docs: retract the aksops +14% (iso-accuracy violation)

[Poytr1]

Why

I reported aksops as a +14% JIT win from a --rollouts 3 run. A --rollouts 6 re-run shows that was a small-sample artifact and it is retracted.

aksops n=3, rollouts=6:
  baseline: 100% success (6/6)   T2S ~333,746
  jit:       33% success (2/6)   T2S ~259,922  (mean over only the 2 that passed)

The "saving" was a mean over a different, smaller success set — not iso-accuracy, so not a valid comparison. The rollouts=3 run got lucky (JIT 3/3). The aksops fixture is flaky for the agent (it notices the mock stubs are fake and sometimes short-circuits).

What this changes

• No shape shows a valid JIT win: nullcheck ~0%, shellseq ~0%, migrate −25%, aksops inconclusive.
• Ops/tool-use remains the most promising hypothesis but is unconfirmed — needs a fixture where both arms succeed reliably.
• Adds the lesson: compare at iso-accuracy and check the verified sample size on both arms before believing a mean. (aj bench --compare already prints an iso-accuracy WARN; I bypassed it in a JSON post-process and misled myself.)

Docs-only; corrects a claim already on main (from #15).

🤖 Generated with Claude Code