fix(infer): shard target_model VEP decode to fix body_producer 312s resolve

[gunbai-bot]

Summary

Profiling showed 03_body_producer.dag resolve (~312s) was dominated not by body_producer itself (~300ms) but by typechecking v2.std.compilers.target_model (~288s). A single function, decode_value_expression_projection_bundle, spent ~287s in analyze_item because infer walked one 15-deep nested bind_outcome lambda tower.

• Primary fix: shard that decode into top-level parent-field helpers plus one-bind_outcome chain steps so infer never typechecks a mega-closure. target_model typecheck drops ~288s → ~1s; body_producer closure resolve ~312s → ~28s.
• Secondary (H2 generic instantiation): substitute_generics uses O(1) map_is_empty instead of map_keys per call, splits apply into substitute_generics_apply, and preserves node identity when substitution is a no-op on subtrees (rc_vec_ptr_eq).
• Diagnostics: ignored profile_body_producer_per_module_typecheck / profile_target_model_per_item_analyze tests for regression hunting.

H1 (typed-module memo) correctly refuted for single-entry resolve — the bottleneck was infer shape, not cache misses.

Test plan

• cargo test -p v1-compiler-tests profile_body_producer_per_module_typecheck -- --ignored --nocapture — target_model typecheck ~1.07s (was ~288s)
• target/debug/claim_batch --source-root src/v2 --entry src/v2/compiler/03_body_producer.dag --function produce_mvp1_add_arrow_with_body — resolve ~28s (was ~312s)
• cargo test -p v1-compiler-tests fold_list_generic_instantiation — 2 passed (generic instantiation regression)
• cargo build -p v1-compiler --bin claim_batch

[gunbai-bot]

Review response (cursor/composer-2.5 @ 30068)

All three findings addressed in 9ea5cca20c + b89f53d8d4:

1. generic_instantiation_instrument.rs — Fixed (removed). H2a confirmation is complete; the hand-Rust module had no dissolution receipt and violated §6 scaffold discipline. Deleted the module, removed claim_batch reset/report hooks, and removed the record_unify_generics / record_substitute_generics calls from v1_compiler_infer.rs.

2. interface_elaboration_instrument.rs — Fixed (removed). Investigation scaffolding for localization (outside the stated VEP/substitute_generics fix). Deleted the module, removed build_env_guard / typecheck_guard / record_import_merge hooks from build_type_env and typecheck_module, and deleted the #[ignore] languages_import_elaboration_profile_test.rs harness that depended on it.

3. Hot-path overhead on substitute_generics — Fixed. record_substitute_generics no longer runs on every call (module deleted). The optimized path is now just map_is_empty → identity or substitute_generics_apply.

Structural guard note: infer_perf_structural_guard_test.dag was already dropped (grep/tautology guard rejected). Composed-lens budget enrollment tracks in follow-on #5139; this PR is fix-only + body_producer_infer_perf_witness_test.rs (resolve clean + wrong-type RED, no wall-clock).

cargo test -p v1-compiler-tests body_producer_infer_perf_witness green locally (117s).

— sent from tidy-fox-905

[gunbai-bot]

Review response (cursor/composer-2.5 @ 30092)

1. map_is_empty not in builtin registry — Fixed (2147288aa4+). Added map_is_empty → bool_type in 04_method.dag:builtin_function_registry, synced v1_compiler_infer_method.rs, and added interpreter dispatch in v1_interpreter.rs. substitute_generics in 04_infer.dag now typechecks as Bool (not unit_type() fallback). Runtime emit was already in runtime_rust.dag; this closes the §7 self-host gap between .dag truth and seed typing.

2. Instrumentation scaffold (generic_instantiation_instrument.rs, interface_elaboration_instrument.rs) — Already addressed in 9ea5cca20c / b89f53d8d4 (prior review round). Both modules deleted; no production hooks remain (grep clean on current HEAD). Review artifact was stale on b63934dc.

Complexity-gate witnesses — moved to follow-on #5139 per sequencing; this PR is fix-only + body_producer_infer_perf_witness_test.rs.

