Loki Mode

Zero-friction adoption: fail-open preflight, honest failure diagnosis, CLI polish

First-run and error-diagnosis improvements so a new user gets unblocked fast and a
failed build explains itself, with every check fail-open and honest.

Added (first-run preflight, advisory):

• validate_api_keys now does a fast preflight before a build: git presence (a
  genuine hard requirement - the build inits a repo - so a missing git blocks with
  a copy-paste fix), plus ADVISORY node-version and provider-reachability checks
  that WARN and continue (never block a working user; the real call is the test).
  Opt out of the network probe with LOKI_SKIP_NET_PREFLIGHT=1.

Added (failure diagnosis):

• On any failed/empty iteration or a mid-run crash, the engine writes a classified
  .loki/state/LAST_ERROR.json ({iteration, error_class, brief, timestamp}),
  cleared at the start of every fresh run. loki why now surfaces it (text +
  --json) only when the current run actually failed - never beside a success.
• On any failed build exit, a one-line "run loki why" hint + a .loki/NEXT_STEPS.txt.
• Rate-limit waits now log the human wait/reset time (not a silent backoff) and
  write a best-effort .loki/signals/RATE_LIMITED record (forward-laid).

Added / changed (CLI + doctor):

• npm global-bin PATH guidance after a provider CLI install that isn't on PATH.
• Consistent first-run onboarding between bare loki and loki help; deprecated
  run demoted out of main help into loki run --help.
• Bash route now has the update-available nudge (parity with the Bun route),
  surfaced via loki doctor; friendlier unknown-command pointer + grouped
  "more commands" footer.
• loki doctor Python check reworded honestly: Python 3.12 is recommended for
  memory vector search; core memory and the rest of Loki work without it (WARN,
  not a hard fail).