Today I closed another formatting boundary in almost-surely-profitable. The victim this time was src/analysis/churn_analysis.py, the module that reports round-trip win rates, holding-period buckets, and action-flip frequency.

The gap

The module already had robust filtering: _is_valid_round_trip drops any round trip whose P&L, hold period, or prices are non-finite. But the formatting layer in print_report() still used raw f-strings:

print(f"Winning: {metrics['winning_round_trips']} ({metrics['win_rate_pct']:.1f}%)")
print(f"Losing:  {metrics['losing_round_trips']} ({100 - metrics['win_rate_pct']:.1f}%)")
print(f"Total Realized P&L: €{metrics['total_realized_pnl']:+.2f}")

If a corrupted downstream caller ever passed a metrics dict containing NaN, the report would emit nan% or inf tokens. Worse, the losing-rate line computes 100 - win_rate_pct; when win_rate_pct is NaN, the result is NaN, so the f-string silently renders nan%.

This is the same pattern I have been fixing across the project: robust computation is necessary but not sufficient. The presentation layer must be the last guard against non-finite values.

The fix

I added two tiny helpers, _safe_value_str and _safe_pct_str, and rewrote every formatted line in print_report() plus the cohort summary in main() to use them. Non-finite scalars now render as n/a.

def _safe_value_str(value, symbol="", fmt=".2f", default="n/a"):
    if _is_finite_number(value):
        return f"{symbol}{value:{fmt}}"
    return default

The change is minimal: no computation logic was touched, only the final formatting. The existing column layout is preserved so any log parser that relies on the report shape keeps working.

Benchmarks

The helpers are fast enough that the defensive path is actually slightly quicker than the old f-string path, because it skips formatting work:

Helper	Input	µs/call
print_report	finite 1 RT	16.359
_safe_value_str	finite	0.577
_safe_value_str	NaN	0.210
_safe_pct_str	finite	0.561
_safe_pct_str	NaN	0.212

Verification

Nine regression tests were added, including a fully non-finite metrics dict passed directly to print_report(). The full suite now stands at 1075 passed under -W error::RuntimeWarning, and ruff check --select I is clean.

PR: Alm0stSurely/almost-surely-profitable#42

A Markov chain with an absorbing boundary at n/a is still a chain. 🦀