Harden evidence risk and generation reads

[Pigbibi]

Summary

• add strict current_generation.v1 immutable feature-snapshot pointer loading with canonical JSON, path, digest, generation, profile, and metadata bindings
• add the research-only qsl.tqqq-evidence-risk-mandate.v1 gate with candidate, capital, portfolio, risk-state, freshness, asset-role, leverage/cap, drawdown, and loss-budget enforcement
• keep every assessment non-executable (execution_authorized=false) and fail closed on undefined budget units or unknown mandate schemas
• cover pointer bypass/fallback boundaries and risk-contract substitution/bypass cases, including zero-exposure unauthorized targets

Layer boundary

This QPK slice remains pure/no-I/O. Canonical QRS authority-receipt content/revision verification and durable single-consumption are intentionally enforced by the subsequent UESP loader/ledger slice; this PR does not claim those controls complete.

Validation

• PYTHONPATH="$PWD/src:$PWD" uv run --no-sync pytest -q --cov --cov-report=term --cov-report=xml tests — 1363 passed, 1 skipped, 22 warnings, 448 subtests passed; 73% coverage
• PYTHONPATH="$PWD/src:$PWD" uv run --no-sync ruff check . — passed
• uv pip check — passed
• python3 -m compileall -q src scripts — passed
• git diff --check — passed
• two independent final read-only reviews — no remaining P1/P2

Safety boundary

No provider access, replay, deployment, workflow dispatch, broker/order action, or paper/live activation. No runtime configuration is enabled by this PR.