RecursivePi — full smoke set

Second RecursivePi run (read-only-scout pi-recursive/rlm_query harness), the full smoke set on qwen3.6-35b-a3b, cap 1500s/trial. See harnesses/recursive_pi.py + AGENT_TODOS.md.

Result: 4/4 PASS (mean 1.00), 0 errors, 10m14s total.

task	result	agent time
fix-git	PASS	24s
nginx-request-logging	PASS	27s
openssl-selfsigned-cert	PASS	18s
regex-log	PASS	6m16s

Key behavioral finding: ZERO rlm_query calls across all 4 tasks (no rlm-trace.txt produced on any trial → definitively no child ever spawned). The tool was registered and the read-only-scout delegation prompt was active (confirmed in the transcripts), yet the model solved every task directly.

Interpretation: consistent with the delegation-decay pattern from the 18 SubagentsPi trials — these small models don't spontaneously delegate even when given the tool and told to. The smoke tasks are also all quick, self-contained wins (18s–6m agent time) that don't obviously benefit from decomposition, so there was little to delegate.

Bottom line for the spike: the mechanism is proven (host: rlm_query→child →llama-local→file read) and adds NO regression (4/4 pass), but recursive delegation does not emerge on its own here. To actually exercise it, the next run should (a) target a task that clearly benefits (large-file read / wide repo search / multi-part analysis) and likely (b) force a first-action rlm_query scout call — the same fix that took SubagentsPi delegation from 0 to working.