Add rev11 real-input AZI_CHUNK sweep benchmark, scoped rev12, and rev11-vs-rev12 validator

[nicklasorte]

Motivation

• Benchmark real-input performance of subchunk_agg_check_maxazi_rev11 across AZI_CHUNK values to choose a safe default for further optimization.
• Implement a carefully scoped rev12 that preserves rev11's one-time RNG/pre-generated random matrices and chunking while addressing the next likely bottlenecks.
• Provide an automated statistical validator to ensure rev12 remains statistically equivalent to rev11 and to measure runtime impact.

Description

• Add benchmark_subchunk_agg_check_maxazi_rev11_chunk_sweep_real.m which times subchunk_agg_check_maxazi_rev11 on real inputs over chunk sizes [32 64 128 256 512 1024], uses timeit when available, prints a clean summary table, and returns chunk_sizes, runtimes, best_chunk, best_runtime, and speedup_vs_128.
• Add subchunk_agg_check_maxazi_rev12.m which preserves the one-time RNG/pre-generated random matrices and AZI chunking (default AZI_CHUNK=128, optional override), vectorizes the off-axis gain lookup via a flattened nearestpoint_app pass, and reduces temporary allocation by accumulating in linear mW domain and converting once per MC.
• Add validate_subchunk_agg_check_maxazi_rev11_rev12_statistical.m which runs rev11 and rev12 on identical inputs, captures runtimes, compares metrics (mean, std, min, max, median, p90, p95, p99, and optional p99.9), emphasizes upper-tail checks, applies documented absolute/relative thresholds, prints a structured summary, and fails closed if thresholds are exceeded.
• Keep debug checks gated with DEBUG_CHECKS=false and avoid changing aggregation semantics or helper-function call patterns to preserve statistical behavior.

Testing

• Performed repository-level inspections and sanity checks (files reviewed with sed/nl and git status --short showed the three new files present); these checks succeeded.
• Attempted to run a CLI MATLAB substitute via octave --version which failed in this container, so the MATLAB/Octave runtime benchmarks and the statistical validation could not be executed here.
• No runtime benchmark/validation tests were executed due to missing MATLAB/Octave; the new files are committed and ready to be run in an environment with the MATLAB runtime.

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: eec404382a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve -Inf when aggregate power is zero

This conversion now clamps zero aggregate power to realmin before log10, which returns a large finite negative value (~-3076 dBm) instead of the -Inf produced by rev11 (pow2db(0)). In scenarios with no active BS contributions (or all contributions omitted as NaN), this changes the output contract from non-finite to finite and can alter downstream filtering/statistics that use isfinite to detect no-signal cases.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Treat non-finite output mismatches as validation failures

The validator drops any sample where either rev11 or rev12 is non-finite before computing drift metrics, so regressions that introduce NaN/Inf (or contract changes in non-finite handling) are silently excluded from overall_pass. That undermines the stated fail-closed behavior because the comparison can pass while problematic samples are ignored.

Useful? React with 👍 / 👎.