Retire the ×1.11 deep-tail correction

[mjaskelioff]

Closes #165.

The dedicated measurement (tail-correction.md, 370 000 fresh draws across both designs) reads the deep-tail small-component factor at 0.994 ± 0.021, rejecting ×1.11 at z = −5.7; the cell the factor was originally measured on (M4 RF = 7) does not reproduce. This pass retires the factor from every document that carried it, per the plan agreed in the #151 review thread.

What changes

• M4 μ_design = 0.35 stands at RF = 11 (172 896 msgs, 13.3 copies/honest) — the only decision the factor ever flipped. RF = 12 remains noted as the option for a deployment wanting μ-shift slack at that grid point. Comparison §5 tables A/B unstarred; the frontier verdict's M3 lead at 0.35 becomes ~7 % (the old 7–15 % range spanned the retired RF = 12 reading).
• M3 (17, 7) at μ_design = 0.3 regains its raw +0.004 margin — the "essentially no margin with the correction" reading is gone.
• The "×1.11 corr." / "×1.1 corr." table columns and all corrected-value parentheticals are dropped; budgets stand at the raw law crossings.
• The "roughly one bad graph in ten is a ≥ 2-node islet" severity inference is dropped; the severity sections instead cite mu-sweep.md, which validates the isolated-vertex laws across μ = 0.2–0.4 with no small-component excess.
• sweep_m4_reprovision.py loses its hardcoded 1.1 * law(RF) corrected-row branch.

Editorial rule (as agreed in #165)

Validity caveats are replaced, not deleted: the tail at the operating points stays law-read, not directly measured; what changes is that the second-order term is now measured absent instead of carried at ×1.11. Every rewritten validity note cites tail-correction.md including its §4 limit — the factor is measured at P(bad) ≈ 5–8×10⁻³, two decades above the 10⁻⁴ operating tail, and constancy across that range is the same extrapolation the laws already carry.

Deltas vs the issue text (written before #187 merged)

• Table B premiums are measured against the RF = 9 margin selection, so unstarring M4's 0.35 premium keeps −19 % / +22 % (the issue's −8 % / +38 % was computed against the old RF = 8 baseline).
• mu-sweep.md carries no severity section, so it is cited for what it establishes (law fidelity across μ, no small-component excess) rather than as a severity sweep.
• The m3/m4 transmission_unreliability.md files (new in Select operating points with a 2% disturbance margin #187) already used the measured 0.994 factor and are untouched.

Verification

• validate.py passes (check 6's searches use the raw law — factor-independent).
• Default runs of sweep_m3_reprovision.py, sweep_m4_reprovision.py, sweep_m3_mu_shift.py, sweep_m4_mu_shift.py match the updated tables.
• Relative-link check on every changed file. One pre-existing dangling link surfaced (not introduced or touched here): comparison.md cites pubsub-node/docs/experiments/churn-proposed-points.md, which exists in no commit on any branch.

🤖 Generated with Claude Code