builtins: rgxall1 flat-capture + ct count-by-predicate (rerun6 batch)

[danieljohnmorris]

Summary

Two builtin additions from the rerun6 persona round, both manifesto-aligned token-economy wins on multi-rerun-old shapes:

• rgxall1 pat s -> L t — flat first-capture-group convenience over rgxall. Closes the ext xs:L t>t;hd xs helper that the html-scraper persona kept paying ~30 tokens to re-declare across five reruns (rerun5, line 4988).
• ct fn xs -> n — count-by-predicate, parallel to flt's 2/3-arg shape. Avoids the intermediate L b allocation len (flt p xs) pays on every call. Originating ask: bioinformatics rerun6 tm=ct has-tm seqs over 20k proteins (line 5028).

Named ct rather than the persona's wished-for cnt: cnt is reserved as the loop-continue keyword (src/parser/mod.rs:3507). Two chars beats three, no parser surgery, the loop-control reservation stays intact (regression-covered).

Repro before/after

rgxall1. html-scraper persona, "extract every match of one capture group":

-- Before (rerun5, 40 LOC; ext is the only helper that pays no semantic rent):
ext xs:L t>t;hd xs
titles=map ext (rgxall "<a class=\"titleline\">([^<]+)</a>" html)

-- After (PR):
titles=rgxall1 "<a class=\"titleline\">([^<]+)</a>" html

ct. bioinformatics persona, "how many proteins have a TM helix":

-- Before:
tm=len (flt has-tm seqs)   -- allocates L b sized to the keepers

-- After:
tm=ct has-tm seqs           -- straight counter, zero intermediate alloc

What's in the diff (per commit)

• f3aed43 builtins: rgxall1 pat s for flat first-capture-group extraction — adds Builtin::Rgxall1, tree-walker arm, tree-bridge eligibility, propagate-error allow-list entry. 0 groups → flat L t of whole matches; 1 group → flat L t of capture-1 strings; 2+ groups → runtime ILO-R009 pointing at rgxall. 8 cross-engine regression tests + examples/rgxall1-flat-captures.ilo.
• e86dba9 builtins: ct fn xs for count-by-predicate without the filtered-list alloc — adds Builtin::Ct with predicate signature identical to flt's, 2-arg and 3-arg closure-bind shapes. Tree-bridge wired for VM/Cranelift parity; runtime non-bool predicate errors join the propagate allow-list. No count long-form alias on purpose (caught a real false-positive that would have trampled the user-fn name in examples/unq-numbers.ilo). 10 cross-engine regression tests including a cnt-as-continue coexistence regression guard + examples/ct-count-by-predicate.ilo.
• a0a1ee2 docs: sync SPEC.md, ai.txt, SKILL.md for rgxall1 and ct builtins — every keep-in-sync surface gets new rows in the builtin tables and the new names in the cheatsheet/HOF lists. Site builtins.md lands in a follow-up ilo-lang/site PR.
• 4f91c38 tests: tighten ct non-bool predicate assertion to anchor on "ct:" — rust-review caught that bare stderr.contains("ct") would match the substring inside "expected" and let unrelated wording satisfy the assertion. Anchor on the literal ct: prefix from the ILO-R009 message instead.

Manifesto framing

These two compose: any persona that today writes len (flt … rgxall …) or map (hd …) rgxall … is paying both a helper-decl tax and a redundant allocation. With rgxall1 + ct the canonical "scrape one HTML page, count items matching a predicate" pipeline shaves ~50 tokens and one O(n) alloc per occurrence. Across the agent population shape-of-task that is real compression.

Both ride the tree-bridge — same dispatch contract as rgxall, flt 3, grp 2. No backend codegen changes, no JIT lowering complexity, full cross-engine parity for free.

Test plan

• cargo test --release --features cranelift — full suite green post-rebase
• cargo clippy --features cranelift --all-targets -- -W clippy::all — clean
• cargo test --features cranelift --test regression_rgxall1 — 8/8 pass on tree/VM/Cranelift
• cargo test --features cranelift --test regression_ct — 10/10 pass on tree/VM/Cranelift, plus cnt-as-continue coexistence regression guard
• cargo test --features cranelift --test examples_engines — both new examples run on every engine
• Builtin::ALL tag round-trip preserved (entries appended after Sleep, no reordering)
• Self rust-review pass complete; one false-positive test assertion caught and tightened

Follow-ups

• PR B (ILO-W002 unused-binding warning, security-researcher rerun5 ask) branches off main after this merges, per the two-PR split agreed at the gate.
• Site doc sync (ilo-lang/site builtins.md) lands in a follow-up PR against the site repo once this merges.

[codecov]

Codecov Report

❌ Patch coverage is 71.66667% with 34 lines in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/interpreter/mod.rs	70.66%	22 Missing ⚠️
src/verify.rs	62.50%	12 Missing ⚠️

📢 Thoughts on this report? Let us know!