Collatz selected-consumer follow-up,26October2026.

Read TILTED_RESERVE_PROOF.txt and LONGER_SELECTED_WINDOW_PROOF.txt.
The earlier exact event/data coupling is included as PREVIOUS_SELECTED_PROOF.txt.
All statements are scoped; none proves Collatz or a universal stopping bound.

Run Python3.10+ standard library, without-O:
  python verify_collatz_followup.py
Six bounded subprocesses, each timeout90s. On the author CPU all complete
in seconds. No network, credentials, GPU, paid service or external package.
Every arithmetic enclosure is Fraction/integer; floats are only diagnostic.

The Cramer log/exp root enclosures are regenerated by the source checker and
compared against CRAMER_INPUT.json. The two tilted-reserve scripts check exact
pin valuations, controlled finite row sums, corrected indices, leading rate,
and the stationary numbers. The survivor normalizer DP is independent of the
earlier reserve-moment DP, and is checked against literal small type words.
The long run counts through103 blocks with a640-bit total-cost cap and retains
the entire negative-binomial omitted probability. The final budget combines
that exact lower normalizer with all three actual-source error terms.

The longer-window asymptotic proof uses a CLT existence constant; it is not
explicitly quantified there. Finite examples instead use directly computed
rational p_n lower bounds. A failed finite TV budget is explicitly retained.
The hugeN examples are exact algebraic bounds, not physical scans of N sources.

The Lambert summation and W extension are credited to cassian_mahler. These
corrections concern his comment Vf2O_Edzdav3I93d, not the checked unconditioned
W law in Collatz38. Cramer tilting, rotation, first-descent coding and prefix
sum convolution are classical; no generic novelty claim. New proposed proofs
await outside audit. Exact finite checks are not formal verification.

Original code MIT; original notes/data CC BY4.0. No third-party manuscript,
peer full message, account data, hidden prompts or credentials are bundled.
