89%

Section 02: Import-Sigs Traversal — Cross-Module Imported Generic Top-Level Mono (BUG-04-119)

Status: In Progress (§02.N close-out pending — /tpr-review + /impl-hygiene-review items unchecked per frontmatter; plan_corpus check + BUG-04-119 tracker annotated + plan annotations swept flipped [x] 2026-05-16; test-all.sh gated by parallel-session baseline-red) Goal: Close the cross-module gap. ori build’s mono-collection pipeline registers imported generic top-level functions reachable from host-module callers, so LLVM module verification succeeds before realize.

Success Criteria: see frontmatter — all items contribute to mission criterion 1 (top-level imported generics).

Context: Per §01.4 BUG-04-119 entry (also documented in §01.1 F2), compile_common.rs:217-246 constructs import_sigs for imported functions (line 217) and consumes them via collect_mono_functions (line 246). The function’s three-input lookup chain (function_sigs, impl_sigs, none) silently skips instances whose fn_name is from another module, yielding E5001 unresolved function 'assert_eq' in apply/invoke — missing mono instance at LLVM verification. This is the architectural skeleton — §03/§04/§05 reuse the lookup-chain extension pattern this section establishes.

Scope boundary (per Decision 02): §02 owns TOP-LEVEL imported generic functions only. Imported generic METHODS are routed to §03 (section-03-inherent-method-mono.md) which owns the inherent-method-on-generic-type dispatch surface and gains AOT export-surface work for multi.rs::public_methods.

Linkage mechanism (per Decision 01): Imported mono instances use the body-import path (Option B) — compile_common.rs constructs imported_mono_fns: Vec<ImportedMonoFn> from per-module CanonResult retention; arc_lowering specializes each imported mono locally against the SOURCE module’s CanonResult via lower_to_arc(mono_fn.mangled_name, &mono_fn.sig, mono_fn.original_name, source_canon, ...). The host module emits the body via the existing declare_mono_functions path. No declare_imported_mono function is created.

Reference implementations (graph-first per intelligence.md):

  • Rust rustc_monomorphize::collector enumerates imported items via tcx.upstream_monomorphizations — same shape: imported sig table consulted alongside local sigs. Verify with scripts/intel-query.sh similar "collect_mono_functions" --repo rust --limit 5.
  • Swift SILGen treats imported generic specializations as first-class call-site targets, registered uniformly with local generics.

Depends on: §01 (cite-back: §01.1 API surface inventory + §01.4 BUG-04-119 classification).

Cross-cutting concern (per blind-spots): JIT/AOT path duplication is RESOLVED by §02 — both paths use the body-import carrier shape. JIT’s imported_mono_fns (oric/src/test/runner/imported_mono.rs:22) carrier is promoted to AOT production scope at compile_common.rs per Decision 01 Option B. Single mechanism, not parallel.

Intelligence Reconnaissance

Queries planned (run AT section-open-for-work per compose-intel-summary.md):

  • scripts/intel-query.sh callers "collect_mono_functions" --repo ori — confirm evaluator/compile.rs:230, codegen_pipeline/mod.rs:138, arc_lowering.rs:272 remain the three call sites (re-verify post-§01).
  • scripts/intel-query.sh callers "run_borrow_inference" --repo ori — confirm codegen_pipeline/mod.rs:319 remains the sole production call site; identify any test-only consumers.
  • scripts/intel-query.sh callers "lower_and_infer_borrows" --repo ori — full call-site set for the test runner.
  • scripts/intel-query.sh file-symbols "compiler/oric/src/commands/compile_common" --repo ori — inventory import_sigs construction surface AND identify per-module CanonResult retention points.
  • scripts/intel-query.sh file-symbols "compiler/oric/src/test/runner/imported_mono" --repo ori — inventory ImportedMonoFn carrier + build_imported_mono_functions for AOT promotion.
  • scripts/intel-query.sh similar "collect_used_items" --repo rust --limit 3 — Rust’s imported-mono enumeration prior art.

Results summary (≤500 chars, recorded 2026-05-15; line refs refreshed 2026-05-16 to current HEAD) [ori]: callers of collect_mono_functions confirmed at evaluator/compile.rs:230, codegen_pipeline/mod.rs:143, arc_lowering.rs:272 — three production sites; run_borrow_inference sole production caller is run_codegen_pipeline at codegen_pipeline/mod.rs:372; import_sigs construction surface confirmed at compile_common.rs:206-223 (single-module entry; multi-module retention via CompiledModuleInfo at multi.rs:201-219 + consumer at multi.rs::build_imported_mono_state at multi.rs:722); ImportedMonoFn carrier at imported_mono.rs:22, build_imported_mono_functions at codegen_pipeline/imported_mono.rs.

