feat(telemetry): give every failure an error code and report where a run stopped

[djgould]

clerk init fails on 21.4% of runs (1,596 runs / 30 days), and the telemetry can't say why or where. This makes both legible.

• Error codes on all 87 CliError sites. 19 were uncoded and collapsing into a generic cli_error — 125 of 342 init errors (36.5%). Adds 14 ERROR_CODE entries across init, update, deploy, env, users, and switch-env. Agent-mode JSON gates on error.code, so these failures now emit structured JSON where they previously fell through to plain text.
• A stage dimension recording how far a multi-step command got, emitted on success, error, and abort — a drop-off funnel, not an error-only field. clerk init is instrumented across ten markers: flags → detect → strategy → link → bootstrap → install → scaffold → keys → skills → done, plus already_set_up for the clean-rerun short-circuit.
• Registry failures split by cause. clerk update --channel typo reaches npm fine but reported registry_unreachable, so agents retried a permanent input error. Now usage_error for a missing dist-tag, update_failed for a bad response, registry_unreachable only for transport/timeout.

Disclosure notice and README both updated — this collects a new field.

What it does not fix

The 107 unexpected_error runs (31.3% of init errors) are plain Error throws that never reach CliError; stage will say where they die but the error stays anonymous. Capturing error.constructor.name would decompose them — deliberately left out of this PR.

Review

Carries a prior multi-model adversarial review (Claude + Codex, 19 agents, 10 findings, 0 refuted). All 8 distinct findings are applied, with one deviation: finding 7 proposed deleting the dead users/shared.ts; this takes the leaner option the reviewer offered — dropping the NOT_IMPLEMENTED code so the PR doesn't mint a public error code for dead code, leaving the out-of-scope file alone.

The three new stage tests were mutation-checked: each fails when its marker is removed.

Verification

typecheck, lint, format:check clean; 2673 pass / 0 fail. E2E not run (needs 1Password-injected secrets).

🤖 Generated with Claude Code

[changeset-bot]

🦋 Changeset detected

Latest commit: 6e9f1d3

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
clerk	Minor

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR