Add AutoPolicy 001 exact budget planner

[Labeeb2339]

What

• add deterministic exact scalar and robust-minimax mixed-precision budget solvers
• add strict, content-bound candidate/plan artifacts and the cliffquant autopolicy CLI
• add a measured 150-matrix Experiment 001 candidate bundle plus an exact 24-transformer-block allocation view
• publish a 10-point logical-byte/proxy frontier, representative 7.087 bpp plan, graph, method, result, and prior-art notes
• bump the package and installed CLI to v0.2.0

Recorded result

At a 440,721,408-byte planning budget, the exact block-constrained solution uses 440,303,616 logical bytes (7.087 bpp): 6 W4 blocks and 18 W8 blocks. Its worst-environment additive distortion proxy is 218.811 versus 1078.217 for all-W4, a 79.706% reduction in this planning proxy.

This is not a mixed-bit checkpoint export or a downstream quality, checkpoint-size, runtime-memory, latency, or novelty claim.

Important negative result

The exact robust DP is not practical on the full 150-matrix nontrivial frontier with the current implementation. The release therefore exposes the failure explicitly and uses a documented exact 24-block constraint; it does not silently approximate.

Verification

• 508 passed, 6 skipped
• 1,500 scalar and 1,000 robust randomized problems matched exhaustive oracles
• Ruff lint and format checks passed
• portable release manifest revalidated all 154 evidence files
• validator independently rebuilds all 24 blocks and re-solves all 10 frontier points
• source distribution and wheel built successfully
• isolated installed-wheel version, console entry point, solve, and verify smoke passed