Intermittent-power benchmark results with realistic power failures

[byeongjee]

Summary

The first full ckpt intermittent matrix (13 benchmarks x milp/rockclimb/schematic/schematicO3 x 10 traces) showed almost no power-failure activity: MILP finished with 0 recoveries nearly everywhere while RockClimb saw many, and 25 runs never completed. Root causes: the replayed Mementos traces are far slower than the benchmarks (an outage only every 5-50 s), and the 3.3 V wait threshold was reachable only at trace peaks (the series schottky diode drops the supply below it except when the trace exceeds ~3.5 V, 2-15% of the time).

Changes:

• Wait threshold 3.3 V -> 3.0 V (vcc_wait.c), with config_board.json capacity recomputed for the lower full-charge level.
• Replay traces time-compressed 10x (preprocess_traces.py --speedup, regenerated benchmarks/traces/*.csv): power failures now occur 0.25-1.9/s, in the range RockClimb (RTAS'22) evaluates against.
• cnt_wait counter: boundaries/boots that find VCC below the threshold and sleep are now counted (this is the event RockClimb's cost model calls a wait; real brownouts remain cnt_recovery) and reported as runtime_wait_count in the intermittent CSV.
• Docs: actual 560 Ohm series resistor (was documented as 5.6 kOhm).
• Results in results/intermittent/ (per-benchmark CSVs, summary.csv, report.md, resumable driver + summarizer).

Results (520 runs)

algo	complete	correct	mean waits	mean recoveries	mean t (s)
milp	130/130	130/130	9.0	0.1	0.87
rockclimb	129/130	129/129	231.7	9.3	32.23
schematic	130/130	130/130	38.6	1.2	4.95
schematicO3	130/130	130/130	11.3	0.1	1.14

Every completed run returns the uninstrumented baseline's result. The single incomplete is stringsearch_fixed/rockclimb on trace 4 (finishes in 49-187 s on the other nine traces).

Test plan

• uv run pytest tests/ unit suites pass (incl. test_preprocess_traces.py).
• On-device: aes/crc verified across all traces and algorithms before the full matrix.