ci: impute a weight for new PostgreSQL tests instead of refusing them - #821

#815 packs the PostgreSQL shards by measured duration and fails closed on
any workflow entry with no `measured_seconds`. Weights are harvested from
`cargo-postgres-timing:` lines in a FINISHED CI log, so a test that has
never run cannot have a measurement -- and it cannot run until its PR is
green. The guard therefore made every new PostgreSQL test unmergeable.
#802 is the first one to hit it:

  1 workflow entr(ies) have no measured_seconds; regenerate weights
  before trusting the balance:
    ontology-canonical-adapter-postgres-employment-reassign-as-runtime-role-pg

That guard was written for a different failure: a map that has DRIFTED,
where entries silently lost weights they once had. Both look identical to
a per-entry check, but they differ by scale -- a PR adds one or two, a bad
regeneration drops dozens -- so the distinction is now made on share.

  - Unmeasured entries are imputed at the mean of their OWN package where
    it has a measured sibling (a new suite usually resembles the suite it
    joins), and the global mean otherwise. Never 0: a free entry lands in
    the lightest bin, which is exactly where a heavy newcomer hurts most,
    and treating unmeasured as free is how the entry-count scheme drifted
    in the first place.
  - Above MAX_UNMEASURED_SHARE (10%) the map is treated as drifted and
    still fails closed. At the observed 209 entries that allows 20 -- far
    more than any PR adds, far fewer than a regeneration that lost its
    timings.
  - A map with NO measurement anywhere fails closed separately: there is
    no basis to impute from, and imputing from nothing would invent a
    balance nobody measured.
  - Imputation bases are computed over measured entries only, so an
    imputed value never feeds back into the mean used to weigh the next
    one.
  - The CLI names every imputed entry and the total seconds it added, so
    an estimate is visible and gets replaced at the next timings harvest
    rather than hardening into a fake measurement.

Measured on #802's exact case: imputed 35.4s from its package mean,
partition ok, 5 shards, spread 1.00x, exit 0.

Mutation-proven on the real 209-entry map:
  strip 15% of measurements -> exit 1, "14.8%, limit 10%, the map has
    drifted";
  strip every measurement   -> exit 1, "no basis to impute from";
  one new entry (the #802 shape) -> exit 0, imputed and named.

26 partition tests pass; 12 gates swept, 0 failed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>