smpp-js/AGENTS.md

# AGENTS.md

Guidance for LLM agents working in this repository. Human-facing documentation lives in
[README.md](README.md); the remaining work is tracked in [todo.md](todo.md).

## What this is

A ground-up TypeScript rewrite of `larvitsmpp` 0.4.0, published as `@larvit/smpp` 1.0.0. The branch
started from an orphan commit — no history from 0.4.0 is carried over. The 0.4.0 source is still
readable on the `master` branch of the same repository and is the reference for protocol behaviour,
not for structure or style.

The library's value is its very small API. Do not grow the public surface without being asked.

## Hard rules

These are not preferences. Breaking one is a defect.

1. **Nothing throws.** Every fallible function returns (or resolves to) a DTO carrying an optional
   `err`. No `throw`, no rejected promises, no exceptions as control flow. Node APIs that throw are
   wrapped at the boundary and converted into a result. Programmer errors (bad arguments) are
   results too.
2. **Log messages are static strings.** Every dynamic value goes into the log metadata. Never
   interpolate, never concatenate.
   - GOOD: `log.debug('sendSms() - splitting message', { parts: msgs.length, to });`
   - BANNED: `log.debug('sendSms() - splitting into ' + msgs.length + ' parts');`
3. **No `error` event.** Node makes an unhandled `error` event throw, which would break rule 1.
   Sessions emit `sessionError`, servers emit `serverError`.
4. **No casts, no non-null assertions.** `as`, `as unknown as` and `!` are all banned. Parse untyped
   input once through a type guard at the boundary; everything past it is typed. `noUncheckedIndexedAccess`
   is on, so every lookup into a record or buffer is `T | undefined` until you handle it — that is the
   point, not an obstacle to route around.

## Architecture

```
src/
	index.ts             Public surface. Named exports only, no default export.
	client.ts            client() -> { err, session }
	server.ts            server() -> { err, server }, server owns the listener + close()
	session.ts           Session: dispatch, events, and the collaborators below
	sms.ts               The live handle emitted as the 'sms' event (sendResp/sendDlr)
	dlr.ts               Delivery receipts: text and TLV parsing, receipt status codes
	dlr-merger.ts        DlrMerger: per-segment receipts counted into one MessageDlr
	error-from.ts        errorFrom(): whatever was thrown or rejected, as an Error
	expiring-groups.ts   ExpiringGroups: the capped, expiring store both of those share
	incoming-requests.ts Every request the peer sends: messages, receipts, links, unknown commands
	link-gate.ts         LinkGate: where a request with no link to go out on waits for the next one
	link-timers.ts       LinkTimers: the enquire_link heartbeat and the idle timeout
	log.ts               SmppLog, the logger contract, and silentLog — the default
	message.ts           Encoding detection, splitting, bit counting, SMPP date formatting
	pdu.ts               pduToObj / objToPdu / pduReturn — synchronous, result-returning
	pdu-framer.ts        PduFramer: a byte stream cut into complete PDUs
	pdu-transport.ts     PduTransport: the socket a session reads complete PDUs off
	pending-requests.ts  PendingRequests: sequence numbers, correlation, timeout, abort
	reassembly.ts        Reassembler: capped, expiring multipart groups
	reconnect-loop.ts    ReconnectLoop: backoff, retry timer, stopped-ness
	result.ts            Result<T> — the shape every fallible call returns
	send-sms.ts          submitSms composition and the submitSmParams builder
	send-window.ts       SendWindow: the maxOutstanding semaphore
	session-options.ts   SessionOptions, ReconnectOptions, bind direction and the session defaults
	sms-id.ts            The notation a peer writes message ids in, normalised for comparison
	udh.ts               User data header: the concatenation fields of a long SMS
	uuid.ts              uuidv7() — the ids the library generates for messages
	defs/
		commands.ts      The 33 commands, their ids and ordered parameter lists
		constants.ts     consts + constsById, and the SMPP version constants
		encodings.ts     GSM 03.38, LATIN1, UCS2, detection, data_coding resolution
		errors.ts        errors + errorsById (ESME_*)
		tlvs.ts          TLV definitions, tlvsById
		types.ts         Wire types: int8/int16/int32/string/cstring/buffer/arrays
```

Dependency direction is one way: `defs` knows nothing above it, `pdu` uses `defs`, `session` uses
`pdu`, and `client`/`server` use `session`. Nothing reaches back up.

