fix: multi-line fn body with record-constructor tail no longer trips ILO-P020

[danieljohnmorris]

Summary

A bare record constructor (cr field:val field:val) at the start of a continuation line in a multi-line function body was getting mis-classified as a new top-level fn declaration. The body terminated, the parser tried to read cr country:name revenue:rv as a fn header, and ILO-P020 fired because that line has no > at all. The persona-side workaround was paren-wrapping the constructor (cr country:name revenue:rv) — a four-character cost on every multi-line function returning a record, which is a very common JSON / API shape.

Reported by db-analyst rerun8 friction #3 (severity red). Manifesto framing: record-as-tail multiplied across every agent reaching for a record-shaped return is a measurable token-cost paper-cut that this fix removes for the whole population.

Repro

type cr{country:t;revenue:n}
build-country-rec p:L _>cr
  name=p.0
  rv=p.1
  cr country:name revenue:rv

main>n
  p=[ "uk", 100 ]
  rec=build-country-rec p
  rec.revenue

Before:

ILO-P020 incomplete function header for `cr`: header runs off the end of the line
ILO-T008 return type mismatch: expected cr, got _

After: 100 (clean run on tree, VM, Cranelift).

Root cause

Parser::is_fn_decl_start_strict in src/parser/mod.rs is the guard that prevents parse_body_with from terminating a body when the next ;-separated statement looks like Ident param:type ... but is actually a record-constructor expression. It scans forward looking for > (the return-type separator that confirms a real header) and stops only at ;, }, or EOF. Newlines have already been filtered out of the token stream (their positions are recorded in decl_boundary for ILO-P020 anchoring), so the scan would walk straight across a top-level declaration boundary, find the > of the NEXT function's header, and incorrectly return true.

What's in the diff

• src/parser/mod.rs: per-token decl_boundary check inside the scan loop. A real fn header always has its > on the same logical line as the name, so finding a boundary before > means this is not a header.
• tests/regression_p020_record_tail.rs: cross-engine coverage (tree + VM + Cranelift) for (a) 1/2/3-field record tails followed by another fn — the original repro; (b) record tail with no trailing fn (EOF branch); (c) record tail on single-line body after explicit ;; (d) regression guard for real fn decls still being recognised correctly (the case the strict check was originally added for).
• examples/record-tail.ilo: agent-facing example, picked up by the examples_engines.rs harness so it runs on every engine on every commit.
• SPEC.md: record-constructor row added to the "safe non-last function body endings" table. ai.txt and skills/ilo/SKILL.md regenerate from SPEC.md via build.rs and pick up the same row.

Test plan

• minimal repro now runs clean on tree, VM, and Cranelift
• new cross-engine regression test covers the four shapes
• examples/record-tail.ilo runs on every engine via the harness
• doc regeneration (ai.txt, SKILL.md) clean after cargo build
• full cargo test --release --features cranelift green
• CI all checks green

Follow-ups

None. The fix is narrow to one helper with one new early-return; no adjacent bugs surfaced.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!