Follow-ups for #776: RVM suspendable partial-rule parity, else_bodies metadata safety, json.patch JSON-Pointer validation, AST/RuleInfo semver

[anakrish]

Follow-up items from the review of #776 (json.patch builtin + partial-rule multi-body fix). None of these blocks correctness of normal policy evaluation — #776 passes the full opa.passing suite (2875/2875) and all targeted tests — so the plan is to merge #776 and address these here in a follow-up PR.

Findings surfaced by the repo's code-review and deep-review skills, deduplicated and calibrated below.

Should fix (before RVM suspendable / serialized-program modes are relied upon)

1. RVM suspendable path skips later successful bodies of partial rules

• Location: src/rvm/vm/rules.rs — rule_frame_after_success (current_body_index = definition_bodies.len();)
• Issue: feat: json.patch builtin with Rego set support + partial-rule multi-body fix (picks up #442) #776 fixed multi-body partial-rule accumulation in the run-to-completion path (execute_rule_definitions_common) but the suspendable frame path still jumps to the end of the definition after the first successful body, dropping contributions from later independent partial-object/set bodies.
• Reachability: set_execution_mode(Suspendable) is currently only used in tests; default eval_rule uses RunToCompletion, so this is not a production allow/deny bug today. It is a dual-path divergence that will produce wrong results if suspendable/host-await evaluation is used.
• Fix: Mirror the run-to-completion logic — for PartialSet/PartialObject advance to the next body unless the next body is marked else; keep first-success for complete/function rules. Add a suspendable-mode test with multiple independent partial bodies.

2. else_bodies metadata safety on deserialized / hand-built programs

• Location: src/rvm/program/types.rs (#[serde(default)] pub else_bodies), consumed in src/rvm/vm/rules.rs via ...unwrap_or(false)
• Issue: A deserialized older program (or hand-built RuleInfo) with absent/short else_bodies defaults every marker to false (= "not else"), so partial-rule else bodies can run as independent bodies after a prior body succeeds.
• Fix: Validate else_bodies shape against definitions at load/execute time and reject incompatible metadata, or bump the serialized program version with a safe compatibility default.

Nice to have / spec-conformance

3. Tighten JSON-Pointer validation in json.patch

• Location: src/builtins/json_patch.rs — parse_path, array_index
• The leading-/-optional relaxation is intentional (matches OPA, documented in the PR) — keep it. But the following are more lenient than RFC6902 and could yield a wrong non-Undefined result on malformed input:
  • Invalid escape sequences (~2, bare ~) are passed through literally instead of failing.
  • Signed / leading-zero array indices (+1, -0) parse as valid indices.
  • Multiple leading slashes are collapsed (//a → a).
• Fix: Validate ~ is followed only by 0/1; validate string array indexes against the RFC6902 grammar; add negative test cases (a/b, //a, /~2, /+1, /-0) asserting Undefined.

4. Malformed-patch vs resource-limit ordering

• Location: src/builtins/json_patch.rs apply (builds full EditTree before validating op shape) + objects.rs wrapper
• Issue: A malformed patch on a very large target can surface LimitError (correctly propagated) before the documented malformed-patch Undefined. Low impact.
• Fix: Pre-validate operation shape (op/path/value) before constructing the edit tree, or document that resource limits preempt malformed-patch Undefined. Add a large-target malformed-patch memory-limit test.

Semver / schema (safe to defer, crate is pre-1.0 @ 0.11.0)

5. Public RuleInfo gained a required field

• Location: src/rvm/program/types.rs — pub else_bodies: Vec<Vec<bool>>
• Breaks external struct-literal construction. Consider #[non_exhaustive] / constructor-only.

6. AST JSON schema changed without version bump

• Location: src/ast.rs (RuleBody.is_else), src/engine.rs (get_ast_as_json, version: 1)
• New serialized field on RuleBody while the AST version stays 1. Bump/document the schema version, or skip_serializing_if the default.

---

Filed as the tracked follow-up for #776. Skill reports and full traces available in the review comment on the PR.