**Parameter order is wire order.** The key order inside `cmds.*.params` is the order the fields are
written to and read from the buffer. Never sort those alphabetically — the alphabetical-ordering
convention applies everywhere else, but here it corrupts every PDU.

## Toolchain

Run everything through the container; never invoke node or npm on the host.

```bash
docker compose run --rm node npm install
docker compose run --rm node npm test
docker compose run --rm node npm run build
```

- Tests are `.ts` and run directly under Node's type stripping — no build step in the dev loop.
- Source imports use `.ts` extensions; `rewriteRelativeImportExtensions` emits `.js` into `dist`.
- `erasableSyntaxOnly` is on, so no enums, no namespaces, no parameter properties. Use `as const`
  objects plus union types.
- The published floor is Node 18, but the dev container runs Node 24 (type stripping needs it). CI
  compiles the tests and runs them on 18/20/22/24, so the floor is verified rather than asserted.
- `typescript` is pinned to the 6.x line because `typescript-eslint` peer-requires `<6.1.0`. Move to
  TypeScript 7 once that constraint lifts.

## Defects found in 0.4.0

Confirmed by reading the 0.4.0 source. The rewrite fixes all of them; each needs a regression test
naming the behaviour, and the wire-affecting ones are cross-checked against a reference
implementation (see todo.md).

| Defect | 0.4.0 behaviour |
| --- | --- |
| LATIN1 never decodes | `decodeMsg` loops `consts.ENCODING` without breaking, so `data_coding` 0x03 lands on the alias `ISO_8859_1`, which has no decoder, and silently falls back to ASCII |
| Short segments | `splitMsg` accumulates a full segment then pushes `msgPart.slice(0, -1)`, so every segment is one character short: 152 GSM characters instead of 153, 66 UCS2 instead of 67. Long messages are split into more segments than they need, and each extra segment is billed |
| DLR month off by one | `smppDate()` uses `getMonth()` (0-based) without `+1`, so January renders as `00` |
| Non-standard DLR status | Receipts emit `stat:UNDELIVERABLE`; the spec's field is 7 characters (`UNDELIV`) |
| Flash destroys UCS2 | `flash: true` overwrites `data_coding` with 0x10, discarding the UCS2 alphabet, which needs 0x18 |
| Shared concat reference | The concatenation reference counter is a module-level global shared by every session in the process |
| `send()` never times out | Each call adds a listener keyed on the sequence number; a peer that never answers leaks it and the promise never settles |
| `tls: true` is not TLS | Constructs a bare `new tls.Socket()` with no handshake instead of `tls.connect()` |
| Alphanumeric sender TON | `sendSms` hardcodes `source_addr_ton` to 1 (international) even for alphanumeric senders, which require TON 5 |
| Text-only DLRs refused | `deliver_sm` without both `message_state` and `receipted_message_id` TLVs is rejected with `ESME_RINVTLVSTREAM`, so Kannel-style receipts are unusable |
| Unbounded reassembly | Incomplete long-SMS groups are capped by nothing and swept only when other traffic arrives, after 24 hours |
| Dead DLR aggregation | `longSmsDlrs` is allocated to merge per-segment receipts and then never used |
| Trailing NULL truncation | `types.buffer.size()` subtracts one whenever the value's last octet is `0x00`, so the PDU is allocated one octet short while `sm_length` still reports the full length. Any UCS2 message ending in a character like U+4E00 or U+3000 goes out corrupt |
| Dormant filters | `defs.filters` is declared on commands and TLVs but never invoked anywhere. Dropped in the rewrite; SMPP time formatting is exported as `smppTime` instead |
| Unchecked reads | Wire reads index straight into the buffer, so a short or malformed PDU throws out of the codec. Reads are bounds-checked and return results now |
| Unrangechecked writes | Integer params are handed to `writeUInt8`/`writeUInt16BE` unvalidated, so an out-of-range value throws from inside Node |
| `submit_multi` missing `sm_length` | The field is commented out of the command table, so `short_message` never round-trips for that command |
| Per-parameter defaults never applied | `calcCmdLength` reads `paramType.default` (the wire type's) rather than the parameter's, so `interface_version: 0x50` on the bind commands did nothing and every bind declared version 0x00 |
| `source_telematics_id` width | Defined as a 2-octet integer; SMPP 3.4 5.3.2.8 makes it 1 octet, unlike `dest_telematics_id`, which really is 2 |
| Binary payloads decoded as text | `data_coding` 0x02, 0x04, 0x14 and 0xF4-0xF7 are 8-bit binary and land on the GSM 03.38 table, which rewrites every octet outside it. They resolve to LATIN1 now, so the payload survives as bytes |
| `ESME_RINVBCASTCHANIND` typo | Defined as `0x011`, three hex digits; the spec value is `0x0112` |

