Treat missingness as informative: add a `missing_as_bin` policy

[orgoca]

Problem

The artifact's missing-value policy is "baseline" (median imputation) or "reject". In credit data, missingness is signal: no trade line, no inquiry history, thin file. Median-imputing it destroys that signal and silently maps a thin-file applicant onto the score of a median-file applicant.

Standard practice in scorecard development is to give missing its own bin. We currently cannot express that.

Related: compileml/compile/extract.py already detects XGBoost's missing-value branches and warns that the compiled artifact has no equivalent — then discards the information. That warning is the ticket for this issue.

Proposed approach

Add a third policy, missing_as_bin:

• Spec (docs/ARTIFACT_SPEC.md §8): define how a node routes an explicitly-missing input. Simplest expressible form is a per-node missing_left: bool, preserving the current two-way structure and integer determinism.
• Compile: capture the teacher's missing direction when available (XGBoost gives it directly); for sklearn, require the caller to declare it.
• Runtime: route NaN/None per the recorded direction instead of imputing.
• Exports: SQL needs explicit IS NULL handling; COBOL needs a validity flag convention.

Acceptance criteria

• Spec section written before implementation — this changes the artifact contract and needs a schema version bump
• Runtime routes missing deterministically; a test asserts a missing value and a median value produce different scores
• SQL export parity test includes NULL rows
• Existing "baseline" and "reject" policies unchanged

Files

docs/ARTIFACT_SPEC.md, src/compileml/runtime/decide.py, src/compileml/runtime/score.py, src/compileml/compile/extract.py, src/compileml/export/*

Scope

Substantial, and it is a spec change — the schema version goes to 3. Worth designing in an issue thread before opening a PR.