COBOL export: emit calibrated PD and reason codes

[orgoca]

Problem

The COBOL export currently emits score and band only. The SQL export carries the full governed output; COBOL does not.

That is a material limitation of the headline claim. The mainframe path is the differentiating deployment story, but it cannot currently produce a calibrated probability or an adverse-action reason — which means the regulatory artifact that motivates the whole reason-code machinery cannot be generated where the decision actually runs.

Proposed approach

PD is straightforward: the calibration table is already integer piecewise-linear. Emit it as a paragraph performing the same div_rha interpolation the runtime and SQL export use. The spec (§6) already defines the algorithm.

Reason codes are harder and depend on #15 — with the current O(p²) formulation, emitting attribution in COBOL is impractical. Once per-tree decomposition lands, each feature's contribution becomes a bounded computation that COBOL can express.

Suggest splitting: PD now, reasons after leaf-time attribution.

Acceptance criteria

• export_cobol emits PD-PPM matching decide() exactly, verified by the existing GnuCOBOL run-parity harness
• Driver-harness mode prints PD alongside latent and band so CI diffs it
• Docs updated — docs/howto/deploy.md currently states the limitation and should state the new scope

Files

src/compileml/export/cobol.py, tests/test_export.py, docs/howto/deploy.md

Scope

PD alone: medium and self-contained. The CI harness already compiles and runs the generated program, so the parity test is nearly free.