PDK import resource contract (written before implementation)

This milestone imports standard technology/cell LEF data and one flat mapped Verilog module. No reference implementation source is consulted. There is no timing/path/cut/SAT exploration in the importer.

Inputs are limited to 64 MiB per LEF file, 16 MiB Verilog/floorplan, and 144 MiB aggregate source bytes. Each parser has explicit token/object/geometry/depth limits. Geometry uses exact decimal-to-integer conversion; off-grid values fail. Mapped-expression depth defaults to 32, leaving room for the enclosing JSON snapshot within serde_json's recursion limit. Deeper syntax is rejected explicitly. The linking budget is 250,000 nets, 1,000,000 pins/expanded bits, 500,000 alias operations, 250,000 retained geometry shapes and 250,000 instances. All indices must fit u32. Work uses bounded signal expansion and union/find, never simple-path enumeration; array lengths are checked before allocation.

Peak working estimate for the default linked-design caps: nets at 256 bytes = 64 MB; pin/bit entries, endpoint maps, CSR indices, typed arena and validation state at 512 bytes * 1 million = 512 MB; input buffers + parser state <= 384 MB; retained library geometry <= 128 MB; bounded metadata serialization, old/new snapshots, a temporary AST-to-connection reconciliation and mmap pages <= 256 MB. One corner and two workers (no per-worker library/database replicas) give a conservative 1.5 GiB runtime allowance. The existing 16 MiB published metadata limit can reject a large linked design earlier. The first dual-PDK run has <1,000 connected pins and <10 MiB combined LEF data, with an expected <512 MiB runtime. Immutable parsed technology data is shared and only used masters are retained, not a deep library copy per instance.

New connection tables use fixed-size little-endian records; counts and byte lengths are checked before mmap. Validation and CSR lowering are O(nets + pins) with explicit caps. Union/find with size balancing and path compression is amortized O((bits + aliases) alpha(bits)); alias operations are capped separately. Cycles of aliases do not recurse. Driver/coverage checks use counts, not enumerated paths.

Five-crate builds use CARGO_BUILD_JOBS=2, RUST_TEST_THREADS=2, RAYON_NUM_THREADS=2, an 8 GiB virtual-address limit, /usr/bin/time -v and timeout 180s. Inspect free -h before builds and dual-PDK suites. External engines retain their existing 2 GiB address-space and 30-second command budgets. OpenROAD's allocator requires an 8 GiB virtual-address allowance, but stays inside a 2 GiB/no-swap container memory limit; this is not a larger physical-RAM allowance. Run the two PDKs serially. Record end-to-end time including parsing, linking, publication/mmap, export and oracles; do not report parser-only time as complete-flow performance. A suite/process above the constitution's RSS limit is a defect, not a reason to lift the limit.