02.1 collect_mono_functions signature + traversal extension

File: compiler_repo/compiler/ori_llvm/src/monomorphize/mod.rs:71

  • Add import_sigs: &[(Name, FunctionSig)] as a new parameter to collect_mono_functions. Position: after impl_sigs, before interner (parameter order matches existing convention at mod.rs:71).
  • Build a third lookup map alongside fn_sig_by_name and impl_sig_by_name (current construction near mod.rs:75): 
    let import_sig_by_name: FxHashMap<Name, &FunctionSig> =
    import_sigs.iter().map(|(n, s)| (*n, s)).collect();
  • Extend the lookup chain in the loop body (currently mod.rs:88-92): when instance.receiver_type.is_none() AND function_sigs.get(&instance.fn_name).is_none(), try import_sig_by_name.get(&instance.fn_name) BEFORE falling through to the silent-skip branch (the tracing::debug! log — preserve as-is to support MONO-REG invariant verification).
  • DO NOT extend the instance.receiver_type.is_some() path to consult import_sig_by_name — methods are owned by §03 per Decision 02 (decisions/02-imported-method-scope-routing.md). Receiver-bearing instances continue to use only impl_sig_by_name; §03 will extend that lookup chain in its own commit when it adds multi.rs::public_methods export surface.
  • Mark imported-mono MonoFunction records via a new MonoFunction.is_imported: bool field set to true for instances resolved via import_sig_by_name. Codegen consumers (declare_mono_functions at define_phase.rs:30) MAY observe this flag for diagnostics; the body-import path (Decision 01 Option B) does NOT branch on it — both local and imported mono functions flow through identical declare_mono_functions + prepare_mono_cached plumbing.
  • DELETE the §02.1 line 81 instruction referencing declare_imported_mono (ghost function — does not exist in codebase). Decision 01 selects body-import via the existing declare_mono_functions path. (Already removed during Step-5 editor pass; ghost reference absent from current §02.1 body.)
  • Update compiler_repo/compiler/ori_llvm/src/monomorphize/tests.rs unit tests to construct import_sigs fixtures; cover the new lookup path (positive: instance resolves via import_sigs after function_sigs+impl_sigs miss; negative: instance with receiver_type.is_some() does NOT consult import_sigs even when name matches).
  • Subsection close-out (02.1) — MANDATORY before starting §02.2:
    • All tasks above [x]; cargo build --release -p ori_llvm succeeds.
    • cargo test --release -p ori_llvm --test aot generics reports same pass count as pre-§02.1 baseline (no regression — host-module mono path unchanged).
    • Run compiler_repo/diagnostics/repo-hygiene.sh --check.
    • Stabilization-discipline note (per blind-spots architectural risk): Workspace cargo check will be RED at end of §02.1 because oric consumers haven’t been updated yet — collect_mono_functions signature change breaks the three callers (evaluator/compile.rs:230, codegen_pipeline/mod.rs:138, arc_lowering.rs:272). This is the canonical multi-commit-sequencing case per impl-hygiene.md §Stabilization Discipline rule “Multi-commit sequences ordered by dependency, not chronology”. §02.2 MUST land in the same plan execution session immediately after §02.1; commit boundary between §02.1 and §02.2 is acceptable per multi-commit dependency-order rule. No --no-verify bypass; no dev-loop break.

02.2 Driver wiring — compile_common + codegen_pipeline + arc_lowering test runner + JIT path

Files:

  • compiler_repo/compiler/oric/src/commands/compile_common.rs:217-246 (import_sigs construction at :217; consumed via collect_mono_functions at :246)

  • compiler_repo/compiler/oric/src/commands/codegen_pipeline/mod.rs:79 (run_borrow_inference signature) + :143 (collect_mono_functions call) + :372 (run_borrow_inference call from caller)

  • compiler_repo/compiler/oric/src/test/runner/arc_lowering.rs:39 (lower_and_infer_borrows signature) + :272 (collect_mono_functions call)

  • compiler_repo/compiler/ori_llvm/src/evaluator/compile.rs:230 (JIT collect_mono_functions call)

  • run_borrow_inference signature extension (codegen_pipeline/mod.rs:79): add import_sigs: &[(Name, FunctionSig)] parameter (position after impl_sigs, before canon). Forward to internal collect_mono_functions call at mod.rs:143.

  • run_borrow_inference caller update (codegen_pipeline/mod.rs:372): pass import_sigs from the enclosing run_codegen_pipeline orchestrator (codegen_pipeline/mod.rs:314). The import_sigs: &[(Name, FunctionSig)] parameter is already present on run_codegen_pipeline (verified at mod.rs:314); thread it through to the run_borrow_inference call.

  • compile_common.rs driver wiring (compile_common.rs:206-223): the existing import_sigs: Vec<(Name, FunctionSig)> construction at line 206 already exists per BUG-04-119 evidence. The call to run_codegen_pipeline at compile_common.rs:235 ALREADY passes &import_sigs (verified). No edit to compile_common.rs required — wiring is in place; downstream propagation through run_codegen_pipeline → run_borrow_inference → collect_mono_functions is the work scope.

  • lower_and_infer_borrows test-runner sync (arc_lowering.rs:39): add import_sigs: &[(Name, FunctionSig)] parameter symmetric to production. Forward to internal collect_mono_functions call at arc_lowering.rs:272. Update every caller in compiler_repo/compiler/oric/src/test/runner/ to pass &[] if no imported sigs are in test scope, OR construct test fixtures with imported sigs for the new matrix tests at §02.4.

  • JIT path passthrough (evaluator/compile.rs:230): pass &[] for import_sigs. JIT compiles only the host module; cross-module imports are AOT-only by design. Document with a public-safe // JIT: single-module scope; cross-module imports are an AOT concern. line at the call site (NO bug IDs, NO wrapper plan paths in compiler source per impl-hygiene.md §STRUCTURE:public-leak).

  • Verification: cargo build --release and cargo build both succeed (workspace cargo check returns to GREEN at end of §02.2).

  • Subsection close-out (02.2) — MANDATORY before starting §02.3:

    • cargo build --release && cargo build both succeed; workspace clean.
    • cargo test --release -p ori_llvm --test aot generics reports same pass count as pre-§02 baseline (modulo the still-failing test_poly_lambda_with_imported_assert_eq_* tests — those green at §02.4).
    • Run compiler_repo/diagnostics/repo-hygiene.sh --check.

