Overall, this PR looks solid! You've added robust platform detection for ANE compatibility, which should make the code more portable across Apple Silicon generations. The structure is clean, and the estimates for TFLOPS seem reasonable based on known hardware specs.

A few suggestions:

• Consider adding unit tests for the detection logic, especially edge cases like unknown chips or older macOS versions.
• The fallback values are sensible, but maybe log a warning when falling back to ensure users are aware.

Great work on keeping it backward compatible while supporting newer MIL versions! If these tweaks make sense, we can iterate.

Overall, this PR looks solid. It adds necessary platform detection for ANE compatibility across different Apple Silicon chips and macOS versions. The code is well-structured and integrates cleanly with existing files. Suggestions: Consider adding more robust error handling if platform detection fails, and document the TFLOPS estimates sources.

Full Review — All 18 Changed Files Inspected

🔴 Bug #1 (Compile-Breaking): ane_print_platform() signature mismatch

training/ane_compat.h:190 declares:

static void ane_print_platform(const ANEPlatform *p) {

But all 14 call sites across every .m file call it with zero arguments:

ane_print_platform();  // no pointer passed

Affected files: inmem_peak.m, test_ane_advanced.m, test_ane_causal_attn.m, test_ane_sdpa5.m, test_conv_attn3.m, test_full_fused.m, test_fused_bwd.m, test_fused_qkv.m, test_perf_stats.m, test_qos_sweep.m, test_weight_reload.m, tiny_train.m, tiny_train_old.m, train_large.m.

Recommended fix: Change the function to use the global directly (avoids touching 14 files):

static void ane_print_platform(void) {
    ANEPlatform *p = &g_ane_platform;
    // ... rest unchanged
}

🟡 Bug #2 (Logic): Hardcoded peak TFLOPS in inmem_peak.m:110

if(ms>0)printf("...", l, w, gf, ms, tf, tf/0.019*100);
//                                       ^^^^^ hardcoded 19 TFLOPS

This defeats the purpose of the dynamic detection. Should be:

tf / (ane_peak_tflops() / 1000.0) * 100
// or simply:
tf / ane_peak_tflops() * 100

(depending on units — ane_peak_tflops() returns TFLOPS and tf appears to also be in TFLOPS, so tf / ane_peak_tflops() * 100 should work)

🟢 What's Done Well

• Zero residual hardcoded strings — Verified via grep: no remaining program(1.3), <ios18>, "3510.2.1", "3505.4.1", or coremltools-version.*9.0 across any of the 18 files.
• Every .m file properly includes ane_compat.h (directly or transitively via stories_config.h → ane_compat.h) and calls ane_detect_platform() before use.
• Chip detection ordering in _ane_identify_chip() is correct — longer substrings checked first ("M5 Ultra" before "M5") to avoid false matches.
• MIL_HDR macro transition from string literal to ane_mil_header() function call is type-safe (both produce NSString*).
• Lazy-init guard pattern in ane_mil_target(), ane_mil_header(), ane_peak_tflops() is a good defensive pattern.
• CRLF→LF normalization done correctly on ane_mil_gen.h and stories_config.h (the large diff on those files is mostly line-ending cleanup, logic is preserved).
• train_large.m and tiny_train.m correctly use ane_peak_tflops() for utilization calculations.

📝 Minor Notes

• training/stories_io.h still has CRLF line endings (134 lines) — not touched by this PR, but it's in the include chain. Worth normalizing in a follow-up.
• Thread safety: The g_ane_platform_detected flag isn't atomic, but since all callers are single-threaded CLI tools calling from main(), this is fine for now. Worth noting if you ever move to concurrent initialization.
• _ane_peak_tflops() default fallback returns 15.8 (M4-class) for unknown chips — sensible choice for current-gen default.

Summary

The architecture is sound and the migration is thorough. Two bugs need fixing before this can ship: the compile-breaking ane_print_platform signature mismatch, and the hardcoded peak TFLOPS in inmem_peak.m.