RecursivePi + soft first-action nudge — full smoke set

Fourth RecursivePi run: the read-only-scout harness with a soft, non-blocking first-action nudge (a followUp reminder to use rlm_query, sent up to 3× per trial until the model delegates, never blocking a tool). Full smoke set, qwen3.6-35b-a3b, cap 1500s. This replaced the hard force block from smoke__…010908.

Result: 2/4 PASS (mean 0.50), 49m19s — WORSE than both prior configs.

task	result	total	nudges	rlm_query	children ok
fix-git	PASS	15m54s	3	1	6
nginx-request-logging	PASS	5m01s	3	1	6
openssl-selfsigned-cert	FAIL	2m37s	3	1	6
regex-log	FAIL	25m45s cut	0	0	0

The persistent nudge DID get the model to delegate on 3/4 tasks (unlike the ignored one-shot nudge) — 3 nudges each, one rlm_query, children reached llama-local. But it made things worse: - openssl PASS→FAIL (delegation + nudge churn disrupted a task with a known fragile hidden criterion). - Everything much slower (nginx 1m→5m, fix-git 15m54s). - regex-log is NOISE: 3.4 KB transcript, one stop, no tool turn, nudge never fired — the first generation stalled (likely llama-swap slot contention after the long fix-git/nginx trials before it) and the trial hung to the 25m cap. Not a delegation effect, but it counts as a fail.

Three-config comparison (same 4 smoke tasks, qwen3.6-35b-a3b)

config	pass	wall	delegation
no push (nudge off) …004419	4/4	10m14s	0 tasks
hard force block …010908	3/4	35m41s	4 tasks
soft persistent nudge …020217	2/4	49m19s	3 tasks

Robust conclusion: pushing rlm_query delegation on these tasks monotonically degrades both correctness and wall clock. The model's spontaneous non-delegation (no-push run, 4/4) was correct — a fresh-context read-only scout adds latency on the single shared slot and buys nothing on tasks that are quick, action-oriented, and already fit the context. Recursive delegation should only be expected to help on genuinely read-heavy tasks (large-log parse, wide codebase search), which the smoke set does not contain.