Targeted hint on ILO-T003 ternary branch-type mismatch

[danieljohnmorris]

Summary

Before: a ternary like ?h c 1 "x" produced ILO-T003 with the generic hint "both branches of a ternary must return the same type". Agents either guessed which side to convert (often producing a follow-on ILO-T013) or restructured unnecessarily. Common gotcha across personas; addressing pending.md #5h.

After: the verifier reads the two branch types and emits a targeted suggestion.

• n vs t surfaces both conversion directions (str <num-branch> for the text intent, default-on-err (num <text-branch>) <fallback> for the number intent) so the agent picks whichever matches the function's return type. The hint explicitly mentions num returns R n t, which avoids the very common follow-on bug of treating the parse result as a bare number.
• Everything else (bool vs text, L n vs M t n, two named records, R T E vs n, ...) falls back to restructure advice. The only builtin scalar coercions in ilo are str (n→t) and num (t→R n t), so suggesting a coercion outside that pair would just trip ILO-T013. Better to point at [...] / record / O T / R T E framings explicitly.

Manifesto framing: cheaper hint, fewer retries. The agent now reads the diagnostic once and writes the fix, rather than walking through two or three failed compile cycles.

Repro

ilo 'f c:b>t;?h c 1 "x"' --vm f true

Before:

ILO-T003 ternary branches have different types: n vs t
  hint: both branches of a ternary must return the same type

After:

ILO-T003 ternary branches have different types: n vs t
  hint: both directions are available: `str <num-branch>` makes both
        text (cheapest if the function returns `t`), or parse the text
        side with `default-on-err (num <text-branch>) <fallback>` to
        make both number (since `num` returns `R n t`). Pick whichever
        matches intent, or restructure to wrap each branch in a
        list/record/`O T` to keep both shapes

What's in the diff

• fix(verify): targeted hint on ILO-T003 ternary branch mismatch — new ternary_mismatch_hint helper in src/verify.rs; the existing Expr::Ternary mismatch path calls it instead of hard-coding the generic string. Pure diagnostic, no codegen / parser / AST surface change.
• test: regression coverage for ternary type-mismatch hint — six new tests in tests/regression_ternary_type_hint.rs pinning the four shapes (n-vs-t in either branch position, bool-vs-text fallback, list-vs-map fallback, matching-branch false-positive guard). Adds examples/ternary-types.ilo showing the canonical correct shapes; the examples_engines harness exercises it across every available engine on each CI run.
• doc: note ternary branch-type hint in SPEC ternary section — short paragraph in SPEC.md next to the existing ILO-T038 paragraph; ai.txt regenerates from SPEC.md via build.rs.

Test plan

• cargo test --release --features cranelift --test regression_ternary_type_hint — 6/6 pass
• cargo test --release --features cranelift --lib verify — 429/429 pass (existing ternary_branch_type_mismatch_error still green)
• cargo test --release --features cranelift --test examples_engines — passes, picks up the new ternary-types.ilo
• cargo fmt --check clean, cargo clippy --all-targets -- -D warnings clean
• Full cargo test --release --features cranelift — only failures observed are pre-existing on main and unrelated (spec_reserved_short_names_match_builtin_registry, body_is_thin_bootstrap)

Follow-ups

• None required. If a future PR widens the set of builtin scalar coercions (e.g. bool→text), update the ternary_mismatch_hint strategy to surface the new direction.

[codecov]

❌ 1 Tests Failed:

Tests completed	Failed	Passed	Skipped
7575	1	7574	0

View the top 1 failed test(s) by shortest run time

ilo::skill_md::body_is_thin_bootstrap

Stack Traces | 0.006s run time

thread 'body_is_thin_bootstrap' (65130) panicked at tests/skill_md.rs:254:5:
SKILL.md body is 10925 bytes; bootstrap shape should stay well under 8 KB
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}