fix(backtest): preserve negative Calmar in Crypto adapters

[Pigbibi]

Scope

• F2 / RA-06: preserve the sign of CAGR when calculating Calmar for ordinary live-pool and combo backtest results.
• One production formula line plus regression coverage in an existing test file. No promotion authority, provider, replay, dependency, lockfile, or old artifact changes.

Validation

• Fresh base: 8778bbb.
• Existing targeted baseline: 11 passed, 4 subtests. New negative-CAGR regression: 2 failing subtests on old source, then GREEN.
• Targeted runner/accounting tests: 24 passed, 12 subtests. Full offline suite: 179 passed, 95 subtests.
• Full Ruff, pip check, scope, compile, added-line credential-pattern and diff checks passed.
• Regression checks both profiles with negative/positive/zero CAGR and undefined zero-drawdown Calmar, retaining ordinary validation_identity=None.

Existing environment limitation

uv sync --locked --offline --extra research fails on the unchanged base: pyproject requests numpy>=2.4.6 while uv.lock records numpy2.3.5. Per bounded authorization, validation used a fresh isolated environment with a full offline pyproject dependency resolution, not --no-deps or --frozen: QPK c812ed70f83d61bdf1816fa5ca112b0f6976c6b6, numpy2.5.2, pandas3.0.5, pytest9.1.1; all 11 installed packages compatible. The repository lockfile is unchanged and is not claimed valid. CI likewise installs from package declarations; any unrelated lock repair remains outside this PR.