Scaled affine correction qualification and matched trial

The saved correction audit reproduces the failed right-side step using the saved center Jacobian. The old bounded solver reports success but returns a higher affine cost than the feasible zero correction. This is a numerical failure before nonlinear propagation.

Candidate

lll.ilvis0_scaled_branch reparameterizes each affine correction coordinate by its available range before applying the existing bounded-variable least-squares method. Every column and original physical bound remains present. The mapped solution must be feasible, no worse than a known feasible reference, and meet a range-scaled affine stationarity check. The outer native stationarity, cost, position, continuity and source-to-final checks remain unchanged. Neither fixed timing nor a damped zero step is substituted for the full problem.

The local affine check uses at most 100 iterations and tolerance 10^-10. Its additional stationarity check uses 10^-8 and projects bounds within 10^-6 of each coordinate's allowed range. This qualifies the affine solve only; it does not replace the native 10^-4 stationarity or any physical-position limit. All candidate modules are private copies; frozen dependencies remain unchanged.

Software qualification

One 33-test batch, 120-second ceiling: all 21 existing representable-clock/native motion controls applied to the new candidate, nine additional affine controls and three worker/scope checks. The affine controls include known bounded optima, rank deficiency and both exact saved failed systems. At most 1,024 callback/native software calls, 81 compiled-motion calls and six clock comparisons. No originals or observed native calls are permitted during software qualification.

Only a complete qualification pass permits observed preparation. Freeze the qualified sources, input hashes, native kernel, environment, task identities and this plan before launch. Preserve unsuccessful exploratory solver tests; the final candidate retains BVLS with coordinate rescaling rather than changing both parameterization and solver family.

Observed limits and acceptance

Scope remains only 370cec66878c4c1cf84b4ceb, profiled-removal specified-field globe, complete original primary and repeat, plus two fresh smooth controls. Two controls × three calls and two target checks × 26 calls: four charged identities, 58 native evaluations, zero optimizer starts. Maximum 402,048 clock queries, six bounded affine solves, four total numerical threads, no retries.

The corrected, already qualified receiver-context adapter remains unchanged. Do not alter intervals, clocks, measurements, covariance, physical bounds or processing assumptions. The two completed earlier trials remain frozen, including their errors, failures, charges and unused allowances.

Require controls first, full two-sided native gates, source reproduction and cumulative source-to-final stability. Require relative pair cost difference at most 10^-6 and per-axis track RMS at most 0.01 m. Preserve every failed start and charge before calling native code. A pending transaction blocks continuation; completed or charged identities cannot restart.

Even a pass needs a separate saved review of all six primaries, both required case repeat pairs and controls. No production promotion, significance claim or change to 51 globe / 0 flat / 2 unresolved follows from qualification alone.