Receiver-coordinate refitting controls

Purpose and scope

Qualify the receiver-projection interface for nonlinear fitting using generated motion whose receiver trajectory is known independently of the navigation predictor. This is software development, not another synthetic Earth-model campaign or a fit to archival observations. Preserve all completed studies.

Reuse the existing closed-form geographic controls: twelve seconds of level eastward motion with changing body yaw and pitch, independently integrated body increments, off-grid receiver epochs, and known calibration and timing errors. The three physical models each generate their own truth. The nominal and variable-interval constructions include interior and near-bound calibration. Short controls verify the fitting interface, not Earth-model discrimination.

Implementation

1. Add lll.ilvis0_projected_fitting.ProjectedProblem, wrapping the full native residual/Jacobian with ReceiverProjection. Preserve every active parameter, the full prediction interface, original initialization and covariance. Expose retained-coordinate physical errors separately from unscored coordinates.
2. Reuse the existing bounded-step solver with exact derivatives and fresh cost, stationarity and physical-position stability checks on retained coordinates. Explicitly label omitted coordinates as unchecked. Report numerical rank and abstain scientifically even when the numerical solution passes. This adapter is not a replacement for the qualified full-coordinate archival policy.
3. Test direct marginal-objective/derivative agreement for all 26/27 columns, independent-trajectory recovery, different starts, exact bounded optima, lost rank, invalid inputs and evaluation exhaustion. Do not demand unique parameter recovery when the retained observations cannot identify them.
4. Freeze sources, environment, finite limits and test identities before the final qualification batch. Publish the batch outcome and individual test records. Development test runs remain separate from this sealed batch.

Finite qualification limits

• At most 24 generated-motion optimizer starts: three models × two clocks × two coordinate subsets (horizontal/height) × two declared starts.
• At most 80 tangent sweeps per generated-motion fit, including its fresh final check; at most two additional value predictions per fit.
• Six all-column native derivative controls, at most 56 sweeps each.
• At most 2,400 native generated-motion sweeps total, charged before calls.
• Algebraic adapter/solver checks use constructed arrays, no native sweeps.
• One qualification test batch, 600 seconds, four numerical threads, stopping at the first failed control, with no retries or allowance expansion. Compile local kernels once as needed. Unrun controls remain explicitly untested.
• Zero original-recording reads and zero observed-data optimizer starts.

Retain all columns after projection. Rank uses the declared floating-point SVD cutoff; it is not an empirical identifiability threshold or a confidence interval. For generated motion require independently known retained-position errors below 2 mm and objective recovery against the known-parameter reference; report unscored-axis errors without using them to manufacture a pass or a failure. Direct-marginal comparison must use the identical covariance and candidate. Finite differences use the separate native value kernel; compare value agreement in physical units at the established 0.3 micrometre tangent-control tolerance, rather than requiring bit-identical accumulation between different kernels.

Development exposed a failed SciPy trust-region prototype on the near-bound flat horizontal control: 80 sweeps did not meet stationarity, despite a 0.02 mm retained-coordinate error. Its unscored height error was about 6.9 m. Preserve that prototype and diagnostic separately; it has no qualification or observed results. The sealed candidate reuses bounded steps and unchanged fresh-check tolerances, rather than relabeling that failure or relaxing the acceptance gate. Use the wider native arithmetic already used in archival recovery. It did not eliminate the development height-only stopping failure; retain that result. The explicit qualify_ilvis0_projected_fitting.py matrix is run by the study worker, separately from routine CI tests of the adapter and failure reporting.

Acceptance and next step

All frozen controls must pass. Any failure blocks advancement; preserve its evidence rather than retrying the sealed batch. Passing qualifies this interface and these controls only. A small observed refit proposal must separately freeze its stopping/stability rule, model/case identities, starts, repeats and budget. No archive rollout, changed 53/0/0 findings or calibrated significance follows.