feat(soxl): add bounded three-asset learning entry

[Pigbibi]

The frozen SOXL V3 runner only accepts its original configuration, so the AI research path cannot use it to compare parameters. Add a separate SOXL/SOXX/BOXX learning CLI that validates the existing P1 input and calls the exact UES strategy version in its locked environment.

The new entry shares the corrected next-session ledger with the frozen runner. It restricts each request to development sessions through 2025-07-31 and one explicit parameter: blend_gate_mid_soxl_weight. Every batch includes the 0.65 baseline, at most two additional values no greater than the baseline, and fixed 5/10/15bps cost cases. Metrics use actual return intervals, positive drawdown magnitude and sample standard deviation; returned cost/parameter bindings must match the request.

Results contain learning summaries only, with no raw daily replay and no promotion authority. Original candidate validation, configuration, evidence and live/shadow permissions remain intact. This PR adds no AI call, workflow, data acquisition, dependency or experiment store.

Validation:

• 61 relevant tests passed, including original replay/materializer, V6/V7 and observation consumers.
• Ruff 0.15.9 and git diff --check passed.
• A synthetic full CLI run exercised the original P1 materializer and clean UES 7756fe32585e85cf1d09a163203a02e3eee39fe1 locked environment, producing three cost summaries without raw replay output. A separate synthetic mid-tier comparison confirmed the parameter changes actual numerical results.
• No real-data replay, model/provider call, historical recomputation or production/shadow operation was performed.