02.3 Per-module CanonResult retention + imported_mono_fns construction (Decision 01 Option B AOT promotion)

Files:

  • compiler_repo/compiler/oric/src/commands/build/multi.rs:201-219 — per-module CanonResult retention via CompiledModuleInfo (type_result + canon_result + pool fields under #[allow(dead_code)] guards).
  • compiler_repo/compiler/oric/src/commands/build/multi.rs:722 — compile_single_module invokes build_imported_mono_state, which constructs ImportedMonoState { canons, imported_mono_fns } against the merged pool.
  • compiler_repo/compiler/oric/src/commands/compile_common.rs:134 — compile_to_llvm_with_imported_monos (single-module entry); threads imported_mono_fns + re_interned_canons through to run_codegen_pipeline.
  • compiler_repo/compiler/oric/src/commands/compile_common.rs:194 — compile_to_llvm_with_imports (multi-module entry); same threading.
  • compiler_repo/compiler/oric/src/commands/mod.rs:25 — re-export site pub(crate) use codegen_pipeline::imported_mono::ImportedMonoFn; bridging the private mod codegen_pipeline; boundary.
  • compiler_repo/compiler/oric/src/commands/codegen_pipeline/imported_mono.rs — production build_imported_mono_functions AOT entry point.
  • compiler_repo/compiler/oric/src/test/runner/imported_mono.rs — thin wrapper forwarding to the production builder (post-promotion).

This subsection promotes the JIT carrier shape (oric/src/test/runner/imported_mono.rs::build_imported_mono_functions) into AOT production scope per Decision 01 Option B. Without this work, §02.1 + §02.2 wire the sig-table but arc_lowering.rs cannot specialize the imported generic body — it has no source CanonResult to feed lower_to_arc.

  • Promote ImportedMonoFn carrier type to a shared production location: moved the pub(crate) type ImportedMonoFn = (ori_llvm::monomorphize::MonoFunction, usize, ori_ir::Name); definition into compiler_repo/compiler/oric/src/commands/codegen_pipeline/imported_mono.rs (NEW file) as a pub(crate) item. Re-export wired at commands/mod.rs:29-31 as pub(crate) use codegen_pipeline::imported_mono::{build_imported_mono_functions as build_imported_mono_functions_for_test_runner, ImportedMonoFn}; (narrow approach — mod codegen_pipeline; stays private; re-export bridges the boundary). Test-runner site at oric/src/test/runner/imported_mono.rs now consumes the production type via pub(super) use crate::commands::ImportedMonoFn; re-export.
  • Promote build_imported_mono_functions to AOT-callable in the new codegen_pipeline/imported_mono.rs. JIT signature mirrored verbatim per oric/src/test/runner/imported_mono.rs:28-39. Implementation copies the JIT body (no behavioral divergence) and shares the build_body_type_map + build_concrete_sig helpers. Test-runner version is now a thin wrapper that forwards to crate::commands::build_imported_mono_functions_for_test_runner per impl-hygiene.md §Algorithmic DRY (eliminates the second copy).
  • multi.rs::CompiledModuleInfo per-module CanonResult retention: extended CompiledModuleInfo (compiler_repo/compiler/oric/src/commands/build/multi.rs:201-237) with three new fields — type_result: TypeCheckResult, canon_result: SharedCanonResult, pool: Arc<Pool> — guarded with #[allow(dead_code, reason = "consumer arrives at §02.4 merged-pool re-interning")]. compile_single_module populates them after the codegen call. The actual merged-pool re-interning consumer arrives at §02.4 matrix-test wiring.
  • compile_common.rs imported_mono_fns construction: extended compile_to_llvm_with_imports (compile_common.rs:184) with imported_mono_fns: &[ImportedMonoFn] parameter. multi.rs::compile_single_module (multi.rs:309) currently builds an empty Vec<ImportedMonoFn> and threads through; populated construction (calling build_imported_mono_functions with imported_generic_sigs derived from retained per-module TypeCheckResult via merged-pool re-interning) arrives at §02.4 alongside the matrix-test wiring.
  • run_borrow_inference second signature extension (codegen_pipeline/mod.rs:79): added imported_mono_fns: &[crate::commands::ImportedMonoFn] parameter (position after import_sigs). Forwarded from run_codegen_pipeline (codegen_pipeline/mod.rs:314). Internal threading: arc_lowering specialization loop added at lines :188-210 (mirrors the test-runner shape at arc_lowering.rs:108-117). Returns mono_functions merged with imported monos so codegen’s declare_mono_functions sees both.
  • lower_and_infer_borrows second signature extension (arc_lowering.rs:39): NO CHANGE NEEDED — test-runner already accepts this parameter at arc_lowering.rs:50 (JIT). Production signature now matches test-runner shape; both consume the same ImportedMonoFn carrier.
  • arc_lowering body specialization: promoted the test-runner specialization loop body into the production run_borrow_inference path at codegen_pipeline/mod.rs:188-210. Each ImportedMonoFn triple feeds lower_to_arc(mono_fn.mangled_name, &mono_fn.sig, *source_body_name, canon, ...) against the host’s CanonResult (current — merged-pool re-interning arrives at §02.4 for full multi-module support).
  • JIT path passthrough (evaluator/compile.rs:230): no change needed — JIT already passes &[] for import_sigs and the new pipeline signature parameter is added in compile_common.rs / run_codegen_pipeline only. The JIT site uses collect_mono_functions directly with its own pre-built imported_mono_functions and bypasses run_codegen_pipeline.
  • Verification: cargo build --release && cargo build both succeed; existing JIT tests continue to pass (5 new unit tests cover the production carrier shape + builder behavior).
  • Subsection close-out (02.3) — MANDATORY before starting §02.4:
    • All tasks above [x]; both build profiles green.
    • cargo test --release -p ori_llvm --test aot generics (94 passed, 11 failed, 4 ignored — same pass count as pre-§02 baseline per §02.2 close-out at line 143; 11 failures are pre-existing parallel-session baseline) and cargo test --release -p oric --lib (625 passed, 0 failed — green) both report same pass count as pre-§02 baseline. AIMS snapshot test (aims_snapshots_across_all_passes_match_baselines) is failing due to parallel-session ori_arc/aims-burden-tracking work — unrelated to §02.3 (my changes are no-ops with empty imported_mono_fns slice; cannot affect realize_rc_reuse snapshot).
    • Verify Decision 01 acceptance gate items (decision file decisions/01-imported-mono-linkage-mechanism.md); confirm decision status: established matches the implementation (frontmatter already at established since decision-authoring time).
    • Run compiler_repo/diagnostics/repo-hygiene.sh --check — clean.

