6. shape

A report, not a manipulation. It never realizes, never forces, never recurses into anything that could block, throw or have effects.

The rule is not a compute budget — this is one-off REPL work, where 5s of local compute beats a context-filling round trip every time. The rule is: touch nothing that could block, throw, or have effects; otherwise scan everything.

• Vectors and maps are never lazy → scan them completely. No sampling, no :full flag, and therefore no confidence hedging: keys {:sid :uri :mime} is a fact.
• Lazy seqs, eduction, bare Iterable, channels, Quiescent Tasks: name them, don't touch them. LazySeq can report realized? on its head for free. To look inside, the caller composes: (shape (vec (take 3 …))) — the decision to realize is explicit and belongs to the caller.
• Opaque objects — nippy byte arrays, RocksDBWrapper, MachineLatch — get their class and size named, never their contents walked. (Rama subindexed reads hand back unserializable wrappers; walking one can block or throw.)

Report in the vocabulary of the predicates used to narrow. The test for every candidate field is: does knowing this change what I type next? Leaf class names (java.lang.Long) fail it. Nil presence passes it, prominently.

(shape (res 12))
;; vector 4  ⚠ mixed types — not sortable
;;   number?   2
;;   string?   1
;;   keyword?  1

(shape (res 7))
;; vector 17323
;;   map? — keys {:sid :uri :mime :parsed :tokens}
;;     :mime    string?  #{"application/pdf" "text/html" "text/markdown"}
;;     :tokens  number?  nil in 412
;;     :parsed  byte[]   ~180 KB each

• Predicate names verbatim, so the output is the code. Always the most specific predicate covering the whole collection — number? not int? if there are doubles, or the caller writes a filter that silently drops half.
• Counts alongside, so the caller knows whether a filter returns 2 things or 20,000 — i.e. whether to print it or shape it again.
• For maps, keysets replace predicates, and for low-cardinality values, the value set itself — that is what lets a filter be written without a second round trip. Stop enumerating distinct values at ~12.
• Nil presence called out separately, never folded into the type list. It is what breaks the follow-up expression.
• Mixed-type collections flag "not sortable" — sort on [1 2 "hello" :world] throws, and that is exactly a follow-up written blind.
• Homogeneous collections collapse to one line: vector 100000 — number?. If generality taxes the common case it loses on net.
• Memoize the digest on the registry entry. The value is pinned and immutable, so shape → schema → a filter costs one scan.

Do not emit suggested expressions. "Try (filter number? (res 12))" guesses intent, bloats output, and is usually wrong.

Do not aim for a schema by default. A schema must account for every element, so its complexity tracks the data's irregularity. What is wanted is a navigational description. Where an actual schema is the goal (writing a spec), malli.provider already does it and already samples — verified present in malli 0.20.1, already on fluent-platform's classpath. Resolve it dynamically; don't take the dependency.

		use
(shape x)	full scan of realized structure	default rendering for large values
(schema x)	malli.provider	opt-in, when writing a spec

When a value exceeds the print budget, emit its shape plus the id — never …-truncated data. Truncated data shows neither structure nor content; shape plus a handle shows the structure and tells the caller what to filter on.

Interim, built 2026-08-07 (wishlist item 3 from the first in-situ smoke test): until shape replaces the dumb cap, the cap announces itself — the response footer counts what the caps hid (;; truncated — 9872 elided · (res 8) is the whole value) via a mirror of the printer's own walk: counted collections answer exactly and for free, an uncounted seq is probed one already-realized cell past the cap, a subtree below *print-level* is reported hidden rather than sized — so the mirror can neither block nor force anything the printer didn't. Deliberately a footer annotation, not an inline ... n more rendering: inline requires a custom printer, which is exactly the machinery shape retires this cap with.