## Multipart sends and the send window

`sendSms` puts every segment of a message on the wire together instead of waiting for each response
in turn. This is not an optimisation: this library's own server holds segments until the whole
message is reassembled before it answers any of them, so sending them one-after-a-response
deadlocks. It follows that a message with more segments than `maxOutstanding` cannot be delivered to
a server that defers responses that way — real SMSCs answer each `submit_sm` immediately, so this
only bites when both ends are this library.

## GSM 7-bit is sent unpacked

Over SMPP the ESME puts one GSM character per octet in `short_message` and the SMSC packs it into
septets. The 140-octet limit applies to that packed result, not to what goes on the wire here, which
is why a concatenated segment is 153 characters plus a 6-octet UDH — 159 octets in `short_message`,
and entirely correct. Do not "fix" this to 134; that number is the packed payload size and would
truncate every long message by a fifth.

UCS2 is not packed, so there the two coincide: 67 characters = 134 octets, plus the 6-octet UDH is
exactly 140.

## Conventions

- Hard tabs. Alphabetical ordering for keys, imports and lists unless order is logic-significant.
  Two deliberate exceptions: command parameters are in wire order (above), and the `errors` and TLV
  tables are ordered by their numeric id so they can be diffed against the spec and gaps stay visible.
- Comments are the exception, not the default — see the root `CLAUDE.md` rules. Do not write file
  preambles or restate what the code says.
- Test data uses real randomised UUID v7 values, never `aaaa-0000` placeholders.
- `message_id` values the library generates are UUID v7.
- A test that needs a dummy peer must `resume()` its sockets. An unread socket never processes the
  peer's FIN, so `server.close()` hangs forever — that is a test bug, not a library one.
- Everything a test opens gets its teardown registered as it is opened, never closed on the test's
  last line: an assertion that throws skips that line, and the listener it leaves behind keeps
  `node --test` alive until CI's ten-minute cap. `test/teardown.ts` covers a session, a server and a
  listener; anything else takes a bare `t.after`. Its close aborts rather than drains, so a test that
  fails holding the send window still ends.
- `t.after` hooks run in registration order, so registering at creation tears the outermost resource
  down first. A teardown that waits on a listener must destroy that listener's own connections before
  it waits, or be registered after the hook that does — `net.Server.close()` does not call back until
  every connection on it is gone.
- `assert.equal` from `node:assert/strict` narrows its first argument, so a following `?.` on the
  same value is flagged as unnecessary. Assert once with `assert.ok(x)` and use plain access after.