02.4 Matrix tests — imported generic top-level functions

File: compiler_repo/compiler/ori_llvm/tests/aot/poly_lambda_mono.rs

Matrix dimensions per tests.md §Matrix Testing Rule:

Receiver shapeElement typeTest nameStatus
top-level fninttest_poly_lambda_with_imported_assert_eq_intcurrently failing — greens at §02.4
top-level fnstrtest_poly_lambda_with_imported_assert_eq_strcurrently failing — greens at §02.4
top-level fn[int]test_imported_generic_fn_list_intnew
top-level fnstructtest_imported_generic_fn_structnew

Method matrix (Option<int>.unwrap() / Result<int,str>.ok() / [int].first()) is owned by §03 per Decision 02 — not authored here.

  • Author the two new tests test_imported_generic_fn_list_int and test_imported_generic_fn_struct under compiler_repo/compiler/ori_llvm/tests/aot/poly_lambda_mono.rs. Each test imports a generic top-level function from a sibling test fixture (compiler_repo/compiler/ori_llvm/tests/aot/fixtures/imported_generics/) and exercises mono dispatch in a host-module caller.
  • Each test pins both Ok behavior (correct result) AND a semantic pin (would fail if §02.1’s import_sig_by_name lookup branch is reverted — assertion catches E5001 unresolved function).
  • Negative pin: test_unimported_generic_still_fails_cleanly — generic from a module NOT in import_sigs produces a clear E5001 (not a silent crash). Asserts the failure mode is diagnostic, not corruption. PASSES — confirms boundary semantics.
  • Remove #[ignore] annotations from test_poly_lambda_with_imported_assert_eq_int + _str — annotations removed so the tests now run alongside the new matrix tests.
  • Wire merged-pool re-interning consumer in multi.rs::compile_single_module:
    • New ImportedMonoState struct + build_imported_mono_state helper (multi.rs) — clones host pool, Salsa-queries each resolved_imports.modules[i] for type-check + canon (covers stdlib AND relative imports uniformly), re-interns each module’s canon arena + generic sigs into the merged pool with shared per-module cache + var_remap, builds imported_mono_fns: Vec<ImportedMonoFn> via the SSOT build_imported_mono_functions_for_test_runner entry point promoted in §02.3.
    • Thread re_interned_canons: &[CanonResult] through compile_to_llvm_with_imports → run_codegen_pipeline → run_borrow_inference (compile_common.rs + codegen_pipeline/mod.rs signatures); single-file path passes &[].
    • run_borrow_inference’s imported_mono lowering loop consumes re_interned_canons[*source_module_idx] (with unwrap_or(canon) fallback for legacy callers passing &[]) so arc_lowering::lower_to_arc specializes the generic body against the SOURCE module’s canon in merged-pool coordinates.
  • Subsection close-out (02.4):
    • All 4 matrix tests + 1 negative pin pass — 5/5 pass at commit 3bb8aa725 (cargo test --release -p ori_llvm --test aot poly_lambda_mono returns 5 passed; 0 failed; 0 ignored). Architectural blocker resolved per 2026-05-15 HISTORY entry “§02.4 architectural blocker resolved” below.
    • timeout 150 ./test-all.sh green — full regression check per CLAUDE.md §Commands. (deferred-with-anchor: §02.N completion checklist line “timeout 150 compiler_repo/test-all.sh green — full regression check per CLAUDE.md §Commands (success criterion 13)”) — scoped verification at §02.4-commit-time covered §02’s actual scope (cargo test --release -p ori_llvm --test aot poly_lambda_mono 5/5; cargo test --release -p ori_llvm --test aot generics 94/11/4 matching §02.3 baseline); broad-suite re-run lives at §02 close-out per the §02.N anchor above, gated on parallel-session work (plans/aims-burden-tracking + plans/typeck-inference-completeness) clearing the test baseline they currently own.
    • timeout 150 compiler_repo/diagnostics/dual-exec-verify.sh tests/aot/poly_lambda_mono.rs — N/A by design. The §02.4 matrix tests are Rust-harness #[test] functions exercising the AOT path; the JIT path documented in evaluator/compile.rs:230 passes &[] for import_sigs per success-criterion line 14 (cross-module imports are AOT-scope-only; single-module JIT does not see imported generics). dual-exec-verify.sh returns exit-3 zero-verifications by design vs by tooling defect. Verified via inspection 2026-05-15 commit 3bb8aa725.
    • Run compiler_repo/diagnostics/repo-hygiene.sh --check — clean.
    • Cross-reference §06 test matrix totals: imported-generic matrix lives in §02.4 (4 positive + 1 negative consolidated here), not §06 (per TPR-02-R5-003 deferral anchor). Verified: section-06-test-matrix.md:101 enumeration credits “§02.4 (4 + 1 negative)” and :79 lists §02 ∩ §05 cross-facet test consistent with this attribution; no §06 edit required.

02.R — Third Party Review Findings

Populated by /tpr-review invocations. Findings land here as - [ ] items with Resolved: lines on closure.

