fix(crux-judge): scan_once and main back at grade B — CB-200 616 again (#5173)

[noahgift]

What

pmat comply CB-200 (TDG grade gate, min_grade = "B") went from 616 to 618 definitions below the floor between v0.70.3 and cc71bde. Of the sub-B definitions new in that range, two sit just under the floor: aprender-crux-judge pyjson::Doc::scan_once and main, both B- (TDG 74.5, cyclomatic 18). This PR refactors those two back to grade A, which returns the count to 616. The [tdg] baseline is not touched.

• scan_once: the six bare-word constants move to scan_literal + literal_value. They are still tried before a number, so -Infinity is not read as -.
• main: argument decoding (utf8_args), the sandbox command (sandbox_cmd) and the collect exit mapping (collect_cmd) become helpers under one run dispatch. Exit codes are unchanged.
• New test bare_word_literals_decode covers every literal plus negative numbers. Before this PR, no test covered the literal arms.

Red before / green after (pmat 3.42.0, pmat query x then pmat comply check --format json, fresh index each)

RED   cc71bde029: CB-200: TDG Grade Gate | Fail | 618 definition(s) below minimum grade B
GREEN 5d39bff0c1: CB-200: TDG Grade Gate | Fail | 616 definition(s) below minimum grade B
GREEN 2f449abb78: CB-200: TDG Grade Gate | Fail | 616 definition(s) below minimum grade B
ref   v0.70.3   : CB-200: TDG Grade Gate | Fail | 616 definition(s) below minimum grade B

Grades after: main A+, utf8_args A+, sandbox_cmd A+, collect_cmd A, run A-, scan_once A, scan_literal A+, literal_value A-.

Checks

• The crate is excluded from the workspace and has its own lockfile, so the commands use --manifest-path tools/aprender-crux-judge/Cargo.toml. With it: cargo fmt -- --check rc 0, cargo clippy --locked --all-targets -- -D warnings rc 0, cargo test --locked 40 passed (39 existing + 1 new).
• cargo fmt --all -- --check rc 0.
• Guard tree (scripts/ci_guards.sh guard-tree) on 2f449ab: 176 checks, 2 failed. The two are check_baseline_ratchets.sh and check_complexity_ratchet.sh, each refusing because its baseline was recorded under pmat 3.41.1 and the local pmat is 3.42.0 (instrument mismatch, unrelated to this diff).

Note: pmat's comply index cache is keyed per checkout and is not refreshed by a new commit. The GREEN line above was measured after dropping that cache. Without that step the count read 618 on the fixed tree.

Closes #5173

ont-delta: none — code refactor in one tool crate; no doc, spec or contract changes

🤖 Generated with Claude Code

[noahgift]

APPROVE seat 2, w71-e3b, 2f449ab

Read the diff at 2f449ab (2 files under tools/aprender-crux-judge).

• main → utf8_args + run: the dispatch order is unchanged (eval/extract/lint → oracles, certify/check → certify, sandbox → sandbox_cmd, else parse → collect). collect::<Result<Vec<_>,_>> stops at the first non-UTF-8 argument, so usage_error still fires once, for the first one. The exit codes are unchanged: decline 2, crash 1, u8::try_from(rc).unwrap_or(1).
• scan_once → scan_literal: the six words are still tried before match_number, and -Infinity before a bare -. The literals are ASCII, so idx + lit.len() advances correctly. A first char that matches no word falls through to the number path, as before. The new bare_word_literals_decode test covers every arm plus negative numbers.
• It fixes the code and leaves the [tdg] baseline alone: no raise and no ratchet, as row P3 requires. The RED 618 → GREEN 616 lines are in the body.
• One note, not a blocker: the two local guard-tree reds are pmat 3.41.1 vs 3.42.0 baseline mismatches and do not come from this diff. CI's guard-tree on this head is what counts.

[noahgift]

seat 1 (P3), aprender-a7 APPROVE @ 2f449ab

Re-ran on a detached checkout of this head (pmat 3.42.0):

• pmat query grades on this head: main A+, utf8_args A+, sandbox_cmd A+, collect_cmd A, run A-, scan_once A, scan_literal A+, literal_value A-. These match the body, and none is below B.
• cargo fmt -- --check rc 0, cargo clippy --all-targets -- -D warnings rc 0, cargo test 40 passed, including bare_word_literals_decode. All three ran with --manifest-path tools/aprender-crux-judge/Cargo.toml. I did not reproduce the comply count of 616 myself; I checked the two definitions it names.

Behaviour, read against the old code:

• scan_once: scan_literal tries every bare word with the same full-string starts check, before match_number. So -Infinity still wins over a number, and -1/-1.5 still fall through to match_number; the new test pins both. Exactly one literal can match at any index, so the order of the array does not matter.
• main/run: the dispatch order is unchanged (eval/extract/lint, then certify/check, then sandbox, then parse). The sandbox extra-argument refusal, the non-UTF-8 usage exit, decline→2, crash→1 and u8::try_from(rc).unwrap_or(1) are all preserved.
• [tdg] baseline is untouched.

Notes, none blocking:

• Both the RED and GREEN lines in the body read CB-200 ... Fail. The fix restores the count to the v0.70.3 figure, 616; it does not turn CB-200 green. That is what this row asked for, but "GREEN" there means "back to 616".
• The two guard-tree refusals (check_baseline_ratchets.sh, check_complexity_ratchet.sh: baseline recorded under pmat 3.41.1, local 3.42.0) are not caused by this diff. If either sits on the merge path, they need their own row.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5180 head=2f449abb786897f029da40e745b91d85d81fe659 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.