## Decisions

- **A close arriving after our own `unbind` is a clean unbind, not an error.** Maintainer's call,
  2026-08-26: most SMSCs drop the socket instead of answering, so the documented shutdown would
  otherwise always report a failure. It does mask a socket that died mid-unbind for an unrelated
  reason, which is accepted — the peer sees the same TCP close either way.
- **The published surface is frozen at what `src/index.ts` exports today.** `Session` is exported and
  publicly constructible, which is why `SessionOptions` and `ReconnectOptions` are public too — that
  is correct, not a leak, and it has been raised twice. The collaborators `session.ts` delegates to
  (`Reassembler`, `PendingRequests`, `SendWindow`, `ReconnectLoop`, `LinkTimers`, `LinkGate`,
  `DlrMerger`, `submitSms`) stay unpublished so they can be reshaped.
- **The sub-3.4 optional-parameter rule is a predicate, not a chokepoint.** `acceptsOptionalParams()`
  is consulted by the library's own senders; `session.send({ tlvs })` is passed through as written,
  because silently stripping a caller's explicit TLVs off a deliberately public low-level surface
  would be worse than sending them. The guarantee is "what this library sends honours the rule",
  never "the session cannot send optional parameters to an old peer".
- **Both ends feed `peerInterfaceVersion`, and a peer that declared nothing is pre-3.4.**
  `acceptBind()` records what the ESME declared in its bind request; the client's `bind()` records
  the `sc_interface_version` the SMSC answered with. A peer that declared no version is recorded as
  `undeclaredInterfaceVersion` (0x00) and is sent no optional parameters — the spec reads an absent
  `sc_interface_version` as an SMSC that supports none. `undefined` is left to mean one thing only:
  no bind has been accepted on this session yet.
- **The library speaks SMPP 3.4 on the wire, and `defs/` keeps the 5.0 tables as a superset.**
  Maintainer's call, 2026-08-26: 3.4 is what SMSCs actually run, while the wider tables let the codec
  parse and build whatever a peer sends. The declared version is an option on both `client()` and
  `server()`, so an implementation that needs 5.0 throughout can have it. The threshold at or above
  which a peer may be sent optional parameters is fixed at 0x34 by the spec and is not the same
  constant as the declared version.
- **Bind direction is enforced on the library's own senders and on everything incoming, not on
  `send()`.** A receiver-bound ESME carries no `submit_sm` and a transmitter-bound one no
  `deliver_sm`; `sendSms()` and `sendDlr()` refuse locally, and an arriving PDU is answered
  `ESME_RINVBNDSTS`. `bindAllows()` is a predicate on the same footing as `acceptsOptionalParams()`,
  so the deliberately public low-level `send()` stays a passthrough. Only those two commands are
  policed, because they are the only ones the library sends and dispatches by direction.

- **The logger is a five-method contract this library declares, not a dependency.** `SmppLog` in
  `log.ts` is what the code actually calls (`debug`, `error`, `info`, `verbose`, `warn`), so an
  application can satisfy it with an object literal and `@larvit/smpp` ships with no runtime
  dependencies. `@larvit/log` implements it structurally and stays a devDependency, where
  `test/tls.test.ts` passing a real `Log` as the server's logger keeps that compatibility compiled.

- **`esm_class` decides what a `deliver_sm` is, and the body is only read when it names nothing.**
  Message type `MC_DELIVERY_RECEIPT` (0x04) makes it a receipt whatever the body parses to, so a
  receipt in a format `dlrFromPdu()` cannot read reaches `dlr` with `smsId` undefined instead of
  arriving as an inbound SMS. Any other named type — delivery or user acknowledgement, conversation
  abort, intermediate notification — is not a receipt and its body is not scraped. A message type of
  0, or one of the ten the spec reserves, keeps the scrape: SMSCs that send text-only receipts leave
  `esm_class` at 0, and reading that as the spec's "default message type" would lose every one of
  them. A non-empty `receipted_message_id` TLV marks a receipt on the same footing there, since
  nothing but a receipt carries one. What gets scraped is the decoded `short_message` with any UDH
  stripped; a receipt this library recognises never reaches the reassembler, so an SMSC that splits one across segments gets a
  `dlr` per segment rather than one merged report.
  The `message_state` TLV is authoritative only where it names a state in the table — SMPP reserves
  0x80-0xFF for MC-vendor-specific values, so an unnameable one keeps its raw `statusId` and leaves
  `statusMsg` to the body.

