feat(inference): ADR-064 Phase-0 decode slope measurement harness

[ohdearquant]

Summary

• Add scripts/bench_decode_slope.sh — runs decode at multiple context lengths (64-1024), fits linear model per_tok_ms = slope*ctx + intercept, outputs JSON with slope_ms, intercept_ms, r_squared, tok_per_sec_64
• Add make bench-decode Makefile target for convenient invocation
• Add debug-only runtime assertion in MetalKvCache::new verifying buffer sizes match expected f32 layout (gates feat(inference): wire f16 KV cache into default Metal decode path #154 f16 migration)

Implements #168 (decode slope harness) and #170 (KV layout assertion).

Bench output (validated on M2 Max)

{"slope_ms":0.003771,"intercept_ms":6.1545,"r_squared":0.918216,"tok_per_sec_64":156.4,"n_points":3,"contexts":[64,128,256],"per_tok_ms":[6.4791,6.5125,7.1616]}

Test plan

• RUNS=3 CONTEXTS="64 256" make bench-decode produces valid JSON on stdout
• cargo build --release --features "metal-gpu,f16" -p lattice-inference succeeds
• Debug assertion passes on current f32 KV layout
• Verify JSON is parseable by downstream tooling (ADR-064 CI integration)

🤖 Generated with Claude Code

[ohdearquant]

Codex Review — Round 1

Verdict: REQUEST CHANGES → Fixed (all findings addressed in commit 5085d21)

Findings Addressed

1. [Major] Input validation — CONTEXTS="64 64" and CONTEXTS="8 64" now properly fail with JSON error and exit 1. Added: set -e, context validation (>N1, no dupes), bc output guard, awk denom==0 guard.

2. [Medium] KV assertion safety — Now skips when num_full_layers == 0 and verifies ALL K/V buffer pairs (not just k_bufs[0]).

Verification

$ CONTEXTS="64 64" ./scripts/bench_decode_slope.sh
WARNING: skipping duplicate context 64
{"error":"need at least 2 valid contexts (integers > 8)"}
# exit=1 ✓

$ CONTEXTS="8 64" ./scripts/bench_decode_slope.sh  
WARNING: skipping invalid context 8 (must be integer > 8)
{"error":"need at least 2 valid contexts (integers > 8)"}
# exit=1 ✓

$ RUNS=2 CONTEXTS="64 256" ./scripts/bench_decode_slope.sh
{"slope_ms":0.003179,"intercept_ms":6.0260,...,"tok_per_sec_64":160.5}
# exit=0 ✓

Ready for re-review.