docs(benchmark): final 89-task figures for the merged-system arm

[OsherElhadad]

The Terminal-Bench run completed after #59 published its interim section. The complete result is materially stronger, and one framing claim needed correcting downward.

The headline improved

cgfinal is the first arm to beat baseline on the raw 89-task total — $94.95 against $100.81 — and it solves the most, 65 of 89 (73.0%). No arm in the original four-way study managed either.

	baseline	cg (old)	headroom*	rtk*	cg (merged)
solved / 89	56	58	64	55	65
total billed	$100.81	$102.55	$114.75	$118.83	$94.95 (−5.8%)
cache-write	4.01M	6.53M	12.37M	7.05M	3.90M
cache-hit	98.15%	96.86%	94.1%	97.3%	98.16%
own LLM cost	$0	$3.26	$0	$0	$0

* cited from the original study, not re-derived — their trial artifacts are pruned from disk.

Beating the raw total is the harder bar, which is worth saying explicitly: six degenerate baseline trials flatter the baseline column, and that is exactly why the correction box exists further down the page.

Both reference arms reproduce their published totals exactly (baseline $100.81, cg-old $102.55). That is the check that the harness and cost model are unchanged across the two studies — without it, a 5.8% improvement could just be drift.

One claim corrected downward

The clean-set aggregate is −15.4%, not the interim −16.4%, on the final 82-task set (62 solved vs 54; steps 2,865 vs 3,117; $83.52 vs $98.70).

And path-tracing alone still carries about half of it — −8.3% without it — so the median per-task figure of −7.7% remains the number to quote for a normal task. That framing is unchanged from #59 and is the point of publishing both.

Two limitations only the completed run revealed

• caffe-cifar-10 is unscored (reward=None after 2,361 s of its 4,800 s budget). It is in the clean set for cost and out of it for the solve count. Baseline also failed it, so the comparison is not skewed by the exclusion — but the asymmetry needs stating rather than quietly resolving one way.
• Five arm exceptions are genuine 4× wall-clock timeouts, counted as reward-0 exactly as in the original study.

Unchanged

Everything else in the section stands: cache-write/cache-read back to 1.86%, identical to baseline (the previous arm ran 2.86%); #40 explicitly unverified by this arm because its only callers are the three excluded components; cachesplit structurally inert on TB's Agent-SDK traffic; and the four published regressions, including system-administration at +17.2% cost and −2 solved.

mkdocs build --strict passes.