- **A listener that rejects is routed by Node's `captureRejections`, not by hand-dispatching.** Both
  emitters construct with `captureRejections: true` and implement
  `[EventEmitter.captureRejectionSymbol]`, which lands a rejected `async` listener on `sessionError`
  or `serverError` beside the synchronous guard in `emit()`. Dispatching `rawListeners()` from
  `emit()` instead needs a cast to call them with the event's argument tuple, which hard rule 4
  forbids. A rejection reason is `unknown` and `String()` throws on a null-prototype object, so both
  handlers normalise through `errorFrom()` rather than inline — a route out of the handler would land
  on a bare `process.nextTick` with nothing to catch it.

- **Both emitters re-declare their listener methods to accept a promise.** Maintainer's call,
  2026-08-27: `EventEmitter` types every listener as void-returning, so the
  `session.on('sms', async sms => …)` the README documents reads as a misused promise in any strict
  consumer. `declare on: …` and its six siblings re-type the inherited methods to return `unknown`,
  which emits nothing, needs no cast and leaves the runtime method on the prototype. Overriding them
  as real methods instead cannot work: the `super.on()` call needs a cast to satisfy the conditional
  `Listener` type. The cost is that a subclass can no longer reach those seven through `super` or
  override them as methods — re-declaring them the same way is its way out. `unknown` rather than `void | Promise<void>` because a listener may return
  anything — `session.on('close', () => set.delete(session))` returns a boolean.

- **A reconnect keeps the delivery-receipt merges; everything else the link held is dropped.**
  `onDeliverSm()` answers each receipt before the group it belongs to is complete, and `teardown()`
  runs on every path — an idle timeout and a failed rebind, not only `close()` — so clearing the
  merges there loses receipts no peer has a reason to send again. They are cleared where the session
  is over instead: `close()`, or a drop with no reconnect loop left to bring the link back. Inbound
  segments stay in `teardown()`, because they go unanswered until the message is whole: the peer
  still holds them, and answering it on a later link with the old segments' sequence numbers would
  correlate with nothing. Surviving a process restart is a separate, public-surface question, and is
  in todo.md.

- **A message id base is merged at most once.** A receipt carries nothing but `<base>-<n>`, so a
  straggler for a message whose group is gone cannot be told from a receipt for a later message the
  peer handed the same ids — an SMSC whose id counter restarts with its process is the realistic
  case. `DlrMerger` remembers the bases it has finished with, capped and expiring exactly like the
  groups, and refuses to open one a second time: the later message gets no `messageDlr`, and an
  earlier one whose receipts are still arriving is dropped rather than left to collect the later
  one's. Every segment still reaches the application as a `dlr`. The rule covers the bases the
  merger opened — `expect()` ignores a lone id, so a single-part message never claims one.

