2026-08-06 — In-situ smoke test

In-situ smoke test (2026-08-06, fluent-platform, the 0.1.0 jar from Clojars — first run in a real app). A clean pass across the whole surface in a 1909-namespace JVM: the running lifecycle end-to-end (live progress line at the await deadline, await blocking to :done, res/out delivering after), cancellation, failure rendering with machinery-subtracted frames surfacing a genuine app-side NPE cleanly, multi-session roster, namespace stickiness, eviction announcing itself under load, and app log lines interleaving into ;; out — it is the real JVM. The detach story composed with the harness: a 10-minute await held the call open, Claude Code backgrounded it at its own 120 s threshold and delivered the settled result as a task notification, while the registry id stayed readable from parallel evals meanwhile — two independent handles to the same work, neither stepping on the other. The vehicle was real: two Rama module updates fanned out in parallel via [(q/task (update-module! …)) (q/task (update-module! …))] — deep grounding making a returned vector of tasks the idiomatic single-eval fan-out, now a line in the manual's async section. The truncation/res split validated exactly as designed: a 10k-element vector rendered cut at ~128 elements while (count @(__r/res 26)) → 10000 — truncated rendering plus a composable handle to the untruncated value. Two doc amendments came out of it: the eval description now states last-value-only up front ("batch independent probes in a vector or labeled map" — the tenant guessed wrong first; last-value-only itself stays, because a per-form subindex would leak into the whole registry model — an id is a handle to one task, one status — to buy what a caller-labeled map already expresses better), and the description's orphan sentence was corrected to "by the time the eval settles", matching the manual. The tenant's __r wishlist, ranked by felt friction, noted for the next phase: (1) global var search — outline-shaped hits with :ns attached, docstring search with :matched evidence; the one it actually missed ("where does normalize-str live?" required guessing namespaces); (2) usages — who calls this, across loaded namespaces (already on the roadmap via clj-kondo); (3) truncation annotated with what was elided (... 9872 more]) — folds into shape; (4) classpath-aware namespace discovery ({:all true} marking loadable-but-unloaded); (5) Java class outline — clojure.reflect reshaped into sorted members-and-signatures data; (6) an out tail option for chatty long evals. Declined by the tenant itself, in agreement: value pretty-printing helpers (truncation + res covers it) and reload/system-lifecycle helpers — project conventions, not the lens's business.