (docs: tutorial notebook 07_verify_receipt.ipynb #99) is the closest to feasibility because geno-lewm-verify and geno_lewm.provenance already ship (provenance: implement content-addressed model IDs + manifest schema #74, provenance: implement input commitment #75, provenance: implement checksum receipt writer/reader #76, provenance: implement geno-lewm-verify CLI (checksum mode) #77 merged this session). A tutorial showing manifest/receipt round-trip is constructible against the current package; what's missing is a real (non-mock) receipt produced by an actual scoring call, which still depends on the ML stack.

To pick up: wait for the relevant upstream issue (#62 / #65 / #59 / #66 / etc.) to land, then implement.

[AbdelStark]

Blocked on batched VCF scoring.

Evidence:

• This notebook requires scoring a toy VCF end-to-end and demonstrating batched throughput.
• geno_lewm/cli/score.py is still a Phase 1 stub that points to surprise: publish clean-machine geno-lewm-score transcript #65.
• GenoLeWMRuntime.score_vcf() in geno_lewm/deploy/runtime.py currently validates paths/batch size, enters the fail-closed network guard, then raises backend-not-ready because no scorer backend/checkpoint is installed.
• Upstream scorer issues surprise: validate scoring APIs against released artifacts #62/surprise: publish clean-machine geno-lewm-score transcript #65 are already marked status:blocked.

Why not write the notebook now:

• A notebook with a fake result table would not exercise the VCF scoring path.
• A notebook that calls the current runtime would fail at the backend-not-ready guard.

To unblock:

• Land score_vcf, geno-lewm-score, and a tiny VCF/checkpoint fixture that can produce real local receipts.