- **A deliberate shutdown drains; an unusable link and an abort do not.** `close()` and `unbind()`
  refuse further sends and wait on the send window, not the pending map — the map misses a segment
  still queued behind a full window, and finishing a half-sent multipart message is the point. The
  window counts slots, never outcomes, so it says when to stop waiting and nothing about what
  happened: it empties on a drop too, where `teardown()` settles everything the link was carrying,
  which is why `drain()` reads `closed` before it reads the count. The wait covers what this end
  sent — a request the peer sent us is answered through `sendReturn()`, which never enters the
  window, so a server session waits for none of its inbound work; that half is in todo.md.
  `shutdownTimeout` bounds the drain alone, and `0` waits forever like every other timeout here;
  `unbind()` then waits `responseTimeout` for its own response, and sends that PDU through
  `request()` past both the window and the drain gate because it must go out either way. A stream
  the framer or the codec cannot read takes `teardown()` instead, and `close({ signal })` on an
  aborted signal and a peer's own `unbind` take `end()` — nothing on a dead link can answer, an abort
  means stop now, and a peer that has declared itself finished will not answer what it still owes us,
  so draining any of the three would only hold a socket open for the timeout. `SmppServer.close()`
  reports each session's unfinished drain through `serverError`, because its own result says
  nothing but that the listener stopped. `shutdownTimeout` stays a session option rather than a
  `close()` argument: `server()` builds sessions on the caller's behalf, so the option
  is the only composition point, and `close({ signal })` already covers a hard deadline.

- **The TLS tests build their own self-signed certificate in DER** (`test/tls.test.ts`) instead of
  adding a devDependency or shelling out to openssl. Maintainer's call, 2026-08-26: the dev image
  `node:24.18.0-bookworm-slim` ships no openssl binary, so a shelled-out fixture would pass in CI
  and fail on every developer machine, and a committed key leaks in a public repository. Valid while
  the dev image has no openssl.

- **`close` means the session is over, and a drop the loop will retry is `disconnected`.** Maintainer's
  call, 2026-08-31: with reconnect on by default a `close` on every transient drop left an application
  unable to tell a retry from the end, and no second one follows because `teardown()` is a no-op once
  `closed`. `teardown()` picks the event by whether the reconnect loop is still live, and `end()` stops
  that loop before tearing down, so every deliberate shutdown emits `close`. Without the split an
  application that opens a replacement client on `close` ends up holding two binds on one account.

- **Only a link that outlasted `maxDelay` resets the backoff.** Coming up is not proof it works: an
  unreadable stream is found after the bind returns, so resetting there gave a link that died on
  arrival a fresh `minDelay` every cycle — one TCP connect and bind per second, forever, which is how
  an account gets blocked for bind flooding. `ReconnectLoop` records when it brought the owner up and
  resets only if the link then lasted longer than the longest wait it would ever schedule. A drop
  after a healthy link still retries at `minDelay`.

- **A stream this library cannot read is a dead link, not a dead session.** Maintainer's call,
  2026-08-31: a framing or codec error tears the link down through `teardown()`, so the reconnect
  loop retries it on a fresh socket with a fresh framer — which is what a desynced stream needs, and
  the common cause. `sessionError` still carries every failure, so a peer that only ever sends
  garbage is visible in the log rather than silent, and the backoff grows to one attempt per
  `maxDelay`.

- **`disconnected` counts failed links, not outages.** A retry that opens a socket and then loses it
  clears `closed` through `attach()`, so the next `teardown()` emits again: each emission is a link
  that went down, which makes the event deliberately not one-to-one with `reconnected`. Suppressing
  the later ones would leave a failed rebind with nothing but a log line.

- **`session.sock` is a getter over `PduTransport`, and stays public.** Maintainer's call, 2026-08-31:
  `session.ts` had reached its line cap, so the socket-to-PDU seam todo.md named was opened —
  `PduTransport` owns the socket, the framer and the parse, and hands the session raw bytes, framed
  PDUs, parsed ones and an unreadable stream. Reading `session.sock` is unchanged; assigning it no
  longer compiles, which never rewired the handlers and so never worked. The transport stays
  unpublished like the other collaborators.

- **A client re-binds after a drop unless it is told not to.** Maintainer's call, 2026-08-31:
  surviving a dropped link is most of what the session layer is for, and behind an opt-in an
  application that never read the options table got none of it. `reconnect` takes
  `{ minDelay, maxDelay }` to retune the backoff and `false` to turn it off, so absent means on and
  there is one spelling for each. Only `client()` reconnects — a `server()` session is a connection
  the peer opened, and nothing at this end can reopen it. The retry timer is `unref()`'d, so a
  process with nothing else left to do still exits between attempts.

