Move firm_generation off populace.calibrate's private optimizer internals

Where this sits

#612 increment 2 (PR #634) moved the UK local/rowwise solve onto the public calibrate() front door, retiring the three private populace.calibrate.solve imports from uk_runtime/local_solver.py. That leaves exactly one consumer of the underscore names in the repo: uk_runtime/firm_generation.py:33-38 imports _optimize and _torch_constraint_matrix and calls them at firm_generation.py:1073-1088.

Why it did not ride along

The firm problem does not fit the front door as it stands: it solves over a synthetic firm population with no Frame (firms are not a populace entity), an externally built torch constraint matrix, and per-target loss weights (_target_loss_weights(layout, config)) — the one shape the UK local doctrine explicitly forbids and calibrate()'s target-loss vocabulary only partially expresses at the problem level. Bolting a firm-shaped seam onto the UK PR would have smuggled a public-API design decision into a migration commit.

The work

Design the public seam that covers the firm problem — either a problem-level entry point in populace-calibrate (matrix + targets + per-target loss weights + initial weights, no Frame, documented as not carrying the kernel weight-kind transition), or a firm-entity Frame expression if firms should become a first-class weighted entity — then migrate firm_generation.py onto it and seal the underscore names (zero importers outside populace-calibrate).

A refactor of _optimize / _torch_constraint_matrix currently breaks the firm generator silently; there is no contract on a leading underscore.

Refs #612, #493.

🤖 Generated with Claude Code

Start with uk_runtime/firm_generation.py:33-38 and 1073-1088, then inspect the public calibrate() entry point and the referenced changes in #612 and #634. Define the public seam for the no-Frame firm problem, migrate firm_generation.py to it, and verify that no underscore imports remain outside populace-calibrate.