Qualification and finite representable-boundary diagnostic

Numerical question

The completed boundary study used 82 native calls and passed both controls and both profiled-removal endpoints. Its two unsubtracted endpoints found a timing slope change but stopped at the minimum spacing guard. Their saved side points are distinct and match their intended clock states. The audit used twelve saved-clock queries, no native predictions and no original IMU data.

This candidate changes the geometry qualification, not the scientific stopping limits. It must not simply disable the guard or accept any slope change. It checks explicit interval membership, distinct floating-point values, an eight-ULP parameter margin, distinct adjacent clock states and agreement of all four side points with those states. An ULP is one representable floating-point spacing at the parameter value.

Implementation

lll.ilvis0_representable_boundary executes private copies of the existing adapter, source-preserving wrapper and two-sided stopping code. One checked substitution replaces the fixed absolute minimum spacing with the direct representability rule. Cost, stationarity, position, continuity, bounded affine corrections, source-to-final movement and repeat equations stay unchanged. Frozen helpers and earlier results are never edited.

The direct clock check reserves six of the existing 1,024 geometry queries; the transition enumerator retains its 200,000-query cap. Search remains within one microsecond, with at most twelve derivative evaluations and 26 total native calls per target. A failed or collapsed interval, wrong clock state, unsupported correction or unstable native prediction remains a failure.

analysis/tests/ilvis0_representable_worker.py uses a private instance of the previous diagnostic executor. It selects only the exact two audited unsubtracted sources and two passing controls. The profiled-removal pair must already have passed its original checks and repeat comparison before preparation is permitted. It is sealed as prior evidence rather than rerun.

Frozen software qualification

One 31-test batch, a 120-second timeout, at most 1,024 software callbacks, 81 compiled-motion calls and six compiled clock comparisons. Four total threads; zero original archive reads or observed calls. Freeze sources and the finite software allowance before running it. Preserve any failed batch.

Controls cover narrow and broad known minima, maxima, discontinuities, nuisance failures, collapsed/under-resolved branches, invalid partitions, wrong clock states, budget exhaustion, source mismatch, charged native errors, nonfinite values, unchanged frozen modules, exact task scope and repeat agreement. An independent compiled packet-clock comparison checks all side states. Three compiled increment-integration controls use independently specified straight trajectories and test observation agreement at the declared physical resolution.

The motion controls do not demand nanosecond recovery of a time parameter when its position effect is below the existing millimetre stopping resolution. They require the accepted predicted observations themselves to agree with the independent analytic track. This is not a demonstration of precise timing, instrument calibration or observational confidence coverage.

Observed diagnostic allowance, only after qualification

The target is stretch 0e73879f98c3b701fd7c42df, zero-field globe, bias1_unsubtracted, primary and repeat. Use each completed source's full parameter vector and hash-checked residual/Jacobian artifact, not an interrupted identity. Keep the selected interval, original measurements, receiver weights, nominal 200 Hz clock, physical bounds and model unchanged.

Seal exact tasks, qualified sources, this plan, original hashes, saved artifacts, kernel and numerical environment before launch. Hold one lock per distinct source directory, require one GiB free storage, charge calls before execution and retain failures. An incomplete transaction blocks continuation. Never restart completed work, retry a charged identity or reuse an older study's unused allowance.

Both fresh controls must pass first. Each target must pass source reproduction, the revised explicit geometry check and the unchanged full two-sided native checks. Cumulative movement from the saved source must pass. The endpoints must agree within relative objective difference 10⁻⁶ and per-axis track RMS 0.01 m. Report all outcomes and search traces.

Even a complete pass requires a separate saved companion/model review covering both processing assumptions and all three physical candidates. No label change, threshold calibration, wider corpus run or causal rotation claim follows from this diagnostic alone. Current published findings remain 50/0/3.