- **The notation a peer writes message ids in is named per place, and normalisation never reaches
  inside a `<base>-<n>` id.** An SMSC may answer `submit_sm_resp` in hex and write the receipt's
  `id:` in decimal, so one transform over both sides cannot make them equal — `smsIdFormat` names
  `receipt` and `submitResp` separately and reads both into a plain decimal value before `smsIds`
  and `dlr.smsId` are compared. `submitResp` covers the `receipted_message_id` TLV too, which SMPP
  3.4 5.3.2.26 defines as the id the `submit_sm_resp` carried: naming one notation for whichever id
  a receipt yields would break the peer that sends both, whose TLV correlated before the option was
  set. Omitting a place is what leaving it alone means, so there is no
  `raw` notation, and a caller-supplied formatter is refused because it would make the promise that
  those two are comparable unverifiable — `onRequest` and the PDU on the `dlr` event are the escape
  hatches, and the `onReceipt` hook in todo.md is the seam if one is wanted. An id no notation reads
  is left exactly as it arrived, which is what keeps `DlrMerger` working: a `<base>-<n>` id parses
  as no number and so reaches `expect()` and `collect()` unchanged. Normalising the base instead
  would break that pair. The option is on `client()` only — a `server()` session generates its own
  ids and writes its own receipts, so both places are already one notation.

- **A send waits for the next link only if it never reached the socket; one that did is counted, not
  resent.** Maintainer's call, 2026-09-01: re-queueing everything unanswered would resend a
  `submit_sm` the SMSC accepted and answered into a dead socket, which is delivered and billed
  twice, while a request that never left this process can be lost for free. `LinkGate` holds a send
  that has no link, so a send issued between links and a segment still queued behind a full window
  when the drop hit both go out on the new one. Once a request has been written, every way it can
  fail — the link dropping under it, `responseTimeout` expiring, the caller's own abort — means the
  peer may have taken it, so `attempt()` wraps all three in `UnansweredError` and `collectSent()`
  counts them into `SendSmsResult.unanswered`. Counting only the dropped-link case, as the first cut
  did, would have called the commonest one safe to resend. A count rather than a boolean because
  `sendSms()` aggregates segments into one `err` slot, and required rather than optional so every
  construction site answers. `UnansweredError` stays unexported: `unanswered` is the one spelling on
  the public surface, and a `send()` error that is neither a build failure nor a pre-write abort
  means the same thing.
  The hold is bounded by `responseTimeout` rather than an option of its own — that is already the
  answer to how long one request may wait. It bounds the hold and the answer separately, and the
  wait for a `maxOutstanding` slot is bounded by nothing, so `responseTimeout` is not a deadline for
  the call; `SendOptions.signal` with `AbortSignal.timeout()` is, and both the gate and
  `pending.wait()` honour it.

- **The gate decides whether a link can carry a request, and a bind is what makes it one.**
  Maintainer's call, 2026-09-01: `attach()` clears `closed` the moment a socket is handed over, one
  round trip before the bind is answered, so gating on `closed` let a send arriving in that window
  go out unbound and come back `ESME_RINVBNDSTS` while a send that arrived a millisecond earlier was
  held correctly. `LinkGate` owns the answer instead — `shut(returning)` on every teardown,
  `open()` only once `comeBackUp()` has a bound link — and `Session.linkDown()` reads it rather than
  `closed`. The bind itself cannot wait for what it creates, so `send()` lets the three bind
  commands past the gate and the window, the same door `unbind()` takes through `attempt()`. That
  keeps the exemption a predicate on the command, like the `_resp` guard beside it, rather than a
  second `send()` on the public surface or a changed `ReconnectOptions.onConnected`.
  The gate is told what happened and never reads back into the session: a collaborator that has to
  ask does not own its decision, which is how the first cut ended up answering the same question two
  different ways at admit and at release.