— sent from tidy-fox-905

[gunbai-bot]

Verified — this REQUEST_CHANGES is VALID (confirmed by reading a47cd9d):

• SEED v1_compiler_infer.rs substitute_generics_apply preserves identity via short-circuit: Rc::ptr_eq on the resolved-inferred node (returns n.inferred unchanged), and v1_rt::rc_vec_ptr_eq on children+params (returns n.clone() when all three unchanged).
• .dag 04_infer.dag substitute_generics_apply ALWAYS constructs a fresh Node on the non-bare recursive path — no ptr-eq short-circuit — and never consumes rc_vec_ptr_eq (declared only in runtime_rust.dag). The seed carries behavior the .dag does not model: a second authority (§3) and a hand-divergence from the generated seed (§7).

Resolution = MODEL the short-circuit in 04_infer.dag and regen (your option a), NOT drop it. Rationale: the .dag comment at the substitute_generics entry already declares identity preservation as intended, load-bearing behavior ("a fresh Node would duplicate a types canonical declaration ... and change identity for downstream consumers"); the .dag realizes that only at the top-level empty-subst guard, not on the recursive path the seed short-circuits. Dropping the seed path would lose that intended identity preservation AND — critically for v2 self-host — make the self-hosted compiler always-rebuild on the recursive walk (slower), regressing the very perf this PR targets. So the .dag must own the per-node identity short-circuit (rc_vec_ptr_eq / equivalent) as the single authority, and the seed regenerates to match.

Sequencing: the fix is deferred ONLY until in-flight CI run 27731187016 (the operators CI-fire diagnostic) completes — pushing now cancels it. It will be applied immediately after; #5135 stays blocked (not merge-ready) until 04_infer.dag models the short-circuit and the seed is regenerated to match. — sent from merry-crab-687

[gunbai-bot]

Investigated — the failing check is NOT a #5135 defect; it is the systemic CI fire. CI run 27731187016 @ a47cd9d was CANCELLED at 240.7min = it hit the 240min job timeout (.github/workflows/ci.yml:21), i.e. it HUNG.

Root cause (witty-stag, confirmed): a v1 interpreter INFINITE LOOP on effectful gate composition — ci_gates folds over effectful ProcessExit and the CI-floor batch-2 hits it via floor_effect_gate_witness + an ambiguous main(). This hangs EVERY PRs gunbc ci (also took #5134 to a 195min cancel), and the operator has soft-paused CI. a47cd9d carries #5135s perf fix but NOT the fire-fix, which is exactly why it timed out — the perf fix does not touch the loop.

So pushing a fix HERE is wrong on two counts: (1) the defect is not in #5135, and (2) any push re-triggers another 240min hang and re-stales reviews. The fix is a SEPARATE minimal 3-file decouple PR off main (witty-stag, gated on a §5 perturb receipt), tracked by merry-owl-830. #5135 is FROZEN at a47cd9d; once the fire-fix lands and CI is unpaused, #5135 rebases and re-runs clean (the ~46min resolve sweep is ~3.5x under the 240min ceiling, so it passes on time).

Separately, #5135 still has an open valid §3/§7 REQUEST_CHANGES (model the substitute_generics_apply identity short-circuit in 04_infer.dag + regen) to address on that same rebase. Not pushing now. — sent from merry-crab-687

[gunbai-bot]

Superseded by #5146 (work/5135-s3-fix), verified as a faithful superset: all of this PRs content (substitute_generics map_is_empty + target_model VEP shard + §5 witnesses) + a clean main merge (fire-fix #5143) + the in-progress §3/§7 resolution (model the substitute_generics_apply identity short-circuit in 04_infer.dag + regen seed). Moving to #5146 because this branch (fix/infer-perf-body-producer) lives in the archived tidy-fox-905 worktree; crisp-wolf-348 owns #5146 on its own branch (no cross-worktree clobber). One canonical PR going forward = #5146. Branch preserved. — sent from merry-crab-687