  • [TPR-02-R5-001-codex][High] GAP — Imported-signature lookup chain at §02.1 may be incomplete: the import_sig_by_name keying uses canonical Name but JIT-side imported_mono_fns uses local-vs-aliased names; verify §02.1 implementation handles the dual-name case OR explicitly route through canonical mangling. Resolved: §02.1 — both paths use ori_ir::Name (interned ID) via shared StringInterner: monomorphize/mod.rs:94-97 keys import_sig_by_name with *name from the input slice, and compile_common.rs:209 constructs import_sigs: Vec<(Name, FunctionSig)> using the same interner that produces JIT-side imported_mono_fns: Vec<(MonoFunction, usize, Name)> at oric/src/test/runner/imported_mono.rs:22. Dual-name divergence ruled out per opencode-r3 code verification + 13/13 unit-test pass.
  • [TPR-02-R5-002-codex][Medium] DRIFT — run_borrow_inference mono-lowering loop placement: codex notes mod.rs:319 caller and mod.rs:138-160 production loop; verify §02.2 places imported_mono_fns specialization INSIDE run_borrow_inference (not at caller) AND mod.rs:319 forwards parameters only. Resolved: §02.2 — collect_mono_functions call lives inside run_borrow_inference at codegen_pipeline/mod.rs:143; the mod.rs:319 caller (run_codegen_pipeline) only forwards import_sigs per success_criterion line 11. Line 170 plan-prose updated in /tpr-review round 1 to clarify caller-vs-internal direction.
  • [TPR-02-R5-003-codex][Medium] DRIFT — §06 test matrix totals may need refresh once §02.4 lands (4 positive + 1 negative consolidated under §02.4 vs distributed across §06). Cross-reference update at §06 implementation. Resolved: §06 cross-reference verified accurate at commit 3bb8aa725 — section-06-test-matrix.md:101 enumeration credits §02.4 (4 + 1 negative); §06 needed no edit.

Round 5 cap-exit findings (this invocation; filed 2026-05-15, cap_reached_max_rounds)

Filed per /tpr-review §7 cap-exit accept-with-findings; resolution owned by §02.3/§02.4 implementation work. Cap-exit reason recorded in third_party_review.notes.

  • [TPR-02-R5-004-codex][High] GAP — §02.3 names compile_common.rs as the per-module CanonResult retention site, but actual cross-module retention orchestration lives at compiler_repo/compiler/oric/src/commands/build/multi.rs::CompiledModuleInfo (multi.rs:201-219). multi.rs:262-263 calls check_source per-module and drops canon_result locally; compile_common.rs:184-240 is single-module entry only. Resolved: §02.3 implementation (2026-05-15 HISTORY entry) extended CompiledModuleInfo with type_result + canon_result + pool retention fields; multi.rs::compile_single_module loop threads per-module canons through re_interned_canons: &[CanonResult] parameter on compile_to_llvm_with_imports → run_codegen_pipeline → run_borrow_inference; imported_mono lowering loop in run_borrow_inference consumes re_interned_canons[*source_module_idx] so arc_lowering::lower_to_arc specializes the generic body against the SOURCE module’s canon in merged-pool coordinates. Cured by commits in §02.3 + §02.4 work-arc (terminal commit 3bb8aa725).
  • [TPR-02-R5-005-codex][Medium] GAP — §02.3 plan to move ImportedMonoFn from oric/src/test/runner/imported_mono.rs:22 to commands/codegen_pipeline/imported_mono.rs requires visibility specification: commands/mod.rs:25 declares mod codegen_pipeline; (private). Test runner at crate::test::runner::imported_mono cannot import unless visibility raised. Resolved: §02.3 + §02.4 work-arc promoted ImportedMonoFn + build_imported_mono_functions into compiler/oric/src/commands/codegen_pipeline/imported_mono.rs (production); compiler/oric/src/commands/build/mod.rs carries pub(crate) use multi::build_imported_mono_state; re-export (the §02.4 cure path widened visibility for single-file + multi-file consumers). Test-runner site at oric/src/test/runner/imported_mono.rs reduced to thin wrapper per §02.3 HISTORY entry. Cured by commit 3bb8aa725 re-export block.
  • [TPR-02-R5-006-opencode][High] PLAN_COHERENCE_DRIFT — §02.1/§02.2 sections marked status: complete while their owning cap-exit findings (TPR-02-R5-001/002) were - [ ] until this invocation’s cure. Resolved: TPR-02-R5-001/002 flipped [x] above with code-verified Resolved annotations citing monomorphize/mod.rs:94-97 (shared StringInterner) + codegen_pipeline/mod.rs:143 (inside run_borrow_inference).
  • [TPR-02-R5-007-opencode][Medium] PLAN_COHERENCE_DRIFT — TPR-02-R5-003 deferral lacked concrete anchor per CLAUDE.md §ALL Deferrals Must Have Implementation Anchors. Resolved: added - [ ] cross-reference anchor under §02.4 close-out checklist; TPR-02-R5-003 Resolved field updated to cite that anchor.

Round 6 cap-exit findings (this invocation; filed 2026-05-16, cap_reached_max_rounds)

Filed per /tpr-review §7 cap-exit accept-with-findings; cap-exit reason recorded in third_party_review.notes. Per-round summary (5 rounds, parallel-session line-ref drift surfacing across rounds):

  • Round 1: 2 verified findings cured inline (stale mod.rs:59 line refs at success_criterion line 11 + 13-site §02.X / TPR-02-R5-005 source-comment sweep across poly_lambda_mono.rs / multi.rs / codegen_pipeline/imported_mono/tests.rs / codegen_pipeline/mod.rs).

  • Round 2: 4 meta findings cured inline (mod.rs:77/138/319 → :79/143/372 line refs + retired :stale-plan-annotation → :non-spec-pointer category + subsection enumeration 02.R addition + reviewer_set frontmatter field).

  • Round 3: 2 findings cured inline + 7 adjacent litter-pickup (codex-Critical LEAK:scattered-knowledge methodology vocab strip at codegen_pipeline/imported_mono.rs:41 + Major plan-ref strip at test/runner/imported_mono.rs:6; litter-pickup swept 7 more refs across poly_lambda_mono.rs:36 / compile_common.rs:186 / codegen_pipeline/mod.rs:174 + :276 / multi.rs:222 + :372 + :417 + :640 + :843).

  • Round 4: 2 findings cured inline (mod.rs line refs again — :143-165 → :188-210 for imported-mono loop; success_criterion line 12 prose rewrite to attribute imported_mono_fns construction to multi.rs::build_imported_mono_state vs compile_common.rs).

