domino oracle: generator-faithful place sampler + InFront name precedence

[yichao-liang]

Summary

Two independent fixes to the open-loop oracle_process_planning demonstrator on pybullet_domino, taking it from 3/5 to 5/5 on seed 0. This is distinct from the agent_bilevel domino path fixed in #42 — it targets the oracle demonstrator's grid-process place sampler and predicate setup.

1. Generator-faithful rank-sum place sampler — ground_truth_models/domino/processes.py::_place_sampler

The sampler read the bare grid cell — a uniform lattice that omits the task generator's inward domino_width/2 corner offset, so a turn block landed on the same cell a straight block would and corner cascades stalled. It now enumerates the generator-faithful candidate placements (_generator_placements, which carry the corner offset) and rank-sums three signals — future-target bridge, planner grid-cell distance, planner angle error — to pick the cascade-correct pose deterministically (top-ranked on ≥2 of 3; falls back to the bare cell only if no candidate exists).

2. Helper-predicate name precedence — approaches/process_planning_approach.py::__init__

The env's position-based InFront (a Predicate) and the grid's derived InFront (a DerivedPredicate) are ==-equal but hash differently, so a plain helpers | base set union kept both; abstract() then evaluated the looser position-based one, hallucinating adjacencies (e.g. the start block "in front of" a staged movable 13 cm away) that let the planner build a physically-impossible single-block bridge to the target. Now drop any base predicate whose name a helper already provides before the union, so the derived InFront fully replaces the position-based one.

Results

Open-loop oracle_process_planning (bilevel_plan_without_sim) on pybullet_domino, 10 test tasks × seeds 0–4 — the same task set, run with vs. without these fixes (baseline = master 95239a7, the parent of this change):

Seed	Without fix	With fix	Δ
0	3 / 10	7 / 10	+4
1	6 / 10	8 / 10	+2
2	6 / 10	8 / 10	+2
3	4 / 10	7 / 10	+3
4	1 / 10	5 / 10	+4
Total	20 / 50 (40%)	35 / 50 (70%)	+15

Both runs have zero crashes or timeouts (num_solve_failures = num_execution_failures = num_*_timeouts = 0 for every seed) — every miss is a cascade that executed cleanly but fell short of toppling the target, so the fix converts 15 of those "policy failed to reach goal" cases into solves. The remaining 15 misses are the long tail (harder layouts where staging geometry overlaps the chain region / occasional Pick collisions), which needs a task-generator change beyond this PR.

yapf · isort (5.10.1) · docformatter · mypy (1.8.0) · pylint all pass.

[yichao-liang]

Added a no-fix baseline to the results table above — same 50 tasks (10 × seeds 0–4), run on master 95239a7 (this change's parent): 20/50 (40%) without → 35/50 (70%) with the fix (+15 solves; zero crashes/timeouts in either run, so the fix turns 15 cascade-fell-short misses into solves).