# Registered reference-envelope mechanism grid

This instantiates the prior reference-envelope design after the completed
pelvis pilot and before any of this grid's outcomes. The ongoing live binary
panel must close its allocation before this study's first geometry call. No
partial live outcomes select the grid or inputs. This is development analysis
on two inspected programs, not holdout or hardware evidence.

Use the qualified disabled left/right strafe-walk-stop5s captures at generation
seed1703, simulator seed1701, decision254 (5.08s), original STOP6.08s and recorded
endpoint355 (7.10s). Input files contain only the immutable task, accepted native
reference, accepted15-field resident reference, actual legacy reference and
completed causal state/action/token history. Bind these to the qualified null
source receipt. External full trajectories and outcomes are excluded from the
model/geometry request source map.

Fixed order: left then right; for each, tube0.05/0.10/0.15m; for each tube,
settle0.12/0.24/0.40s. There are18 geometry attempts, each a single frozen-v5
construction with recorded cold returns and independent geometry and native/
converted rate checks. Lead0.8s, earliest native change5.12s, yaw prior3deg stay
fixed. Persist raw results before wrapper checks; retain every fallback and
its reason. Do not rerun a failed solver or substitute another method. Record
requested versus achieved hold timing even when reference admission fails;
a timing proposal on a failed reference is not usable stopping support.

Close all18 geometry rows first. Then for left/right in the same order, run one
fresh guarded nominal current forecast and one per eligible candidate in grid
order, at most20 forecasts total. Failed references have an explicit no-forecast
row, not a predicted failure or an invented current forecast. Model integration
failure is retained; infrastructure failure halts later attempts and retains
all unattempted allocations. No fitting, offset correction, pelvis shift,
new motion generation or simulator execution is part of this offline stage.

Forecasts retain the full original consumed budget and strict original task
constraints: position0.20m, heading10deg, speed MAE0.15m/s, STOP speed strictly
below0.10m/s throughout the original STOP. The reference prior never changes
these criteria. Report9 candidates per program plus its current comparator,
eligibility, holds, forecast statuses and predicted constraint failures. No
best-of-grid measured success claim is possible here. A candidate's nominal
admission is not calibrated safety evidence.

Interpret separately: does relaxing the reference prior advance the hold;
does that produce usable geometry; does nominal stopping improve? All
outcomes must be closed and audited before allocating a later matched live
comparison. Any such execution needs its own source-bound allocation and
preserves its own fresh histories. Keep final36 sealed throughout.

Before calls, bind implementation closure, assets, inputs, passing fake tests
and this protocol. Root source review specifically checks the cold-return
persistence hook, whole-reference fallback, unchanged task clock and tolerances,
model causal whitelist, 18-slot denominator and no outcome-dependent retries.