tests: efficiency suite — tiny-model regression gates + full-model optimization dossier (#359)

[woolcoxm]

Closes #359.

What this adds

Two layers of efficiency coverage, both parsing telemetry glm.c already emits but nothing previously asserted on or surfaced:

tests/test_inefficiency.py — 8 tiny-model asserted tests (run via make test / test-python discovery, ~0.15s/run on glm_tiny):

• telemetry parses (format-drift guard)
• tiny tok/s ≥ floor (throughput regression)
• PROFILE phases present and non-negative (accounting sanity)
• disk-wait not dominant on a resident model (I/O-path regression)
• CPU greedy determinism (two runs agree)
• CUDA (auto-skip on CPU-only builds): init path (no silent exit 2), dense actually uploads VRAM, CPU-vs-CUDA argmax agreement ≥ 70% (kernel-correctness guard)

tests/test_efficiency_report.py — opt-in optimization dossier for a real model. Turns on every instrumentation flag (PROF, COLI_CUDA_PROFILE, CACHE_ROUTE, DISK_SPLIT, LOOKA) and prints 9 sections — provenance, throughput + tail latency, where-time-goes, attention breakdown, expert cache, disk I/O + phase split, routing quality + predictability, speculation, GPU tiers — each flagging inefficiency with the concrete knob to move tok/s. Never fails CI.

tools/efficiency.py — shared harness. parse_run() captures every telemetry signal; run_engine() wraps the subprocess. Reuses PROFILE_RE/SPEED_RE from tools/benchmark_cuda_fixture.py and extends the tok/s regex to also match the run_text (parenthesized) format the full-model PROMPT path emits.

Makefile — efficiency / efficiency-cuda / efficiency-report targets.

How to run

make efficiency                 # tiny CPU tests (part of make test)
make efficiency-cuda            # tiny CUDA tests (needs: make glm.exe CUDA_DLL=1 && make cuda-dll)
COLI_EFFICIENCY_MODEL=../glm52_i4_g64 COLI_EFFICIENCY_CUDA=1 make efficiency-report

Verification

• All 8 tiny tests pass (3 CUDA tests skip on a CPU-only build, pass on a CUDA build).
• End-to-end on the full glm52_i4_g64 (390 GB), CPU and CUDA: clean exit, no crash.
• The dossier surfaced real optimization targets on g64: 74% I/O-bound at 14.8% hit rate (cache re-fetching evicted experts — 600 loaded/token vs topk=8), and "layer input, skip attention" predicting next experts at 86% recall vs 24% for plain prefetch (signal to enable PILOT).

Notes

• The CUDA tests require a CUDA build (make glm.exe CUDA_DLL=1 + make cuda-dll). They auto-skip otherwise — make test is green on any CPU-only build.
• Tunable floors/thresholds are module constants (tools/efficiency.py, top of test_efficiency_report.py), overridable via env. Documented in tests/README_efficiency.md.
• No engine code changed; this is test/tooling only.