make-mips rerun — the owed cache measurement did NOT happen (guard never engaged); expected merit-fail, this time at the timeout wire

What was run

make-mips-interpreter × 1 at stock settings (limit 200000), budget 8000. This was the rerun owed since …180910: the decisive frontier-hysteresis cache measurement (baseline to beat: 62.4% hit / 79 cache breaks / ~23 min re-prefill from the per-call-pruning run …152728).

What happened instead

• AgentTimeoutError at 1h00m after only 19 turns, peak context ~55k — the trajectory never got anywhere near the 200k prune threshold, so the guard never fired (0 prune lines). Last time the same task ran ~58 min, ~a hundred turns, peak 114k+ with heavy pruning; this time each turn spent minutes in the CPU-capped MIPS test runs and the clock ran out. Same merit-failure (VM too slow), different place on the clock — boundary variance, expected.
• Cache stats are trivially clean because nothing was pruned: 93.2% hit, cacheRead monotonically increasing every turn, zero breaks. That confirms the unpruned baseline behavior but says nothing about hysteresis under pruning.

Consequence

The hysteresis-under-pruning measurement is still owed. It needs a trajectory whose full-transcript estimate crosses ~200k — which is luck-of-the-draw per trial. Rather than re-rolling make-mips repeatedly, let the measurement come opportunistically from the next long run: any trial showing pr:N in make failures at stock settings is the experiment. Read its cacheRead series; expect single-digit breaks and ~90%+ hit.