Summary

Adds bdata scraper heal <collector_id> "<prompt>" — AI self-healing for scrapers. When a scraper drifts (selectors move, a page redesigns), the agent fixes it in place instead of rebuilding, so the saved collector_id keeps working and improves.

• One new command, scraper heal — the maintenance twin of scraper create: POST /dca/collectors/{id}/refactor_template → poll refactor_template/progress, reusing the existing async trigger→poll machinery (poll_until, build_ai_trigger_retry, SCRAPER_BODY_HINTS, the 429 backoff).
• The agent is the detector. The CLI never guesses a scraper is "broken" (a heal is slow/billable/mutating). The agent inspects run output and decides. So run stays read-only — there is no --heal flag on run.
• Closes the loop. On success heal emits a {collector_id, status, completed_steps, prompt, view_url, next_step} envelope (same shape as create), where next_step is a ready-to-run bdata scraper run <id> <url> verify command (--url bakes the real URL in). The intended agent flow is: run → inspect → heal → re-run → verify.
• Failure is non-destructive. A failed heal leaves the existing scraper unchanged and still working; the recovery note says so (distinct from create's "half-built collector" wording).
• Required <prompt> (≤1000 chars, validated fast); carries over --timeout, --max-retries/--no-retry, -o/--json/--pretty/--legacy-output/--timing/-k.

Built test-first; 25 new tests. Design + plan in docs/superpowers/.

Test Plan

• pnpm type-check clean
• pnpm build clean
• src/__tests__/commands/scraper.test.ts — 131/131 pass (incl. validation, all failure paths, retry forwarding, legacy-output, progress-endpoint URL, exit codes, subcommand wiring)
• bdata scraper heal --help shows args/options/examples; bdata scraper --help lists heal
• bdata scraper run --help has no --heal option (design invariant)
• Live run against a real collector (reviewer, with credentials)

Note: the repo's full suite has 8 pre-existing failures in browser/daemon/discover/scrape tests that also fail on main (0.3.0) — unrelated to this change. Agent-facing docs are in a companion PR on the brightdata-plugin repo (scraper-studio skill: Action 3, api-flow, recipe, common-mistake).