fix(inference): avoid float32 overflow in StatsNormalizer

[ShubhJain09]

Summary

• _normalize in StatsNormalizer now does the math in float64 and casts the result back to the dtype it produced before
• The fix covers the mean std and min max and quantiles modes which all had the same overflow
• Added regression tests for each mode and a check that the output dtype stays the same for float16 float32 and float64 inputs

Why

• The stats normalizer fuzz job fails on main because intermediates like 2 * (x - q01) overflow float32 to inf even when the normalized value is an ordinary number
• One crash input had x equal to 0 and q01 around -2.3e38 where the correct output is 1 but the old code returned inf
• This makes the required fuzz check fail on unrelated PRs such as fix(runtime): discard stale inference after watchdog reset and warmup timeout #320

Validation

• The new overflow tests fail on main for all three modes and pass with this change
• Both crash inputs saved by CI now pass the harness and a 120 second fuzz run did over 4 million executions with no crash
• pytest tests/unit packages/*/tests passes and prek run --all-files is clean
• The float64 path adds roughly 1.4 ms on a 640x480 frame and under a microsecond for joint state

Breaking changes

• None

Related issues

• Follow up to the fuzz failure discussed in fix(runtime): discard stale inference after watchdog reset and warmup timeout #320 as suggested by @AlexanderBarabanov

[Copilot]

Copilot review overview

🟢 Approval recommended

The focused implementation resolves the overflow while preserving established dtype behavior with comprehensive regression coverage.

Review effort: Balanced
Findings: None

What changed in this PR

Prevents representable normalization results from overflowing during float32 intermediate calculations.

Changes:

• Performs normalization arithmetic in float64, then restores the expected output dtype.
• Adds overflow regressions and dtype-preservation tests for all normalization modes.

File	Description
src/​physicalai/​inference/​preprocessors/​stats_normalizer.py	Uses float64 intermediates across normalization modes.
tests/​unit/​inference/​preprocessors/​test_stats_normalizer.py	Covers overflow cases and output dtypes.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.