2026-08-10 — Third in-situ smoke test: the manual restructured, and __q retired

A fresh tenant on the same live host system ran the full surface — manual onboarding, vars scoring, a 12 s eval outliving its await with live progress, failure rendering, teardown, cancel-by-id, release! — all passing, and its impressions confirmed the bet's shape: settlement footers, trimmed exceptions and cancel-by-id are what saved effort, and the manual's irreplaceable job is the silent-failure modes (orphans, infinite returns, retention) that no response teaches by example. Asked what was tangential, it named the __q-vs-q rationale paragraph as design document rather than manual, and — pressed with Henrik's alternative — made the case that ended the split: aliasing Quiescent as q in scratch only makes eval idioms verbatim-portable against the skill and committed source, and the foreign-namespace failure mode flips from prevented-by-guarantee to loud-and-self-diagnosing (No such namespace: q, the model's own omission, one roundtrip). The __q injection was a permanent per-eval tax spent preventing exactly that error.

So built the same day: ensure-aliases! asserts __r alone; scratch requires Quiescent :as q (bare refers unchanged); the manual, eval description, api docstrings, pattern namespaces and tests all speak q. The manual keeps one line of residue as advice, not machinery: alias Quiescent __q in a namespace you don't own, the collision-proof spelling of rewiring someone else's alias map. New invariants pinned: scratch resolves q/task, and a visited namespace carries neither q nor __q unless the model put it there. Doctrine revised in §4's tooling-rail record. The manual also went principles-first this week (About → toolkit → async → handles → navigation → sharing), per the same tenant's read that the frame of mind, not the tool list, is what a fresh session needs first.

The tenant's remaining observations, unactioned and noted: the manual payload is the biggest single token cost in the flow (full docstrings for all publics; a first-sentence-only tier for the rarely-touched ones would cut a third); mid-run out peeking rarely beats turn latency for sub-10 s work, suggesting the description nudge toward sizing await to the expected duration instead; (__r/evals) listing the eval executing it as :running costs a double-take.