v1.7 — harness, extended-precision edge limit, and the stability statement corrected

Adds harness/: a fixed-support builder with the archimedean, pole and prime blocks exposed separately; a validator that checks the assembly non-circularly against ordinates from mpmath.zetazero, reporting residuals against explicit tail bounds; a regression harness that sweeps the m-ladder, prints the full inertia triple with its stated tolerance, writes raw arrays and JSON metadata, and exits nonzero on failure; and edge_precision.py, which evaluates the edge limit through a reduction with no trigonometric cancellation, at a precision set from the target δ. Built to a specification supplied by A. Groskin.

With that precision the (m−1)/(m+2) law is confirmed to converge like δ², the O(δ⁵) remainder holds to δ = 10⁻⁶, and the parity ratio is measured as σ₂/σ₁ = C(m) ε⁴. Running the module with --dps 16 reproduces the float64 turnaround, showing it to be arithmetic rather than mathematics.

The prime-edge note is corrected throughout. λ_min converges — 2.39%, 0.75%, 0.39% across the ladder at p_c = 60 — while the inertia counts do not, which is why the harness gates on λ_min and reports the counts without gating on them. Four precision edits due to A. Groskin are applied, the asserted equality d = 2c − 6 is withdrawn, and Bombieri's count theorem is explicitly not invoked, being stated for a different matrix family.

Four retained scripts are corrected in place so that no script prints a claim its own output does not support: the section-B conclusion in prime_edge_jump.py, the ε/δ labelling and the missing precision warnings in check_reply.py, and the withdrawn scalar-verification reading in both prime_edge_jump.py and prime_edge_rank.py