2i-issues Pyrope lowering & cgen emission defects (non-LEC)

A (registered-output interface) and B (hierarchical enums) are the only remaining work — both still hard-error in lowering (the latest patch round cada30102, import + ABC sra/mul ops, did not touch their code paths). Each is valid, documented Pyrope; they have no ordering between them, so pick either and work it end-to-end from its subtask below.

re-verified again 2026-06-20 — direct compile of both repros confirms neither has cleared: A aborts with reg-output-cycle (state register 'count' … homes at stage 0 but its declared landing cycle 0 requires home -1) and B with undefined-read (read of undefined variable 'eagle'), so both rows stay open. The companion TBD-feature audit (2026-06-20) — documented-but-unimplemented features, a category distinct from the A/B lowering defects — now lives in the index's TBD features still pending summary. B is itself a TBD (hierarchical enums) and is intentionally not given a new fail-by-design test there — it is already tracked as subtask B here.

Scope. This page deliberately excludes the LEC-engine precision issues (cvc5 false-refutes on hotmux / FSM-state / memory, the \x01nxt: phantom-output, the lgyosys no-reset-flop disagreement). Those are checker bugs, not compiler bugs, and live in the repo-root issues.md (Group C). Every code pointer below was re-verified by reading the source on 2026-06-20 against a fresh bazel build -c dbg //lhd:lhd.

Status snapshot (2026-06-20)

Depends on

None. A and B are each self-contained and the only remaining work.

Subtasks — pending

• A — registered-output interface form -> (reg name@[0]). Repro unexpected_semantics/regout_form/impl.prp — the canonical counter2 pattern from the pyrope SKILL.md and docs/pyrope/04-variables.md:

  pub mod top(enable:bool) -> (reg count:u8@[0]) {
    if enable { wrap count += 1 }
  }

  lhd compile … --emit verilog:gen.v hard-errors: state register 'count' of 'impl.top' homes at stage 0 but its declared landing cycle 0 requires home -1. The in-body form (reg count:u8=0; value=count) works; only the interface spelling fails. Root cause: the state-register output discharge applies the feedforward invariant home.min + 1 == rec.min at upass/tolg/upass_tolg.cpp:4525 (diagnostic :4526-4534). A state register pins σ(q)=σ(din) with no flop crossing (:4434-4437, documented :4085-4086), so a counter whose din is the same-cycle increment of q has home.min=0; the +1 rule then demands home.min == -1. The comment at :4093-4094 ("discharges against home+1") directly contradicts the no-crossing model at :4085-4086. Fix: in the state-reg branch (:4514-4535) test home.min == rec.min (so @[0] matches home 0) instead of home.min + 1 == rec.min; update the comments. Before landing, confirm no passing test relies on the home+1 spelling for a state-reg output @[N>0].
• B — hierarchical (nested-tuple) enum members. Repro unexpected_semantics/hier_enum/impl.prp — verbatim from docs/pyrope/03-bundle.md:600-616:

  enum Animal = (,bird=(,eagle, ,parrot) ,mammal=(,rat, ,human))
  pub comb top(s:bool) -> (r:u6) { r = int(Animal.bird.eagle) }   // == 0ub000011

  Hard-errors read of undefined variable 'eagle'. Root cause: nested-enum detection at inou/prp/prp2lnast.cpp:4553 calls enum_definition_node_of(e.value_node), which (:4362) returns non-null only for an explicit enum(...) node — not for the bare-tuple value (,eagle,,parrot) the spec prescribes. So the hierarchical branch is skipped, control falls into the "Expression value" branch (:4593-4605) which lowers the tuple via expr_to_node, reading eagle/parrot as variables (read_sites_ :5372) → undefined-read abort (:859-864). The explicit nested-enum(...) form already works, and the runner read side (upass/runner/upass_runner.cpp:168, :2606) already supports hierarchical enums — only the bare-tuple detection is missing. Fix: make the nested test at :4553 (or enum_definition_node_of) also accept a tuple value and parse its members with the existing process_enum_assignment logic (:4657-4699), reusing the parent-bit / __enumval machinery so inner names go to fields, not expr_to_node.

Verification

Each fix should both compile and prove equivalent to the directory's ref.v (which encodes the intended hardware). The re-test harness from the 2026-06-20 round lives in the session scratchpad (retest.sh = compile + both LEC engines; comb_oracle.py = exhaustive combinational iverilog; seq_oracle.py = clocked random iverilog). For one case:

# A, B: must stop hard-erroring
lhd compile <dir>/impl.prp --emit verilog:gen.v

# once they compile: prove equivalent to ref.v with BOTH engines
lhd lec --impl <dir>/impl.prp --ref <dir>/ref.v \
        --impl-top impl.top --ref-top top --set lec.solver=cvc5
lhd lec --impl <dir>/impl.prp --ref <dir>/ref.v \
        --impl-top impl.top --ref-top top --set lec.solver=lgyosys

Add a permanent regression for each: a glob-matched pair under inou/prp/tests/equiv/ (forward LEC) — these auto-generate a prp-<name> bazel target with no BUILD edit. Keep the full prp suite green (bazel test //inou/prp/... //upass/... //pass/...) and watch for no regression in the other (already-passing) unexpected_* cases.

Done when

The remaining work is A and B only. Each must compile and emit a netlist that is iverilog-correct and LEC-equivalent to ref.v under both solvers, with its checked-in regression flipping green; then its row is removed from the index's "Unimplemented / mis-lowering forms" tables. This page is deleted when both A and B clear.