self-host parity: higher-order functions, doc continuation, discharge fence

[jasisz]

Rewritten 2026-08-11. Part 2 of the original issue is fixed and has been removed — leaving it would send whoever picks this up chasing a module-resolution bug that no longer exists. The line anchors in part 1 have been refreshed, and the two divergences that are actually left have been added. See the comment below for the sweep this is based on.

A VM-vs-self-host corpus audit during 0.24 prep found the self-hosted interpreter (self_hosted/*.av, compiled to Rust under src/self_host/) lags the other backends on several language features. 0.24 "Divide" closed the tractable ones. This issue tracks what is left.

Regen after editing self_hosted/*.av: aver compile self_hosted/main.av --target rust --output self_hosted/out --module-root self_hosted --with-self-host-support --guest-entry runGuestCliProgram --with-replay --policy runtime, then copy self_hosted/out/src → src/self_host/ and cargo build.

1. Higher-order functions (Fn(...) parameters)

Still reproduces exactly as filed. The host prints 12; --self-host fails at parse with Expected ',' or ')' in parameter list.

fn dbl(x: Int) -> Int
    x * 2
fn applyTwice(f: Fn(Int) -> Int, x: Int) -> Int
    f(f(x))

Two layers, both unchanged:

• Parser: skipTypeExpr in self_hosted/domain/parsermatch.av:224 handles TkIdent and TkLParen but has no TkFn case, so a Fn(A) -> B parameter type never parses. Add a TkFn branch that skips Fn ( types ) -> ret. (TkFn/TkArrow tokens already exist.)
• Eval: ValFnRef is created at self_hosted/domain/eval/common.av:23 when an identifier resolves to a fn, but never consumed. callWithArgs / callWithArgsNormal (self_hosted/domain/eval/core.av:387 / :394) resolve the callee only as a top-level fn or a builtin and do not receive env, so a call f(x) where f is a local bound to ValFnRef falls through to callBuiltin("f", …) and fails. Both the named-env path (evalCall, core.av:608) and the slot-env path (evalCallSlot, core.av:1047) would need it, and the resolver may also reject a param as a call target.

The self-host is slot-resolved and was not designed for first-class fn values; this is a feature add, not a patch.

Worth weighing before anyone starts. Exactly one file in the 112-file example corpus uses an Fn(...) parameter — examples/formal/oracle_independent_products.av:11 — and it has no main, so it is not even in the parity sweep. examples/core/lambda.av does run under --self-host, but it is a lambda-calculus interpreter over a type Term and exercises no Aver first-class functions at all. So this is a real gap with almost no corpus pressure behind it.

2. Multi-line ? documentation continuation

examples/refinement/natural_app.av fails under --self-host with Expected expression, got Indent. Bisected to a ? doc string followed by an indented continuation line, the shape fn add uses in examples/refinement/natural/natural.av:

fn add(a: Nat, b: Nat) -> Nat
    ? "line one"
      "line two"

exposes [f], exposes opaque [Nat] and verify ... law ... given all parse fine — it is only the continuation. Small and self-contained, and fixing it takes natural_app.av to parity.

3. Literal smart-constructor discharge is fenced off

examples/data/json.av diverges because the self-host refuses 0.28's literal smart-constructor discharge, loudly and deliberately — reject_literal_refinement_discharge at src/main/commands.rs:1292, with the reasoning at :1277-1291. Fail-closed with a message, not a silent miscompile, so this is correct behaviour rather than a bug.

It does raise a standing question this issue should carry: whether the self-host is meant to track host typing rules as they land. The note at commands.rs:1287 puts the cost plainly — mirroring the discharge "is not the three syntactic predicates the literal-divisor rule needed; it is the whole recognizer". That is a tax on every future discharge rule, and it should be decided rather than absorbed.

Verification

Corpus parity sweep over every examples/**/*.av with a fn main, host versus --self-host, diffing stdout:

• 43 exercised, 37 exact match
• of the 6 differing: date.av is clock nondeterminism, and app.av / app_dot.av / pricing_app.av differ only by a host-side stderr warning about unchecked dependency-module verify blocks — their stdout is byte-identical
• leaving 2 genuine: natural_app.av (section 2) and json.av (section 3)

Regen gate: AVER_SELF_HOST_REGEN=1 cargo test --test rust_self_host_regen -- --ignored.

[jasisz]

Rewrote the body. Part 2 — multi-module type and symbol resolution — is fixed, and the issue was actively misleading with it still in.

The sweep it is based on: every examples/**/*.av with a fn main, host versus --self-host, diffing stdout. 43 exercised, 37 exact match. Of the 6 that differ, date.av is clock nondeterminism, and app.av / app_dot.av / pricing_app.av differ only by a host-side stderr warning about unchecked dependency-module verify blocks — comparing stdout alone, all three match byte for byte. So the expected String, got E cross-module generic instantiation this issue was half about is gone, and so is the Unknown identifier 'Refinement' module-prefix failure.

natural_app.av does still diverge, but not for the reason the issue gave. Bisected it: exposes, exposes opaque and verify ... law ... given all parse; what fails is a multi-line ? doc continuation. Unrelated to module resolution, much smaller, and now section 2.

The other survivor is new since this was filed — json.av, where the self-host refuses the 0.28 literal smart-constructor discharge. That one is deliberate and fail-closed, so it is written up as a scope question rather than a defect.

Part 1 is untouched and still reproduces verbatim; only the line anchors moved (core.av:341/348 → 387/394). Added one thing worth knowing before anyone starts on it: exactly one file in the whole example corpus uses an Fn(...) parameter, and it has no main.

[jasisz]

Reviewed at main db6fb85b615cfc04c84cb67fb99b1c56624f8b59. This remains open: skipTypeExpr in self_hosted/domain/parsermatch.av:224 still has no TkFn case, and reject_literal_refinement_discharge remains in src/main/commands.rs:1497. This was a targeted source audit, not a new full parity sweep; it does not establish the current doc-continuation status. The regeneration helper is now python3 tools/regenerate_self_host.py (--check for comparison), rather than manually copying generated source.