[DonaldLucy]

What this adds

Base gLayout builds every transistor on its own: fingers of one transistor share diffusion, but two transistors never
share a diffusion node or a gate strap. This pull request adds such sharing to gLayout as an explicit, checkable plan,
and a benchmark that measures what sharing gains and what it costs.

• src/glayout/sharing/ (new package): SharingPlan (finger order, orientations, diffusion joins, gate straps),
  validate_plan (structured rejection before any geometry is drawn), build_mos_group (draws a group for any
  MappedPDK; the sharing that really happened is measured again on the drawn shapes), and a six-axis enumeration with
  named modes.
• benchmarks/textbook_sharing/: 18 textbook topologies x SKY130, GF180MCU, IHP SG13G2 x 12 size points; row and
  channel assembly; verification with the official DRC decks, LVS, an extracted-size check, R+C extraction and
  post-layout simulation against frozen specifications; a plan-selection study with a plan-level cost rule; language-model
  sessions through a restricted plan interface; every analysis and the result tables.
• README.md: a four-line pointer. No existing gLayout code is changed.

Where to start

The 639 case folders and the run records are packed into two archives (data/cases.tar.xz,
results/run_records.tar.xz); unpack both first (README, quick start). The code is in src/glayout/sharing/ and
benchmarks/textbook_sharing/textbook_sharing/.

1. benchmarks/textbook_sharing/README.md: the idea in one example (a five-transistor OTA), quick start.
2. benchmarks/textbook_sharing/docs/DESIGN.md: how a plan is checked and drawn, the enumeration, the row assembler,
   where area goes, the cost rule, the selection study.
3. Code: src/glayout/sharing/group.py (validate_plan, build_mos_group), evidence.py, enumeration.py;
   benchmark layout.py, assembly/engine.py, verify/, selection/cost.py, llm/plan_parser.py.
4. docs/PARAMETER_SPACE.md (process x size x plan, counts), docs/VERIFICATION.md, docs/DATA.md (every file and
   field), docs/MODELS.md (model protocol and audit), docs/RESULTS.md (all results with their source tables).

Main results (from results/, details in docs/RESULTS.md)

• 648 requests give 547 formal entries in 17 topologies; 13 topologies admit sharing across transistors.
• Sharing diffusion and gates makes 279 of 321 verified layouts smaller (mean -5.5 %, up to -18.4 %). Of the 21 area
  increases over all modes, 19 come from routing and row assembly and 2 from a shared cell whose top-strap pads sit past
  the row ends.
• The plan-level cost rule, frozen on two development sizes, gets the sign of the area change right for 795 of 872
  held-out pairs and finds a plan within 1.3 % of the best of a 64-plan catalogue with one verification.
• Language models given the interface reach layouts smaller than every catalogue plan in 16 of 28 tasks; with an
  interface limited to base gLayout, 41 of 42 sessions end at the unshared area.

How it was checked

All in the pinned container image, on the commits of this branch:

• pytest (17 tests): plan legality and construction, enumeration counts, the program parser, rebuilt areas of the
  running example, data counts, and that every number quoted in the documents equals the result tables.
• python -m textbook_sharing.analysis.all recomputes all result tables and the generated documents from the run
  records; afterwards git status is clean, so they are byte for byte the committed ones.
• scripts/reproduce_clean.sh on 41ed33d: five layouts exported with git archive, built and verified from scratch,
  match the recorded geometry hash, area, status and every verification stage
  (results/analysis/repro_clean_20261012T140050Z.json).
• python -m textbook_sharing.analysis.geometry_check rebuilds all 6159 recorded layouts: 6093 have the recorded
  geometry hash, including every main-grid layout and every model proposal. The other 66 are SKY130 selection
  candidates evaluated before the implant-bridge rule was added; they differ only by 0.35 um^2 of N+ implant, and a
  full re-verification of all 66 with this code reproduces their recorded status, area and stage results
  (results/analysis/geometry_check_diagnosis.json).
• The visual review was done by a vision model under a blinded protocol, not by a person; every flag was resolved by
  objective checks (official DRC/LVS, connectivity checks), stored in each record.

Not in this pull request

• The paper, slides and report (kept separately).
• GDS files, tool logs, renders and extracted netlists (rebuilt by python -m textbook_sharing.run_case ... --from-record),
  and the raw model transcripts.

Known limits

One row-based assembler (another router could move the losses); bounding-box area; one size grid and no Monte Carlo;
the class-AB stage stays diagnostic; magic DRC is advisory; base gLayout in the interface comparison is emulated by one
restricted call; selection candidates outside the reported set were not reviewed visually.

🤖 Generated with Claude Code

[DonaldLucy]

Please don't merge this directly to the main repo. I was trying to open a new branch for this but seems like I don't have the permission