tier-filter-local-ports

name: tier-filter-local-ports description: >- Route only LOCAL-tier function classes to the local model for JS→Rust porting: no JSON usage, no regex literals, no UPPER_CASE_CONST references (Tier-A-plus). Measured: K_broad 0.263→0.500 (+90%), savings 7%→47% on filtered subset. Skip pure-string and codebase-const-heavy functions. Always pass --std-only. Extend the parity battery for complex return types (f64/Vec/HashMap/Option/Result): 26/56 compile-only traces flipped to gold (46.4% flip rate). status: proto workflow: ./workflow.md

Routing everything through the local model (baseline) wastes CPU on functions that fail 95% of the time. The Tier-A-plus filter (no JSON, no regex, no UPPER_CONST) selects ~37.7% of the pool that the local model handles well. On that subset: K_broad = 0.500 (up from 0.263 baseline), savings = 47% vs frontier. The remaining 62.3% goes directly to the supervisor.

Extended parity battery converts compile-only → gold by handling f64/Vec/HashMap/ Option/Result/tuple return types. Parity-K on tier-A-plus: 0.05 → 0.375 (7.5× lift). Real bulk-port run: 38 functions, $0.51 frontier tokens avoided, ~5 min CPU.

Workflow: ./workflow.md

Provenance

Candidate: session W74 research, phase 374
Date: 2026-05-30
Evidence files:
- .planning/research/synthetic-traces/TIER-FILTER-RESULTS.md (K_broad measurements)
- .planning/research/synthetic-traces/PARITY-BATTERY-EXTENDED.md (parity flip data)
- .planning/research/integration-port/REAL-BULK-PORT.md (real 38-fn run, $0.51 saved)
See: PROVENANCE.md