A failure's cause sentence comes from core; each app binds the remedy
Status: decided and built. The rule it serves is the failure-message bullet in CONTRIBUTING.md. See docs/notes/README.md.
A failure more than one app reports -- a partner that never arrived, a shared folder that is missing -- used to be worded at each raise site, with a flag name appended by whichever app had injected its guidance into core. The same cause read differently in the CLI and the console, and a core message could name a flag the run did not take.
Where the two sentences live
Core holds the catalog, packages/core/src/failureCause.ts: a FailureCause union whose members carry facts only (a channel, a path, an errno code, a wait in milliseconds), failureCauseSentence giving the one sentence stating what happened, and markFailureCause / failureCauseOf attaching the cause to an error as a property tag and reading it back through the cause chain. The tag leaves the error's class alone, so exit-code classification is unchanged. Core's sentence names no flag, control or command. The CLI draws on the catalog, and so does the console for a run its server conducted through the CLI: the CLI's terminal event states the cause as data (CLI_EVENTS.md), and the console shows core's sentence with its own remedy. The public web app's failure copy does not, and its WebRTC no-show wording lives in apps/web/src (useInviterExchange.ts).
Each app binds the remedy in a map keyed on every FailureCauseKind, so a cause core adds fails to compile in an app until it has a remedy. The console's is apps/web/src/console/consoleFailureCause.ts, naming the Connection tuning control that sets the wait and the console's shared-folder mount; the CLI's is apps/cli/src/failureRemedy.ts; it names only the flag that bounds the run in hand -- --peer-timeout on an exchange, --accept-timeout on an online invitation -- by reading a second tag the command boundary sets on a partner-never-arrived error. The remedy for an unreachable relay registrar is the same in both apps, since the registrar address is set in the configuration and neither app has a flag or control for it, so core holds it (relayRegistrarUnreachableRemedy) and both maps return it.
Formatting the facts
A wait is stated by formatWaitDuration in the largest whole unit that states it exactly, and every count goes through formatCount (packages/core/src/utils/formatCount.ts), whose explicit locale keeps the digit separator ASCII for the CLI's console check. An OS error code goes in parentheses after the path or host it concerns. A fact the raise site cannot know -- a channel or a wait when the connection holds no configuration -- is left out of the sentence rather than guessed. Internal tags such as a [role] prefix stay in debug logs.