Tech debt: Python routing facade is designed (lp-qp-routing.md) but not built

[jkitchin]

Area: Python routing facade (designed, not built)

State. dev-notes/lp-qp-routing.md specifies a ProblemClass-driven router,
and crates/pounce-cli/src/dispatch.rs already classifies/routes on the CLI
(solver_selection=auto). But the Python surface exposes parallel, hand-picked
entry points -- minimize, solve_qp, solve_socp, sos_minimize,
minimize_global -- with no unifying dispatch.

Debt.

• Python callers must know solver theory to choose; the CLI can auto-route from a
  parsed .nl, Python cannot.
• Two divergent dispatch stories (CLI classifier vs. Python explicit) will drift.
• minimize deliberately cannot route (opaque callable), so a router can't just
  hide behind it -- it needs structured input. That design question is unresolved.

Recommendation. Decide explicitly: (a) build a Python router taking structured
problems and dispatching by ProblemClass (mirroring dispatch.rs), or (b)
commit to explicit entry points and make "Choosing a Solver" the documented front
door. Track this as the designed-but-unbuilt piece it is.

See dev-notes/multi-solver-tech-debt.md (area 2).

[jkitchin]

Resolved — already built, this issue was stale.

The recommendation's option (a) (a Python router dispatching by problem class,
mirroring dispatch.rs) shipped in the very PR this issue references (#70), in
commit caf6bcb ("feat(python): auto-route minimize() to the convex LP/QP
solver"). The issue was filed from the area-2 tech-debt note before that note was
reconciled against the code that landed in the same branch.

What exists now:

• python/pounce/_route.py — classify_and_extract() resolves the "minimize
  only sees an opaque callable" design question by probing: it evaluates the
  callables (fun/jac/hess + constraints) at an anchor plus held-out points,
  fits a linear / convex-quadratic model, and validates it before trusting it.
• python/pounce/_minimize.py — wires it behind a solver_selection option
  (auto default, plus nlp / lp-ipm / qp-ipm) that mirrors the CLI.
• Conservative fallback matches dispatch.rs: nonlinear, indefinite Hessian,
  nonlinear constraint, or any probe that raises → falls back to the NLP solver.
  A genuinely nonlinear problem is never silently sent to the QP solver.
• Documented in docs/src/python.md (auto-routing + route_tol).
• python/tests/test_minimize_autoroute.py — 8/8 passing: convex-QP routing
  with objective-constant fold-back, LP routing, auto==forced-NLP transparency,
  nonlinear/nonconvex staying on NLP, forced-selection raising, and
  finite-difference routing without user-supplied derivatives.

The two divergent-dispatch-stories concern is addressed: CLI and Python share the
same classify-then-conservatively-fall-back contract.

dev-notes/multi-solver-tech-debt.md (area 2) has been updated with a
"Status (issue #106)" paragraph recording this. Closing as completed.