fix: preserve raw text for t-typed CLI args

[danieljohnmorris]

Summary

CLI args declared as t used to be silently coerced to Number when the raw input parsed as a finite float. f arg:t with input 2 arrived at runtime as Number(2.0), so num arg returned nil, which then collapsed the surrounding ?r{~i:..;^e:..} match into a silent nil return because there is no nil arm. Static types were fine, so the verifier never caught it.

This was the longstanding 🔴 friction documented in ilo_assessment_feedback.md since v0.10, and the same friction surfaced again in the v0.11.1 interactive-CLI rerun. Highest-leverage manifesto win here: zero retries for any agent that declares arg:t and passes digit-shaped input. Pre-fix the agent had to invent a workaround (a separate i:n param threaded through the dispatcher purely to keep the type honest). Post-fix the declared type does what it says.

Repro

f arg:t>n;r=num arg;?r{~i:i;^e:0 - 1}

Before:

$ ilo prog.ilo f 2
nil

After:

$ ilo prog.ilo f 2
2
$ ilo prog.ilo f abc
-1

Same result on --run-tree, --run-vm, --run-cranelift.

What's in the diff

• fix: preserve raw text for t-typed CLI args — adds parse_cli_args_typed which consults declared param types during CLI parsing. t params keep the raw string verbatim as Text (the existing quoted-string strip is preserved for back-compat); L _ params still wrap a scalar; everything else parses by the same rules as before. Switches the six production callsites that used to do parse-then-coerce-by-type to the new typed parser. The legacy coerce_cli_args is kept under #[cfg(test)] so its list-wrap tests keep pinning that behaviour from pre-parsed values, but has no remaining production callers. Unit tests cover text-preservation across numeric, bool, nil, and list-shaped inputs, plus number passthrough, list wrap, mixed signatures, and the no-func-name fallback.
• test: cross-engine regression for t-typed CLI args — tests/regression_cli_text_arg.rs exercises all four engines for the num round-trip (both digit and non-digit input), identity on bool-shaped and nil-shaped inputs, and a sanity check that n-typed params still parse as numbers.
• examples: t-typed CLI args preserve raw input — examples/cli-text-arg.ilo documents the now-correct behaviour and is picked up by tests/examples_engines.rs for a third regression layer across every engine.

Test plan

• cargo test --release --features cranelift — full suite green (2866 lib tests + integration crates)
• cargo clippy --release --features cranelift --all-targets -- -W clippy::all — clean
• cargo fmt --check — clean
• Manual repro across all four engines (default, --run-tree, --run-vm, --run-cranelift) — all return 2 for f arg:t with input 2

Follow-ups

None. The fix is surgical: type-aware parsing for t only, every other declared type's behaviour is byte-for-byte unchanged.

[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!