Log the cause when a capture recovery settlement fails

[ppXD]

Summary

• AgentRunLogCaptureRecoveryService.RecoverClaimAsync turned any exception from SettleAsync into a lost lease and logged nothing. A guard RAISE (P0001), a unique violation, a dropped connection and a settlement timeout all left the claim leased and idle until its lease expired, with no trace of why. That is why the retry-clause refusal fixed in Check capture retries against the settling transaction's start #2057 looked like an unreproducible flake.
• The catch now logs a Warning with the run id, the intent id, and the state and last_error_code the settlement was writing. When a PostgresException is in the exception chain, it also logs that exception's SqlState and MessageText. For any other cause both are null, and the attached exception carries the cause. The warning says the claim waits out its lease before a later wave re-claims it. Before it writes, SettleAsync can replace the observed outcome with Superseded or recovery-exhausted, so it records the outcome it writes in a SettlementAttempt. The log therefore names the write the guard actually judged.
• The outcome does not change. A settlement that raises is still counted as a lost lease and keeps its lease. No other typed outcome exists for an unwritten settlement, and retrying inside the wave would only repeat a deterministic refusal.
• Left as is: the lease re-check and DbUpdateConcurrencyException exits in SettleAsync stay unlogged. They are real lease losses, or impossible under the row lock. The RecoverAsync catch still drops its cause, but it settles as a typed recovery-operation-exception retry.

Test plan

• A_settlement_the_database_refuses_is_logged_with_the_write_it_refused_and_still_waits_out_its_lease, a Theory with 2 rows. RefusedCaptureSettlementInterceptor makes the guard refuse only the owned intent's settlement write, retry or terminal, and a due neighbour sits ahead of that intent. The test asserts:
  • a Warning, with RunId set
  • Outcome/OutcomeCode of the refused write: SourceFinalized/complete-backend-unavailable, or ExternalStateIndeterminate/recovery-exhausted with MaxAttempts = 1
  • SqlState P0001 and a MessageText naming the intent
  • LostLease >= 1 and the claim still leased
  • the neighbour settled and not logged
• A_settlement_that_outlives_its_budget_is_logged_without_a_database_cause_and_still_waits_out_its_lease: SlowCaptureSettlementInterceptor holds the owned run's settlement read until the settlement budget cancels it. The test asserts SqlState and MessageText null, an OperationCanceledException, and the same lost-lease outcome.
• Mutations, each red:
  • log call removed: all 3 tests
  • refusal!.SqlState: the budget test, with a NullReferenceException escaping ReconcileAsync
  • logging the observed outcome instead of the written one: the exhausted row
  • dropping attempt.Outcome = settled: the exhausted row
  • refusal seam no longer scoped to the owned intent: both rows, on the neighbour
• Integration: AgentRunLog* + *LogCapture* family, 112 of 112. The recovery flow class 5 runs in a row, 24 of 24 each time. Each new test alone in a cold process, 3 times each.
• Unit: full suite, 11588 of 11589 (1 skipped). dotnet build CodeSpace.sln: 0 errors.

Depends on #2057.