bip360: bound merkle depth at 128 and replace validation asserts in ref-impl

[let-the-dreamers-rise]

Two conformance/robustness fixes to bip-0360/ref-impl/python/p2mr.py, found by differential testing against an independent implementation (byte-for-byte agreement with the reference on all 16 official construction vectors and on 2000 seeded random script trees before divergence analysis). Reported to the BIP authors by email first; opening it here at @EthanHeilman's request so it is trackable — issues are disabled on this repo, so a PR is the closest equivalent.

1. No depth bound — the ref-impl can emit consensus-invalid control blocks

BIP 360 Script Validation: the control block "must have length 1 + 32 * m, for a value of m that is an integer between 0 and 128, inclusive." compute_merkle_root / compute_control_block applied no bound, so a 129-deep tree produced a normal-looking root and a 4129-byte control block (m = 129) — an output consensus must reject as unspendable.

import p2mr
leaf = lambda s: {"leafVersion": 0xC0, "script": s}
tree = leaf("51")
for _ in range(129):
    tree = [tree, leaf("52")]
print(len(p2mr.compute_control_block(0, tree)))   # 4129 before this PR

Now refused with ValueError; m = 128 is still accepted.

2. Validation asserts are stripped under python -O

assert len(tree) == 2 vanishes under -O, so a malformed 3-child branch silently returned the root of its first two leaves — the third script leaf dropped from the commitment with no error:

$ python3 -O
>>> import p2mr
>>> l = lambda s: {"leafVersion": 0xC0, "script": s}
>>> p2mr.compute_merkle_root([l("51"), l("52"), l("53")]).hex()
# identical to the 2-leaf root before this PR, no error

Validation in the construction path now raises explicitly, which survives -O. assert is left in place for genuine internal invariants.

Tests

Adds negative_structure_tests() — ternary branch refused, m = 128 accepted, m = 129 refused for both the root and the control block — run from BIP360_tests(). The existing 9/9 vector suite passes with and without -O.

Two related things I'd rather ask than patch: tapleaf_hash refuses empty scripts (deliberate, or defensive?), and construction silently masks odd leaf versions 0xc1 -> 0xc0 per v = c[0] & 0xfe (would you rather refuse odd versions than rewrite them?). Happy to follow up either way.

Full graded writeup with method and repro instructions: https://github.com/let-the-dreamers-rise/p2mr-assurance-lab/blob/main/FINDINGS.md