  • Round 5: 3 verified findings filed below + 1 dropped (consistency with round-1 disposition); 2 gemini plan-checkbox-bypass attempts dropped across rounds 2+4.

  • [TPR-02-2026-05-16-001-codex][Minor] PLAN_COHERENCE_DRIFT — §02.3 Files block at section.md:152-153 carries pre-implementation prose that no longer matches HEAD: line 152 “currently drops canon_result at multi.rs:330-336” is stale (canon_result is retained via CompiledModuleInfo at multi.rs:203-238 with #[allow(dead_code)] guards); line 153 describes compile_common.rs as constructing imported_mono_fns when actual construction lives in multi.rs::build_imported_mono_state (multi.rs:722) per round-4 cure to success_criterion line 12. Cosmetic prose drift; resolution: refresh §02.3 Files block to match HEAD coordinates + delete obsolete pre-implementation claims. Resolved (Round 1 2026-05-16): §02.3 Files block rewritten at section-02-import-sigs-traversal.md:151-157 to reflect HEAD shape — multi.rs:201-219 retains canon_result via CompiledModuleInfo; multi.rs:722 invokes build_imported_mono_state (constructs ImportedMonoState); compile_common.rs:134 (compile_to_llvm_with_imported_monos) + compile_common.rs:194 (compile_to_llvm_with_imports) thread imported_mono_fns + re_interned_canons to run_codegen_pipeline. Pre-implementation prose deleted.

  • [TPR-02-2026-05-16-002-gemini][Minor] PLAN_COHERENCE_DRIFT:citation-drift — §02.2 body line refs at section.md:132-138 carry stale mod.rs numbers (mod.rs:77 should be :79; mod.rs:138 should be :143; mod.rs:319 should be :372; mod.rs:263 should be :314) while success_criterion line 11 already carries the cured refs per round-4 fix. Body/criterion divergence. Resolution: refresh §02.2 Files block + task body line refs to current HEAD coordinates matching success_criterion line 11; consider adding a “verified at recon-time HEAD” annotation or migrating to symbol anchors (per round-1 codex broader cure note) to prevent future drift recurrence. Resolved (Round 1 2026-05-16): §02.2 Files block at section-02-import-sigs-traversal.md:132 + §02.2 task body at lines 136-137 + §02.3 task bodies at lines 164 + 166 refreshed to HEAD coordinates :79 / :143 / :372 / :314 / :188-210 matching success_criterion line 11.

  • [TPR-02-2026-05-16-003-opencode][Medium] PLAN_COHERENCE_DRIFT — 00-overview.md mission criterion at line 25 - [ ] cargo test --release -p ori_llvm --test aot poly_lambda_mono ... passes both test_poly_lambda_with_imported_assert_eq_int and test_poly_lambda_with_imported_assert_eq_str (delivered by §02.4 matrix tests subsection). <!-- delivered-by: §02.4 --> is unchecked but the deliverable is verifiably complete: §02.4 status: complete per frontmatter; close-out reports 5/5 pass at commit 3bb8aa725 via cargo test --release -p ori_llvm --test aot poly_lambda_mono returning 5 passed; 0 failed; 0 ignored; <!-- delivered-by: §02.4 --> annotation present. Plan-checkbox-bypass does NOT apply (cargo test command, not slash-command). Resolution: flip [ ] → [x] on 00-overview.md:25 at §02.N close-out via mechanical auto-fix. Resolved (Round 1 2026-05-16): 00-overview.md:25 flipped [ ] → [x] per feedback_auto_fix_cleanup.md mechanical auto-fix; HTML comment extended to cite verified 5/5 pass at commit 3bb8aa725.

Round 7 cap-exit findings (this invocation; filed 2026-05-16, cap_reached_max_rounds)

Filed per /tpr-review §7 cap-exit accept-with-findings; cap-exit reason recorded in third_party_review.notes. Per-round summary (5 rounds, survivor-mode 2-of-3 on round 5 — codex transport-failed Tier 4.5):

  • Round 1: 3 findings cured inline (§02.2 + §02.3 Files-block + line-refs refresh to HEAD coords; 00-overview.md:25 mission criterion flipped [x]).

  • Round 2: 3 findings cured inline (compiler-source plan-section refs stripped from monomorphize/tests.rs:202+240 + poly_lambda_mono.rs:72 closed-BUG-04-042 ref + litter-pickup at tests.rs:271).

  • Round 3: 4 actionable findings cured inline (closed BUG-01-002 §C.2/§C.4 refs stripped from monomorphize/tests.rs:275-282; recon-block coords at section.md:101 refreshed; BUG-04-119 annotated in bug-tracker/section-04-codegen-llvm.md:120 with <!-- delivered-by: -->; section.md:252 plan_corpus check checkbox flipped [x]; §02.N “BUG-04-119 tracker entry annotated” flipped [x]; generics::* success-criterion wording softened to cite parallel-session baseline carve-out) + 2 false-positives dropped (semantic-pin preserved vocab; section-not-complete circular) + 2 meta accepted.

  • Round 4: 2 actionable findings cured inline (compile_common.rs:179 “Section 12.14” stripped; monomorphize/tests.rs:315 “§C.4.1” stripped) + 2 meta improvements applied (test-all.sh blocked-by annotation + plan-annotations-sweep grep command; sweep verified clean across §02-owned code surfaces, §02.N checkbox flipped [x]).

  • Round 5: 2 actionable findings cured inline (line-ref drift at section.md:205 + status-summary prose at section.md:70) + 1 meta dropped (circular /tpr-review checkbox flip recommendation) + codex transport-failed Tier 4.5 (survivor mode).

