row 155: the translate cap leaves the contract; the kernel hands it in the plane's settings and the codec reads it from the plane's own state (stacked on #1030)

[MattJackson]

Row 155 (lane-p3-llm), coordinator/contract-ABI ruling. busbar-contract holds no state (BUSBAR-1.6.0.md Part 4; spec l.3701, l.4589). This PR removes TRANSLATE_CAP_READER from crates/busbar-contract/src/codec.rs.

STACKED ON #1030 (DRAFT until #1030 lands, rule 19). It does not need #1028's CodecHost threading. It is stacked because the retired hot-ABI service slot (PlaneHostVtable::SERVICES, services::arm, the services.rs/host.rs docs) is the same hunk #1030 edits for the tap-fault latch, and rule 20 forbids two open PRs on one hunk. After #1030 lands: merge origin/predev, push once, and ask the coordinator to undraft.

The seam (ruling: the cap reaches the plane through its settings)

• Kernel to plane: the root's kernel_sections hands limits.request_body_max_bytes in the limits section a door plane is opened AND refreshed with. The value is cfg.limits.request_body_max_bytes: the LimitsResolved value the kernel already mirrors into proxy::max_translate_body_bytes, so it is the same number.
• Plane state: Shaping reads it, and Shaping::translate_cap() falls back to the same 32 MiB default (TRANSLATE_BODY_MAX_BYTES_DEFAULT, now a plain const in the contract) when none is handed. The door hands each unit's generation cap to ReplyCtx.translate_cap.
• Codec reads from that state, at all four former reader sites:
  • the whole-answer over-cap check (Reply Whole), with the same whole::over_cap 500 bytes;
  • the relay's same-dialect usage copy (RelayCtx/Relay::from_parts);
  • the stream translators' writers. new_stream_translator_on(.., translate_cap) calls StreamTranslate::with_translate_cap and then ProtocolWriter::set_translate_cap, which drives the Gemini tool-args buffer and every Responses accumulator. Each is an AtomicUsize on the per-stream writer.
  • the buffered-to-stream synthesizer: TranslateCodec::translate_response(.., translate_cap), then a new IrHandle::apply_translate_cap (a defaulted shape method), then synthesize_from_response(.., cap).
• Linked engine (busbar-llm, the default build's llm path): it hands the kernel's own value (busbar_kernel::proxy::max_translate_body_bytes, through its existing proxy import line, so the ports-only count is unchanged) through the same parameters. It also installs a kernel-capped stream-translator factory (engine::stream_translator) in place of the plane's default-capped one. The hot linked lane (root/linked.rs) reads the kernel's value directly.
• Contract: TranslateCapReader, the static, install_translate_cap_reader and max_translate_body_bytes are gone. The hot translate_cap slot stays in the layout but is retired: SERVICES and arm no longer touch it. The kernel stops arming the reader.

Same value, same default, same refusal bytes

• The configured value and the 32 MiB default are unchanged, and the over-cap refusal is the same whole::over_cap answer.
• The cap is read ONCE PER CONFIG GENERATION (in the linked engine, once per relay or translator) instead of per piece. Plugin-owned settings are valid to the next refresh generation (spec §11), so mid-response consistency is the correct behaviour: a live limits reload no longer moves the cap under a body already in flight. Nothing in 1.5.5's oracle exercises a reload mid-stream.

RED

1. Contract compile_fail doctests on TRANSLATE_BODY_MAX_BYTES_DEFAULT: max_translate_body_bytes() and install_translate_cap_reader(..) no longer exist, so no reader static remains.
2. crates/busbar-plane-llm/tests/conformance.rs::the_translate_cap_the_settings_carry_is_the_one_the_door_enforces, run linked AND dropped in:
   • opened with no cap, an openai caller's anthropic whole answer translates (200);
   • refreshed onto settings with limits.request_body_max_bytes: 64, the same answer is exactly whole::over_cap("openai")'s status and bytes.
3. tests/exchange_shaping.rs::the_translate_cap_is_the_one_the_settings_carry_else_the_default.
4. crates/busbar/src/root/tests/door_steps.rs::the_limits_a_door_plane_is_handed_carry_the_operators_translate_cap: the kernel hands the operator's value, or its default.
5. The contract-stateless gate: a static put back is a NEW finding.

Ledger (before → after, on top of #1030)

• qa/contract-stateless.toml: 39 → 38 rows (44 → 38 from predev with row 155: the usage-tap fault latch leaves the contract; the door returns the fault beside its units and the kernel counts it #1030). Struck static codec.rs TRANSLATE_CAP_READER. The codec.rs once-cell OnceLock row stays, because ENTROPY_SOURCE and WALL_CLOCK still hold it until row 155: the codec's entropy and wall clock leave the contract; a codec is handed a CodecHost (stacked on #974) #1028.
• qa/abi-freeze.toml: crates/busbar-contract/src/abi/hot 6100 → 6075 (6224 → 6075 from predev).

Overlap

• Test merges are clean against every non-draft open PR that shares a file, including row 155 (L1#1): the llm door keeps a retried unit for the kernel's next attempt, and ends a final one #890, row 155 B2: the protocol seam names no http type; head fields and u16 statuses, the kernel converts at its edges (TODO 34c (a)) #974, row 155 B1c-1: the plane states its provider keys as a neutral descriptor in its Statement (dormant seam) #981, Law 0 root (PR1): the root's prod source names no plane, transport or auth instance #993, loader: the plugin observer, host series and intake are the dispatcher's (door-only) #1009, egress seams: HOSTLESS moves onto the plane's build context (EGRESS_TRUST measured, not moved) #1022 and llm codec tests: F25 REHOME rows 107 and 109 at 1.5.5 (stacked on #857) #884.
• contract: the cold lane's SDK side is deleted; abi/cold 272 -> 43 (row 84) #904 and row 84: memory-ABI connection shapes move out of abi/hot into abi/shared.rs (hot 6224 -> 5965) #898 also move the hot freeze row, and contract-stateless: the plugin-logging capture is the one sanctioned per-thread slot; a second stays red #944 has adjacent contract-stateless rows. Whoever lands second settles those counts (rule 13).
• Draft row 155: the codec's entropy and wall clock leave the contract; a codec is handed a CodecHost (stacked on #974) #1028 (same lane) overlaps the codec/handle/relay hunks and rebases after.

Nothing was run locally (cargo fmt only); ModernCI is the proof.