Give RT/iter the empty input list

[yogthos]

Closes #4.

Not the type-hint codegen bridges. latest/run and sample/run both open with

(let [it (.iterator (seq fs))] ...)

and in both, fs can be empty — a zero-input (m/latest f) has no flows at all, and sample splits its varargs so (m/sample f sampler) leaves () as the continuous inputs. An empty seq is nil, and a method call on nil is a NullPointerException.

It only ever ran because jolt answered for the nil. Through 0.8.1 (.iterator nil) handed back an empty iterator and (.toString nil) the empty string; 0.8.2 refuses both, as the JVM always has. So this is ebb's bug, not jolt's — jolt got stricter and stopped covering it. jolt-lang/jolt#867 gates the corrected behavior so it cannot drift back.

clojure.lang.RT/iter is what clojure.core's own walkers use for exactly this, and it takes nil on both hosts (verified against Clojure 1.12.0 on a JVM and against jolt 0.8.2).

The sample half was unreported

Nothing covered a sampler-only sample, which is the single most ordinary shape of the operator:

(m/reduce conj [] (m/sample vector (m/seed [1 2 3])))
;; before: NullPointerException: Cannot invoke "iterator" because the target is null
;; after:  [[1] [2] [3]]

It has a row now, and it fails without the fix.

Verified

• bin/test under the released 0.8.2 binary: 283 tests, 1199 assertions, 0 failures, 0 errors, three runs.
• Before the fix, same binary: 2 errors (latest-test/zero-input, missionary-issues-test/latest-supports-zero-inputs-126) — the two the issue reports — plus the new sample row.
• JOLT_VERSION pin moved 0.8.1 -> 0.8.2 with the fix, per the cutover convention in ci.yml.

The intermittent one-in-ten hang ci.yml documents is untouched and unrelated.

https://claude.ai/code/session_019tGgRzPx3vtr5siPkKjjLE