Support OPA v1 'if' rules as single rules

[lovesegfault]

Fixes #20.

Problem

A v1 rule like

allow := true if {
    input.admin
}

parsed as two rules: a value head (allow with a := true body) plus a phantom rule whose head variable is if. Since OPA 1.0 rejects bodied rules without if, virtually every current-syntax policy produced these split trees. The same split affected else chains, and else := 2 if { ... } clauses could swallow following rules.

Approach

The head grammar was missing the plain-var value form (var [":=" term] ["if"] from the OPA spec) — the bracket/args head forms already carried [rule-head-comp] [if], but a plain allow := true if had nowhere to go.

• New rule alternative for the v1 comp+if form: value bound in the head (rule_head_v1, aliased to rule_head so consumers see the usual node), then a first body restricted to a braced query or single literal (rule_body_v1, aliased to rule_body). Restricting the first body means a following v0 rule that happens to be named if (a := 1 then if := 2) cannot be fused in — its := 2 is not a valid v1 body, so that GLR path dies and the two-rule parse wins. Dynamic precedence prefers the one-rule parse over the phantom split.
• rule_head drops prec.right: it statically resolved var • := toward the head path, which would break plain constants (x := 1) once the comp+if form exists. The head forms instead share one prec.dynamic(1) so brackets, args, and if stay in the head rather than demoting to body literals (pinned by the new v0_heads.txt corpus cases at end-of-file, where the tie-break used to be arbitrary).
• rule_head_comp accepts full expressions (f(x) := x + 1 if x > 0, roles := input.roles | {"anonymous"} if { ... }), not just plain terms.
• The rule body value branch accepts = alongside :=, fixing plain unification constants (a = 1, a = 1 { true }) and terminal else = 2 clauses.
• else clauses take an optional if (else := 2 if { ... }, else if input.b).

Corpus

• conditionals and ordered (else) previously codified the split-parse trees; they now assert single rules.
• union (Parsing error - union #8) and reference heads (Parsing error - reference heads #7) remain broken per their own issues; their expectations re-pin the current recovery trees (the reference heads recovery actually improves — one rule with the ERROR confined to the dotted segments instead of a phantom if split).
• New v1_rules.txt (9 cases: braced/unbraced bodies, else-if chains, default+conditional, infix values, unification forms, constants staying separate) and v0_heads.txt (5 cases guarding v0 head shapes and the rule-named-if non-fusion).

53/53 corpus tests pass. src/ regenerated with tree-sitter generate.

Known remaining gaps

• Bodiless partial-set rules (deny contains "msg" with no body) still swallow the following statement — unchanged from before; fixing it trips over the grammar's repeat1 body requirement and deserves its own change.
• A v0 function literally named if directly following a constant rule can fuse into it. if is a reserved word in v1, so this only affects legacy policies defining a function named if.

Note: this branch is independent of #22; whichever merges second just needs tree-sitter generate re-run to resolve the generated-file overlap.

[FallenAngel97]

Hello @lovesegfault !

Thank you for your PR.

Can you please rebase it with the latest master? I fixed the yarn.lock, why it was failing - it was resolving for some reason towards the npm of Apple

[lovesegfault]

done! :)