Frozen before new decoder experiments, 2026-11-02, after paper draft.

Question: if the next odd valuation is unresolved at H source bits,
we still know its total cumulative division cost is >=H. If the current
odd-step numerator coefficient is <2^H, then every completion has a
negative coefficient at this next odd endpoint. Can we reject this whole
cylinder without decoding the exact exponent or requesting future data?

Hypothesis: this extra rejection is sound for Haar coefficient selection
and for actual integer no-descent outside the SAME T_H source guard.
Reason to audit: first odd-step crossing may have total valuation >H,
so the old guard statement (which demanded crossing cost<=H) cannot be
quoted blindly. Need new direct proof using ell_k<=H and prefix envelope.

Test menu: H8,10,12,14; n1,2,3,5; all odd residues, all8 high-bit
extensions; literal actual sources at offsets0,10^6 and10^30+17.
Retain unknown terminal reserves; only add a rejection, never promote an
unknown to selected. Compare both decoders; retain counterexamples and
explicit source exceptions. Cap5million pair checks and120seconds.

No new asymptotic horizon is predicted. Under a fair parity model this
may merely recover the already available half-tilt exponential rate with
better prefactors. Do not modify the manuscript's theorem to rely on it
until a complete proof and independent finite checks are available.
