fix(sofi): the relay carries the bytes that hold a leg cell; the §6.39 owner rulings

[cryptskii]

The owner's rulings of 2026-09-29 on CONFORMANCE §6.39's open items, and the SoFi relay defect that tracing MR-SOFI-0255 found.

The relay never carried an exercise to a leg key

The defect. sofi_relay::relay_fulfillment took the id of the value holding a leg cell, then looked its bytes up by entry digest (value_of). An attempt cell names its holder by E (the external commitment), which is no digest of any bytes. So the lookup never matched, and every relay wrote the position pair alone.

The evidence. A throwaway probe on nodes, after a realized trade:

• the leg cell read Held { id: E, Final };
• every carried value's entry_digest differed from E, and each recognized as the exercise with that E;
• value_of found nothing.

The fix (owner's choice): Core keeps the exact bytes it recognized as holding the cell, instead of anyone looking them up again by id.

• CellReading::Held carries value, the leader link's bytes, which evaluate previously dropped.
• AttemptCellRead::value() exposes them.
• The relay carries exactly those bytes.
• sofi_advance::final_root_cell reads a final root claim's bytes the same way.
• value_of is deleted.

Regressions:

• a_held_attempt_cell_carries_the_exact_bytes_that_hold_it (id E). The same F and P under signatures that do not verify arrive first; the read carries the signed exercise's bytes, and E is not their entry digest.
• a_held_root_cell_carries_the_exact_bytes_that_hold_it (id = entry digest). A claim naming the next position's key arrives first; the read carries this key's claim.
• On nodes, a two-hop relay by a device that is not the trader wrote 2 cells before the fix and 4 after (local runs).

§6.39 rulings

• removeContact: deleted in every layer:

  • the JNI export;
  • the Kotlin declaration and wrappers (UnifiedNativeApi, Unified, UnifiedContactBridge);
  • client_db::delete_contact_by_id and remove_contact;
  • sdk_remove_contact;
  • Core's DsmContactManager::remove_contact, with their tests.

  A future removal is a defined retirement or tombstone flow.

• dsm_init_runtime: the undeclared C export is deleted; get_runtime stays lazy.

• dBTC builtins guard: the load-time #[ctor] and the ctor dependency are removed, in all four lockfiles that carry dsm_sdk. No dBTC code changed. assert_builtins_sound runs in no shipped build and belongs to the owner's dBTC redo; its manifest row carries the dBTC exception.

• MR-STOR-0146, now Met.

  • Every SDK spool write goes through B0xSDK::deliver with bytes seal_for made.
  • A node's spool holds only what its submit route receives: spool_insert has one caller.
  • a_transfer_reaches_the_nodes_only_sealed_and_arrives passes on Postgres.

• MR-DSM-0272, now Partial. Payloads are sealed end to end: XChaCha20-Poly1305, a domain-tagged key per message id, and one seal kept per id. But the key comes from a fresh ML-KEM encapsulation per payload, not the step's own Kyber exchange that Amendment A7 names. spool_seal's module doc claimed the step binding; it now says what the code does.

• MR-SOFI-0255, now Partial.

  • The router sends exactly the eight sofi.* methods to their routes, and each route reaches its producer.
  • A test driving all eight on nodes found the relay defect above, and a second one:
    • the owner's sofi.close of a vault that has been traded through never resolves (RouteEvidenceHasNoSource for three trader leaves);
    • a fresh vault closes fine;
    • it is recorded in §6.39 as confirmed, remediation deferred (owner).
  • The eight-route test lands with the close fix; no red test lands here.

Map and manifest

• Entry points. Three removed; 67 Android entry points remain, plus the node's main.
• Tables. One sentinel removed. The load rule has no real example any more; its unit tests remain. The Android counts are updated.
• Intent manifest: 1,074 rows.
  • +8 MR-SOFI-0255, +2 MR-STOR-0146, +2 MR-DSM-0272, and a dBTC fate row for assert_builtins_sound.
  • −1: the guard's row.
  • Every new row reads SATISFIED, and no Android row fails.
  • Locally the node's rows read not-built; CI's --built android,node reads them.

Verification (local, macOS)

• Targeted --release tests:
  • Core route_chain, sofi, economic::register and spool_seal: 54/0.
  • dsm_sdk contacts, runtime and policy.
  • The full node_e2e_tests module on Postgres, except the held test.
  • sofi_advance.
  • The two regressions.
• make lint exit 0 (fmt and clippy); the real-code guard is clean.
• ci/conformance_evidence.py: 779 rows, totals regenerated.
• make requirement-map and requirement-map-check: 0 contradictions, 24 sentinels, 0 failures.
• requirement-map-fixture: 0 unexpected outcomes.
• requirement-map-mutations: 40 cases, 0 failed.
• cargo test -p requirement_map: 97/0.
• Gemini gate: SATISFIED (round 2).