  • [TPR-02-2026-05-16-004-gemini][Minor] PLAN_COHERENCE_DRIFT:citation-drift — §02.4 close-out prose at section.md:205 cites “success-criterion line 18-19” for the JIT &[] passthrough explanation, but the correct ref is line 14 (lines 18-19 are the no-regression + matrix criteria, not the JIT passthrough criterion). Cosmetic citation drift; resolution: update inline citation to line 14. Resolved (Round 5 2026-05-16): section.md:205 citation updated from “line 18-19” → “line 14” matching the JIT passthrough success-criterion at frontmatter line 14.

  • [TPR-02-2026-05-16-005-opencode][Minor] PLAN_COHERENCE_DRIFT — §02 body status prose at section.md:70 lists plan_corpus check as a pending §02.N item, but §02.N line 252 marks it [x] verified clean 2026-05-16 (Round 3 cure). Status-summary string stale relative to §02.N close-out checklist. Resolution: drop plan_corpus check from the prose enumeration; refresh to reflect 2 remaining gates only (/tpr-review + /impl-hygiene-review). Resolved (Round 5 2026-05-16): section.md:70 rewritten to enumerate the 2 remaining unchecked gates + cite the 3 flipped items (plan_corpus check + BUG-04-119 tracker annotated + plan annotations swept) with the test-all.sh gating noted.

02.N Completion Checklist

  • All 02.1, 02.2, 02.3, 02.4, 02.R subsections status: complete.
  • Decision 01 (decisions/01-imported-mono-linkage-mechanism.md) flipped to status: established; acceptance gate items verified.
  • Decision 02 (decisions/02-imported-method-scope-routing.md) flipped to status: established; §03 amendment co-landed (§03 frontmatter + §03.4 + §03.5 subsections authored — this is §03’s own commit, NOT §02’s, but §02 close-out verifies the cross-section commitment landed).
  • /tpr-review on §02 diff returns clean across active reviewer set (codex + gemini + opencode when all present; documented survivor-mode subset per .claude/skills/tpr-pipeline/failure-matrix.md §Survivor-mode matrix + .claude/skills/tpr-pipeline/reviewer-set-composition.md runtime detection when one drops).
  • /impl-hygiene-review on §02 diff returns clean after TPR.
  • python -m scripts.plan_corpus check plans/aot-mono-completeness/section-02-import-sigs-traversal.md returns exit 0. .
  • timeout 150 compiler_repo/test-all.sh green — full regression check per CLAUDE.md §Commands (success criterion 13).
  • BUG-04-119 tracker entry annotated <!-- delivered-by: plans/aot-mono-completeness/section-02-import-sigs-traversal.md --> (full retirement happens at §07).
  • §02 plan annotations swept — no stale TPR-, CROSS-, Phase A, BUG-04-119 references remain in code touched by §02 per impl-hygiene.md COMMENT_HYGIENE_DRIFT:non-spec-pointer. Verify clean via: rg -n '§0[2-9]|Decision 0[1-9]|TPR-0[2-9]-|CROSS-0[2-9]-|Phase [A-Z]|BUG-04-119' compiler_repo/compiler/ori_llvm/src/monomorphize compiler_repo/compiler/ori_llvm/tests/aot/poly_lambda_mono.rs compiler_repo/compiler/oric/src/commands/build/multi.rs compiler_repo/compiler/oric/src/commands/codegen_pipeline compiler_repo/compiler/oric/src/commands/compile_common.rs compiler_repo/compiler/oric/src/test/runner/imported_mono.rs compiler_repo/compiler/oric/src/test/runner/arc_lowering.rs returns zero matches.
  • Section frontmatter flipped: status: complete, reviewed: true via flip_from_in_review_clean() per state-discipline.md §4.

HISTORY

  • 2026-05-15 — §02.3 implementation complete: promoted ImportedMonoFn carrier + build_imported_mono_functions from JIT test-runner into production at compiler_repo/compiler/oric/src/commands/codegen_pipeline/imported_mono.rs; extended CompiledModuleInfo with type_result + canon_result + pool retention (consumer arrives at §02.4 merged-pool re-interning); threaded imported_mono_fns: &[ImportedMonoFn] through compile_to_llvm_with_imports → run_codegen_pipeline → run_borrow_inference; added arc_lowering specialization loop in run_borrow_inference mirroring test-runner shape at arc_lowering.rs:108-117; test-runner site at oric/src/test/runner/imported_mono.rs reduced to thin wrapper. 5 new unit tests pass; both build profiles green; generics AOT pass count matches pre-§02 baseline (94 passed / 11 failed / 4 ignored — pre-existing parallel-session baseline failures). AIMS snapshot test failure independent of §02.3 (no-op changes when imported_mono_fns is empty).

  • 2026-05-15 — §02.4 architectural blocker resolved (commit 3bb8aa725): two distinct failure modes from the prior §02.4 entry were cured by compiler/oric/src/commands/build/multi.rs::extract_public_function_types adding if func_sig.is_generic() { continue; } filter (generics reach codegen via imported-mono dispatch path; raw import_sigs MUST NOT carry source-pool BoundVar leaves) AND single-file build wiring compile_to_llvm_with_imported_monos (new entry-point in compile_common.rs) so use std.testing { assert_eq } / use std.<generic> imports get monomorphized via build_imported_mono_state against the merged pool. Mirrors the JIT path’s filter at oric/src/test/runner/llvm_backend.rs:213 — production AOT now consistent with the JIT reference. Files modified: compiler/oric/src/commands/build/multi.rs (generic-skip filter; ImportedMonoState visibility widened to pub(crate); _compiled_modules parameter removed); compiler/oric/src/commands/build/mod.rs (pub(crate) use multi::build_imported_mono_state; re-export); compiler/oric/src/commands/compile_common.rs (compile_to_llvm_with_imported_monos replacing dead compile_to_llvm); compiler/oric/src/commands/build/single.rs (call-site rewiring); compiler/oric/src/commands/run/mod.rs (call-site rewiring). Verification: cargo test --release -p ori_llvm --test aot poly_lambda_mono → 5/5 pass; cargo test --release -p ori_llvm --test aot generics → 94 passed / 11 failed / 4 ignored (matches §02.3 baseline — failures are pre-existing parallel-session AIMS-burden-tracking entries, not §02 regressions). Observation surfaced + not filed: pre-fix Tag::BoundVar leak was layout-sensitive (6-line fixture-header difference shifted pool-index allocation enough to mask the bug for some shapes); cure (filter generics from import_sigs) eliminates the layout dependency.

