docs(sofi): the recovery object closes the last slot, and a market b becomes constructible

[cryptskii]

What this closes

2c-A froze the settlement bundle and stopped one field short on purpose. MarketTerms (0x0033)
field 6 recovery_material is mandatory and its nested class was 2c-B's, so there was no
encodable market MarketTerms, therefore no encodable market SettlementBundle, therefore no
conformant market b. Registry §6 named it as the single remaining encoding blocker.

This closes it. A conformant market b is now constructible.

0x0031 DsmSuccessorEvidence — substrate, seven fields

The six inputs the relationship chain tip commits, plus sigma_dsm. Its address is computed over
canonical CCB bytes, which is the whole point: the shipping object violated §2.10 twice —
addressing itself over prost bytes, and using prost determinism as its canonical form.

c_dsm_plus is not a field. It is doubly derivable — the chain tip recomputes it from the six
inputs, and it equals trader_successor — so carrying it would state one fact three times, and
the existing verifier already recomputes rather than trusting the carried value.

Two foreign grammars, frozen as grammars

CloseAuthorizationPreimageV1 (discriminator 28, ten fields) and DlvSettleOperationPreimageV1
(discriminator 26, nineteen fields), each enumerated with exact widths. Both use little-endian
lengths and length-prefix their 32-byte values — the opposite of CCB's big-endian digest32,
stated explicitly so an implementation cannot conflate them.

The shipping Operation::to_bytes is the current implementation of these grammars and is
evidence of the byte grammar, never the normative authority.

That distinction is deliberate: it stops a future Rust refactor from silently changing a protocol
signing preimage. Freezing a CCB close preimage instead would have invalidated every existing close
signature.

Validity is ordered, and the comparison set is exhaustive

The encoding-level conjunct on operation_bytes runs before the chain-tip equalities, because
without it a producer who computed trader_successor over arbitrary bytes satisfies the equality
trivially. And 2c-B performs exactly two cross-object comparisons plus the syntax checks —
exhaustive by construction, so everything further is ValidDlvSuccessor and belongs to 2c-C.

Encoding closure is not production enablement

Stated in the amendment's own voice. Four DlvSettle fields are caller-supplied and never checked:
sigma is hard-coded zeros, and settler_public_key is never compared with the key that actually
signed. The chain-tip conjunct pins them so reconstruction is exact; nothing establishes they
are meaningful. Production acceptance stays gated on 2c-C.

A boundary three documents overstated

The owner-close shape was described as "fully constructible today" under 2c-A alone. It is
encodable given already-prepared close_authorization bytes — and the grammar for producing a
fresh one is frozen here. Corrected in registry §4, §5.19 and §7, and in two places in 2c-A.

Size — measured, not estimated

Encoding the objects from the frozen tables, with a pinned fixture:

CloseAuthorizationPreimageV1, signing form        150 bytes
DlvSettleOperationPreimageV1                  100,446 bytes   (2 signatures inside)
0x0031 DsmSuccessorEvidence                   150,447 bytes   (adds sigma_dsm)

three SPX256f signatures                      149,568 bytes   -- the FLOOR

vs 512 KiB authenticated ingress cap             28.7%   ~3.5x margin (close was 10.4x)
vs 128 KiB b0x envelope cap                       115%   EXCEEDS

A market bundle carries three signatures where an owner close carries one. Two consequences:
the market margin is about a third of the close's, and a market bundle does not fit the b0x
envelope at all — which turns "settlement bundles never traverse b0x" from a convenience into a
load-bearing constraint. Same shape as the incident behind this repo's prior 413.

Verification

Namespace re-audited before writing (code tops out at 0x0030; 0x0031/0x0032 in no Rust
constant or proto). §2.8 audit: seven new field numbers, no reuse, no new burn — first ship at
schema 1. Both grammars encoded from the amendment's tables alone, with the close preimage's
arithmetic closing independently at 150 bytes and the cleared signature rendering as 00 00 00 00.
Def 6.26 re-read at its line range. Self-contradiction sweep clean. §4 counts re-derived and sum to
22 live.

Owed at implementation, and stated rather than implied: byte-equality of both grammars against
a running Operation::to_bytes, and the hostile malformed-operation_bytes case, both need a test
binary this docs-only change cannot add. The first commit implementing 0x0031 must assert them —
and if they differ, the amendment is wrong, not the code.

Documentation only.