Extend byte equality to broadcast and initial subasset issuance

[droplister]

Continues the verification architecture (#214, ADR-019): two more compose types move from field-by-field comparison to whole-message byte equality.

Broadcast

CBOR [timestamp, value, fee_fraction_int, mime_type, text] after type id 30. Two format facts checked against counterparty-core (ca2496d) rather than recalled:

• value is always an 8-byte double. The API coerces the param with float() and core calls cbor2.dumps with default options — default, not canonical, meaning finite floats are never shortened to float16/32. Verified empirically under cbor2==5.9.0, the version core pins. The pack-side CBOR encoder gains float64 support (number = float, bigint = int) and refuses non-finite values, which core's compose rejects anyway.
• The timestamp never reaches the packer's params. The form doesn't carry one; composeBroadcast stamps the wallet's own clock into the request. It is borrowed from the decoded message through the Observed channel, bounded to ≤ 1 hour ahead of the same clock that stamped it — an honest response echoes a timestamp taken moments earlier, while a substituted future timestamp settles a feed's open bets before their deadline (broadcast.py settles once timestamp >= deadline). verifyBroadcast applies the same bound on the field-comparison fallback, so a refused borrow doesn't become an allowance. An explicit timestamp=0 (server continues the feed from ledger state) is declined.

Initial subasset issuance

Nine-element CBOR layout under LR_SUBASSET (23), with two byte-level details that differ from the standard layout:

• Flags are ints. Core's subasset branch writes 1 if divisible else 0 where the standard branch passes booleans — different bytes in CBOR.
• The longname is compacted by a new base-68 encoder mirroring the existing decoder in unpack/messages/issuance.ts, emitting minimal big-endian bytes as core's canonical_subasset_compact gate requires.

The numeric asset id is drawn by random.randint at compose time — server-chosen, so it's borrowed from the composed message under a range guard pinning it to the space core draws from. Safety argument (verified in core's validate()): an id naming someone else's asset is consensus-rejected ("issued by another address"); an id naming an asset the user already owns degrades the transaction into a reissuance of the user's own asset to themselves. Neither pays an attacker, and the longname, quantity, flags and description all stay byte-compared.

Still deliberately declined to the field fallback: subasset reissuance (standard layout, ledger-resolved id), ord-inscription composes, and non-text MIME types (core restructures or hex-decodes their content) — the last now declined across standard issuance too, where the packer previously assumed UTF-8.

Oracles and tests

• Compose oracle: broadcast cases with explicit timestamps, plus a subasset case that decodes core's response and borrows the random asset id from it exactly as production does — equality then proves every other byte.
• Round-trip oracle: rebuilds recent on-chain broadcasts and subasset issuances (the issuance decoder now surfaces the wire's mime_type — CBOR layouts only, never invented — which the glue and display can both use).
• Unit fixtures generated with cbor2==5.9.0 plus core's own compaction arithmetic, embedded as byte-exact expectations, alongside borrow-bound and decline tests.

Verification

• tsc --noEmit clean; counterparty unit suite 638 passing.
• Both oracles run against api.counterparty.io:4000: 49/49, including the new cases.
• E2E locally, one file at a time: compose/broadcast/index.spec.ts (10/10), compose/issuance/index.spec.ts (10/10).

https://claude.ai/code/session_01CcjnCrgosSeshymXLBxdGj