2026-08-06 — Fourth smoke test

Fourth smoke test (2026-08-06, fresh tenant, first of the lean-instructions era): the criterion held. Asked only to smoke test from a no-prior-knowledge stance, the tenant ran (__r/manual) before its first real eval — the pull design's make-or-break behavior, n=1 — then learned the entire system through the drilldown without one filesystem read ("discoverability is genuinely good"), ~35 calls and ~6 k tokens all told, against the 2.7 k the eager design charged every conversation up front. Verdict "passes"; the full battery (navigation, ns continuity, error rendering, running/settled lifecycle, cancellation-as-value, cross-eval composition, compel, :by attribution, release!) confirmed the manual's claims empirically, including a deliberate reproduction of the dotimes orphan trap. Three findings, all reproduced live: (1) qfor silently drops extra binding entries — (qfor [i (range 5) :max-parallelism 1] …) macroexpands to the first pair with the rest discarded (verified by macroexpand-1), so a guessed modifier becomes a silent behavioral surprise; a shape assert in the macro is the fix, upstream in Quiescent — queued as a candidate alongside the skill's dotimes item. The want itself (bounding parallelism) is q/gate, which a clued-in model reaches two ways — the Quiescent skill where available, or outlining co.multiply.quiescent — so the manual now says the latter out loud: Quiescent is a namespace like any other, worth outlining before guessing at an option it might not have. (2) Concurrent printlns interleave mid-line in the tee — reproduced verbatim (ticktickticktick 1 from six parallel tasks). Settled as teaching, not machinery: the interleaving is any console's truth (println writes its pieces separately), the tee's never-withholds doctrine is what makes the progress line and since cursor trustworthy, and the tenant read the garbled output correctly — the evidence bar for read-side line assembly is a model drawing wrong conclusions from it, not cosmetics. The manual now carries the working idiom where printing is taught as the cheapest instrumentation: from concurrent work, print a whole line in one write — (print (str "tick " i "\n")), one string one write, where (println (str …)) still writes its newline separately. Both manual additions reached the served surface with no reload — the per-call slurp working as designed. (The file itself moved to api/manual.md the same day, beside the surface it documents.) (3) The footer named a namespace the session never entered — ;; id 36 · ns no.such.namespace on a failed switch, the requested-not-effective ns. Fixed the same day: an explicit ns that does not exist was never entered (resolve-ns refuses inside the task), so the footer default now trusts the request only when find-ns confirms it, falling back to the session's actual place — verified both ways (bogus ns reports scratch, a deliberate throw inside clojure.string still reports clojure.string).