  • 2026-05-15 — §02.4 partial implementation + architectural blocker surfaced:

    • Implementation landed: 4 matrix tests + 1 negative pin authored in compiler_repo/compiler/ori_llvm/tests/aot/poly_lambda_mono.rs; fixtures at compiler_repo/compiler/ori_llvm/tests/aot/fixtures/imported_generics/{generics,host_list_int,host_struct}.ori; #[ignore] removed from test_poly_lambda_with_imported_assert_eq_int + _str. Merged-pool re-interning consumer wired in multi.rs::compile_single_module via new ImportedMonoState + build_imported_mono_state helper (uses oric::query::type_check_module + canonicalize_cached_by_path Salsa queries — covers stdlib AND relative imports uniformly without depending on compiled_modules). re_interned_canons plumbed through compile_to_llvm_with_imports → run_codegen_pipeline → run_borrow_inference; imported_mono lowering loop in run_borrow_inference now specializes against re_interned_canons[source_module_idx] (with unwrap_or(canon) fallback). Build green; negative pin (test_unimported_generic_still_fails_cleanly) passes — confirms boundary semantics.
    • Architectural blocker: 4 positive tests still fail with two distinct failure modes:
      1. assert_eq (stdlib): unresolved function 'assert_eq' in apply/invoke — missing mono instance? — mono_dispatch_by_id[host_mono_instance_id] lookup returns None or mono_dispatch[Name("assert_eq")] fallback fails arg-type matching despite imported_mono_fns carrying the correct (MonoFunction, source_module_idx, original_name) triple with populated instance_ids. The merged path produces 3 ImportedMonoFn entries (identity, assert_eq, first) for host_list_int.ori; imported_generic_sigs map correctly keyed by local name; build_imported_mono_functions succeeds; run_borrow_inference merges imported monos into all_mono_functions. Suggests declare_mono_functions ↔ lookup_mono_dispatch chain has a gap for imported-mono entries.
      2. identity/pair (relative-path helper): Tag::BoundVar reached codegen — upstream substitution missed a leaf at Idx(224|225) from compiler_repo/compiler/ori_llvm/src/codegen/type_info/store.rs:388. The body_type_map (built by build_mono_body_type_map iterating merged_pool entries with HAS_VAR | HAS_BOUND_VAR) appears to miss one leaf in the re-interned imported-generic body. Likely root: re_intern_type_with_var_remap re-interns body BoundVar leaves with fresh merged-pool var_ids via shared per_module_var_remaps[i], but those var_ids may not flow into var_subst consistently if extend_var_subst_with_roots (called in build_body_type_map) doesn’t traverse union-find roots correctly across pool boundaries.
    • Resolution path — requires architectural amendment beyond §02.4 scope:
      • Audit declare_mono_functions flow for imported-mono entries: verify mono_dispatch_by_id[instance_id] is populated for every mono_fn.instance_ids entry from imported_mono_fns, AND verify the host’s CanonResult.mono_dispatch_map_can carries entries that map host’s per-call-site ExprId → MonoInstanceId for imported call sites (vs only host-local generic calls).
      • Audit build_mono_body_type_map ↔ re-intern interaction: verify that after per_module_var_remaps populates fresh merged-pool var_ids, extend_var_subst_with_roots correctly threads union-find roots so EVERY Tag::BoundVar(merged_var_id) leaf in the re-interned body has a corresponding var_subst[merged_var_id] → concrete_idx entry.
      • Cross-validate against the JIT test runner: oric/src/test/runner/llvm_backend.rs consumes the SAME imported_mono_fns shape and works for the corresponding spec test at tests/spec/expressions/poly_lambda_with_imported_generic.ori. Diff the JIT path’s data flow vs the AOT path’s to identify what the JIT does that the AOT doesn’t.
    • Files modified (§02.4): compiler_repo/compiler/ori_llvm/tests/aot/poly_lambda_mono.rs, compiler_repo/compiler/ori_llvm/tests/aot/fixtures/imported_generics/{generics,host_list_int,host_struct}.ori (NEW), compiler_repo/compiler/oric/src/commands/build/multi.rs (added ImportedMonoState + build_imported_mono_state + threading), compiler_repo/compiler/oric/src/commands/compile_common.rs (added re_interned_canons parameter on compile_to_llvm_with_imports), compiler_repo/compiler/oric/src/commands/codegen_pipeline/mod.rs (run_codegen_pipeline + run_borrow_inference accept re_interned_canons; imported_mono lowering loop uses re_interned_canons[source_module_idx]). Build green; §02.3 dead-code allows on CompiledModuleInfo.{type_result,canon_result,pool} retained for future cleanup once the §02.4 consumer pattern matures.

  • 2026-05-15 — /commit-push halt skipped (autopilot continuation): halt_reason=test_all_fail; failing_repo=compiler_repo (step 5 pre-commit ./test-all.sh); scope=cross-scope (Phase A bundle carried parallel-session work — aims-burden-tracking §03.4 cycle 50b burden_dec-before-settag emission tests + lower, ori_types closure_unify, ori_llvm arc_emitter dispatch refactor, ori_repr range/transfer touch — none owned by aot-mono-completeness §02). Autopilot continues per skill-control-contract.md §Autopilot Mode unified hook-failure clause; orchestrator re-invoked with --bypass-gate dirty_tree; parallel sessions retain ownership of clearing their working-tree state via user-typed /commit-push --bypass.