v0.2.62

cli v0.2.62 — four defects a live run found in apprafter status (2.23h, unreleased)

The roll-up shipped in v0.2.60 was run against a real cluster for the
first time. Everything below was invisible to the unit tests, the gates
and the review that passed it.

Fixed

• Two healthy conditions were reported as problems. A clean cluster
  read 2 condition(s) not healthy, naming Ready=True (the parent
  Application is healthy) and UnauthorizedSourceModification=False
  (no foreign writer has been detected). Both readings were inverted,
  by one omission: the polarity table classified six condition types and
  the operator declares eight.

  The completeness test that existed to prevent exactly this could not:
  its list of six came from the same grep that produced the table, so it
  asserted that the build classified what the build already classified.
  It now reads the operator's own COND_* declarations, with a floor on
  the count so a renamed prefix cannot leave it judging an empty list.
  Reverting the fix makes it fail — checked.

• apprafter status printed the whole state store above its report.
  An info!(?state, …) inherited from the skeleton the command
  replaced, where the struct was the output. The smoke test now
  asserts the dump is absent.

• Node disk: could print forever from a poisoned cache. Freshness
  was now.saturating_sub(fetched_at), which is 0 for a stamp in the
  future — inside every TTL, so such an entry never expires. A machine
  here carried a stamp in the year 5138 and printed Node disk: warm
  above every command for two days, on a node with 69% free. A future
  stamp is now a miss.

• The tier read (unset) on a cluster that has one. State.tier is
  written by apprafter init; a cluster stood up the documented way —
  target add then up — never runs it. The header falls back to the
  target's configured tier, with the state file still winning when it
  has one: the target's is a default for the next provision, the state's
  is what was actually provisioned.