parser: anchor incomplete-fn-header errors at the offending function (ILO-P020)

[danieljohnmorris]

Summary

Fixes a token-spend trap reported by qa-tester + devops-sre rerun3: a malformed function header in a multi-function file used to report its error span on the wrong function. Personas wasted session time bisecting which function actually had the bug.

The root cause is in Parser::new: every Token::Newline is filtered out before parsing. When parse_fn_decl reads a malformed header, parse_type / parse_params happily walked across the newline into the next function and emitted the error there.

Manifesto framing: a correctly attributed span lands the persona on the right line first try instead of forcing a token-expensive bisect.

Repro (before → after)

f1 a:n>n;+a 1
f2 a:n>R       -- f2's err-type is missing
main>n;0

• Before: ILO-P003: expected Greater, got Ident("main") at line 3 col 1 (the start of main).
• After: ILO-P007: expected type, got end of line at line 2 col 8 (anchored on f2's line).

f1 a:n>n;+a 1
f2 a:n         -- f2 missing `>type;body` entirely
main>n;0

• Before: ILO-P003 at line 3 col 1.
• After: ILO-P020: incomplete function header for \f2`: header runs off the end of the lineat line 2 col 1, with a hint pointing at thename params>type;body` shape and the indent-the-continuation workaround.

EOF cases (file ends mid-header) used to fall through to Span::UNKNOWN and render as line 1 col 1. Now they anchor at the previous token (last token of the offending function's line) and surface ILO-P020 with the same friendly message.

What's in the diff (per commit)

1. parser: anchor incomplete-header errors at the offending function — Parser records decl-boundary metadata in Parser::new (parallel Vec<Option<Span>> populated before newlines are filtered out). New helper check_fn_header_boundary runs at the two header pinch points in parse_fn_decl (after params, after >). parse_params stops at a boundary so the loop can't drag the next function's name in as another param. parse_type gets a safety net for nested type slots that anchors its existing ILO-P007 at prev_span().
2. diag: document ILO-P020 incomplete function header — registry entry with the three failure shapes (missing return type, missing >, missing err-type after R) and the indent-the-continuation workaround. SKILL.md common-mistakes gets a matching entry as Fix unary negation -x parsing #12.
3. test: cross-line attribution for parse errors in multi-fn files — initial 9 regression cases.
4. parser: extend ILO-P020 to cover EOF mid-fn-header — check_fn_header_boundary treats EOF as a soft boundary alongside decl boundaries. parse_type EOF case anchors at prev_span() instead of Span::UNKNOWN. Existing eof_while_expecting_identifier parser unit test had its assertion updated to accept the strictly-better ILO-P020 wording.
5. test: cover EOF mid-fn-header attribution — 1 additional case + restoration of error_on_last_function_stays_on_last_function to assert proper attribution (not the EOF-fallback shape it was relaxed to mid-development).

ILO-P020 not P019 because P017/P018/P019 are already documented in SPEC.md/ai.txt for the import system. Pre-existing namespace mess; this PR doesn't touch it, just steps around.

Test plan

• New regression test tests/regression_multi_fn_error_span.rs — 10 cases covering missing-err-type, missing->, missing-return-type-after-arrow, middle-of-three attribution, first-fn attribution, last-fn EOF attribution, last-fn missing-> EOF attribution, zero-param-fn variant, valid multi-fn parity, valid indented continuation parity. All passing.
• Parser unit tests (cargo test --lib parser::) — 295/295 passing, including the updated eof_while_expecting_identifier.
• Existing P001-cascade regression — 5/5 passing.
• Full suite: 3064 passing locally. 7 AOT Cranelift tests fail with cannot find libilo.a — Searched: target/release/libilo.a because they hardcode target/ and ignore my CARGO_TARGET_DIR=target-isolated override. CI uses the default target dir and won't hit this — pre-existing infra constraint, not caused by this PR.
• Manual repros verified on the rebased release binary across all four shapes (mid-file/EOF × missing-arrow/missing-err-type).

Follow-ups

• EOF rendering as Span::UNKNOWN (line 1 col 1) is infra-wide — every error site uses peek_span() which returns UNKNOWN at EOF. This PR routes around it for the fn-header case specifically. A follow-up could replace peek_span() with a prev_span()-on-EOF fallback site by site.
• P017/P018/P019 are double-booked between the parser (lambda capture, variadic, my originally-planned slot) and the import system (SPEC.md / ai.txt / main.rs). Worth untangling in a separate cleanup PR; not in scope here.

[codecov]

Codecov Report

❌ Patch coverage is 98.33333% with 1 line in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/parser/mod.rs	98.33%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!