fleet_dispatch: skip tiers that already failed identically, using ledger history + TieredEngine

[lorenzoliuzzo]

Context

Downstream of my-things-core#119 (the TieredEngine composition
primitive). Blocked on that landing first. This is the "consult the ledger
before choosing a tier" half of the original my-things-core#116 — it turns
out to need no new core surface, just a consumer-side pattern in the one
place that already has both attempt history and a retry loop:
fleet_dispatch.py.

Confirmed by reading the file: _last_attempt(ledger, candidate_id)
(fleet_dispatch.py:641-674) already filters kind="dispatch" ledger entries
for one candidate and returns an Attempt with final_message
(e.data.get("final_message")) and attempt_number. MAX_ATTEMPTS
(fleet_dispatch.py:694) already gates escalation to needs_human.

Proposed change

1. Extend the existing attempt-reading path (generalizing _last_attempt, not
   replacing it) to return the last N dispatch entries for a candidate,
   not just the latest — the data is already there, this is just not
   truncating to one.
2. Build a skip set for TieredEngine.run(): if the last N attempts'
   final_message values are identical for a given tier, add that tier name
   to skip on the next attempt for that candidate — the same "identical
   failure N times" signal my-fleet#11's harness-bug check uses, applied to
   tier selection instead of escalation.
3. Wire a TieredEngine into the dispatch worker in place of whatever single
   Engine it constructs today, with skip computed this way per candidate
   before each attempt.

Explicitly out of scope

• The harness-bug signature check itself and its escalation path — that's
  my-fleet#11, a related but separate signal (this issue skips a tier;
  fleet: distinguish 'the harness is broken' from 'the model keeps failing' on repeated identical failures #11 escalates the whole issue to a human). They can share the
  "last-N-identical" helper if one is written generically enough, but that's
  an implementation detail for whoever picks both up, not a reason to merge
  the issues.
• The real local-model tier itself (my-things-core#119's scope note
  applies here too).

Blocked-by: my-things-core#119. Relates to my-fleet#11 (shared
"identical failure" signal, different consumer).