[scttfrdmn]

Fixes #125

Stacked on #133 (#128). This PR targets main so CI runs, which means its diff includes #133's commit until that merges. Review only the top commit, Typed SpawnError rejections…. Merge #133 first. It is stacked because the inline error and the "stopping…" progress need cards that survive the monitor's re-renders, which is what #133 adds.

Client

• SpawnError (src/core/errors.ts). stop / start / hibernate / terminate / extend / signalComplete / launch / refresh reject with it. Fields: operation, instanceId, code, aws, httpStatus, requestId, cause.
  • With aws: true, code is the AWS code: UnauthorizedOperation, IncorrectInstanceState, UnsupportedHibernationConfiguration, ExpiredToken, and so on. It is read from SDK v3's name when $fault/$metadata are present, or from Code.
  • Otherwise code is one of spawn-ts's own: InvalidInstanceId, InstanceNotFound, AmbiguousName, NotSimulated, NoTtl, InvalidDuration, LaunchRefused, NetworkError, Unknown.
  • The message is unchanged, so existing toThrow(/…/) matches still hold.
• An action resolves only when its provider call succeeded. The refresh after it no longer turns a successful stop into a rejection; a failed refresh is reported instead.
• refresh() rejects on a failed DescribeInstances and never resolves []. The new client.listState is { status: "loading" | "ok" | "error", instances, checkedAtMs?, error? }. On error, instances is the last list that loaded.
• EC2Provider.list follows NextToken. Reading only one page was another way to show a wrong list as if it were current.
• No unhandled rejections. A failure nobody awaited (a monitor tick, or a post-action refresh) is emitted as { type: "error", error }. Before, void this.tick() left an unhandled rejection.
• launch types its errors at each throw point rather than wrapping the method. A wrapper adds a microtask per launch, and the fan-out pump's ordering is sensitive to that (one sweep test caught it).
• The post-launch-refresh-fails → terminate behaviour (Go #1298) is unchanged.

Dashboard

• A refused action shows on its card (.card-error, role="alert") as one plain sentence naming the action, from describeSpawnError, plus the code and AWS's message, with a dismiss link. It survives re-renders, is cleared by the next successful action, and is logged as an error line.
• While a call runs, the pressed button reads stopping… / starting… / terminating… and the card's buttons are disabled. Terminate confirms before anything runs.
• A failed refresh keeps the last cards under a .list-error banner: "…[ExpiredToken] Showing the list from 12:00:00; those instances may still be running."
• "no instances" appears only after a list actually loaded. Before the first one it says "checking instances…". After a failure with no prior list, it shows the banner and nothing else.
• .orphan-banner[hidden] (and the new elements) now actually hide. display: flex was overriding the attribute, as the brief noted.

Tests

• src/core/errors.test.ts:
  • code extraction; toSpawnError keeping the code, status, request id, cause and message; network and local codes; the plain sentences.
  • With a fake EC2: stop / hibernate / terminate rejecting with UnauthorizedOperation, UnsupportedHibernationConfiguration and IncorrectInstanceState. A Name rejects with InvalidInstanceId and makes no AWS call.
  • A successful stop with a failing refresh resolves and sets listState to error.
  • EC2Provider.list rejects instead of returning [], and paginates.
  • refresh rejects, keeps the last good list and marks it stale.
  • 8 monitor ticks over a failing account emit at least 8 error events and leave no unhandled rejection (process.on("unhandledRejection")).
• dashboard.test.ts:
  • a refused stop shows the reason and code on the card, logs it, re-enables the buttons, and leaves no unhandled rejection;
  • the error survives re-renders, and a later success clears it;
  • progress text shows while the call runs;
  • a failed refresh keeps the cards with the banner and never says "no instances";
  • a first refresh that fails shows the error, not an empty account.

Breaking for existing consumers

Mildly, and noted in the CHANGELOG:

• errors are SpawnError, though messages are unchanged and instanceof Error still holds;
• a stop that succeeded no longer rejects because its follow-up refresh failed;
• before the first refresh, the Dashboard says "checking instances…" instead of "no instances".