2026-08-07 — The output tail, and the bank goes line-structured

Built (2026-08-07) — item (6) from the in-situ wishlist, taken structurally. The question that shaped it: values truncate safely because (res id) holds the whole, but the output bank keeps only a 32 KiB tail — would a tail option lose the rest? No: out is a non-consuming read, so a window takes nothing from a later full read; the only irretrievable drop is the cap's own, which predates the feature and stays announced. Henrik's structural call: bank entries, not one string — retain or drop whole log lines, string manipulation only at the edges. So the bank became a queue of completed lines plus the pending unterminated partial: trimming pops whole lines oldest-first (the one surviving string cut is a single line outgrowing the entire cap), tail is a subvec window with what lies behind announced (…[6 more lines retained]), and deposit! shed an accident the restructure exposed — the old string-concat bank re-copied the whole retained 32 KiB on every write. The since cursor kept its character-count contract through the restructure, tests unchanged and passing. Surfaced as (__r/out id n), the pending line counting as the newest entry — it is the progress line. Verified live: an 8-line chatty eval tailed to 3 showed the last two lines plus the in-progress print with the window announced, the full read intact behind it. The reload itself demonstrated the transition hazard on cue: the eval carrying the :reload deposited into a bank minted by the old build and NPE'd — old-shape state meeting new code, machinery-subtracted frames pointing straight at it; the next eval minted fresh and ran clean. Suite green: 74 tests, 266 assertions.

Revised the same day: the windowing arity lasted hours. Henrik's ergonomics question — wouldn't out returning the vector compose better than (out id n)? — answered itself the moment it met the doctrine: results are plain data, composition is the option surface, and (out id n skip) is exactly the per-function knob §2 rejects. So out returns the vector of lines (terminators stripped, pending line last), take-last peeks, drop-last pages, filter searches — verified live, all four idioms through the connection — and the drop announcement moved to the vector's first element, positionally honest at the head end where tail-reaching composition never meets it. One trim rule tightened in passing: a completed entry outgrowing the whole cap is now cut from its own front instead of dropped whole — tail-keeping outranks entry purity when the two collide. Suite green: 75 tests, 268 assertions.