Boundary-aware continuation: finite trial

Use the same three preselected unfinished stretches and completed control as the two earlier recovery trials. Start from the original saved first-pass points, preserving every earlier result and using a separate source/input/environment freeze.

• 15 unfinished primary fits, six completed control fits and three prescribed repeats.
• At most 600 residual/tangent calls and two value predictions per start: 24 starts / 14,448 native sweeps, no retries or replacement starts.
• The boundary-aware solver takes bounded linear least-squares steps and verifies actual nonlinear costs, shortening rejected steps. Snap only normalized boundary slack of at most 1e-12; record the initial change.
• Keep stationarity 1e-4, relative cost stability 1e-8 and relative residual stability 1e-6. Every passing fit needs the existing fresh bounded-stability check.
• Preserve physical models, calibration bounds, nominal 200 Hz timing, h1_v3_tau60 covariance, ±1°/hour offsets and both processing assumptions.
• Use the original fixed 0.01 perturbations. Reuse completed reference fits where the original selection specifies; never choose a better starting point after inspecting the newer trial costs.
• All 24 planned starts must pass. Controls and repeats must also agree in relative cost within 1e-6 and in every-axis predicted position within 0.01 m RMS.
• Retain per-evaluation costs, accepted/rejected step fractions, final residuals and Jacobians, epoch errors, source hashes and charged append-only journals.
• One durable worker with two BLAS threads. Charge calls before execution. Interrupted identities stay spent and unresolved; completed trials cannot resume.
• Estimated runtime at the ceiling: 10–14 hours, based on the same recordings' prior evaluation times. Warm starts may finish sooner; the extra bounded linear calculations are not yet timed on the full observation model.

The software comparison passed 40/40 known-answer controls with this candidate, versus 35/40 for the original rectangular method and 38/40 with an exactly snapped start. Software success alone does not resolve recordings. This trial tests transfer to the actual observation model. It issues no calibrated scientific decisions and does not change the original 12 globe preferences, zero completed flat preferences and 41 unresolved.

Success supports the separately frozen all-41 recovery stage. Failure remains recorded and cannot be erased by spending unused allowance on retries.