Export every ensures emission and otherwise: emissions, not just the first statement

[yenda]

Fixes #88.

Bug

When building the set of names a module exports (and the local emitted-trigger set), the checker registered only the first statement of each rule's ensures: block — a compatibility remnant of the old TS regex, per the comment it carried. A trigger emitted as the second or later statement, even when the first statement is itself another emission, was invisible to importers:

• every qualified reference to it drew a false allium.reference.unknownName warning (exit 1),
• its cross-module consumers drew a false info-level allium.rule.unreachableTrigger,
• local listeners of a mid-block emission drew the same false unreachable info.

Triggers emitted only through requires: ... otherwise: SomeError(...) were invisible the same way (they parse as an assignment item named otherwise, which no collector visited).

Repro (fails on main, passes on this branch — transcribed into cross_module_lifecycle.rs):

rule EmissionAfterEmission {
    when: Do2(x)

    ensures:
        FirstEmitted(value: x)
        SecondEmitted(value: x)   -- consumers of this one warned unknownName
}

Why reordering was never a workaround

A block emitting two triggers can export at most one under first-statement-only rules. And the single exportable slot is often pinned by data dependencies — an emission referencing a binding a let above it creates cannot move first.

Fix

collect_leading_ensures_call becomes collect_ensures_emission_calls: every block item whose top-level expression is a bare PascalCase call is an emission — first or not. Value-producing calls nested inside assignments, let bindings or arguments are still not collected, so the emission/black-box-call distinction is preserved. otherwise: values join the emitted set via a new arm in collect_emitted_trigger_from_item.

The parse layer needed no change: allium parse already contains every emission as a structured Call node — the loss happened only in the resolution layer.

Tests

• 3 unit tests (local reachability: second-position, post-assignment, and otherwise: emissions; plus a guard that a trigger nothing emits stays unreachable)
• 2 integration tests (cross-module: consumers of SecondEmitted / AfterAssignEmitted / otherwise-emitted triggers draw no diagnostics; a trigger the provider never mentions still warns)
• Full workspace suite green: 651 tests, 0 failures. No existing test pinned the leading-only behaviour.

Real-world impact

On a ~180-spec production corpus, the v3.5.0 → v3.5.3 upgrade surfaced 4 false unknownName warnings of exactly this shape (a login spec consuming triggers its auth-OTP module emits mid-ensures), forcing per-site suppression comments. With this fix those suppressions become unnecessary.

---

Note: this PR and the config/deferred one both append tests to cross_module_lifecycle.rs and the analysis.rs test module — whichever merges second needs a trivial union rebase; happy to rebase promptly.