chore(e2ebench): main-v2 cache baseline marker (do not merge)

[esengine]

Do not merge. Empty commit off main-v2 HEAD (no file changes) so the e2e bot builds the current main-v2 agent verbatim. This gives a cache-hit baseline to compare the cache-sensitive PRs (#2958/#3027/#3037) against while we chase the cache-hit regression reported in #3091 — in particular the per-task cache hit on the compaction task.

/e2e

[esengine]

Baseline run against main-v2 HEAD.

/e2e

[github-actions]

🤖 Reasonix e2e benchmark

Accuracy: 4/4 (100%) · Cache hit: 79% · Tokens: 160,565 (prompt 158,394 / completion 2,171) · Compactions: 3 · Cost: ¥ 0.0406

Task	Result	Steps	Prompt	Completion	Cache hit	Compact	Cost
compaction	✅ pass	8	102,643	1,061	68%	3	¥ 0.0365
fix-add-bug	✅ pass	4	24,858	420	98%	0	¥ 0.0017
fizzbuzz	✅ pass	2	12,258	244	99%	0	¥ 0.0008
palindrome	✅ pass	3	18,635	446	99%	0	¥ 0.0015

_{Real provider run. Cache-hit % is cached prompt tokens / total prompt tokens.}

agent: PR head (00da2bc) · triggered by @esengine

[esengine]

Baseline captured: main-v2 HEAD = 79% total / 68% on the compaction task / 4-4. Recorded for the #3091 cache investigation; closing the marker.