fix(consolidation): guard observation_history append on 0-row observation UPDATE

[nicoloboschi]

Summary

When consolidation updates an existing observation, _execute_update_action
runs source-liveness checks (a non-locking preflight plus a FOR SHARE
recheck) — but those guard only the source memories. The observation row
itself (UPDATE memory_units ... WHERE id = $5) can still be concurrently
invalidated/deleted, so the UPDATE matches 0 rows. The code then fell
through to _append_observation_history, whose INSERT carries an
observation_id FK onto memory_units. Appending history for a now-missing
row raises ForeignKeyViolationError — a (correctly) non-retryable integrity
failure that marks the whole consolidation operation failed for a row that
simply no longer exists.

This captures the UPDATE command status in the SQL branch and bails out
(return None) before writing history when 0 rows matched. The
store/upsert branch is an upsert and cannot hit the 0-row case, so it needs no
guard. The Oracle DatabaseConnection wrapper reshapes rowcount into the same
"UPDATE <n>" command-tag form, so the int(status.split()[-1]) parse is
dialect-safe (same idiom already used in config_resolver.py).

Relationship to #3094

This supersedes #3094, which diagnosed the same FK race correctly but:

• added the guard twice — once inside the SQL branch and again,
  unconditionally, after the if/else; and
• the second copy references update_status, which is only assigned in the
  SQL branch. On the upsert/non-SQL path (writes_memory_rows_in_sql=False)
  that variable is unbound, so the block raises UnboundLocalError on every
  update through that path — turning a rare, correct-to-skip race into an
  unconditional crash. Its tests only drove the SQL branch, so this was not
  caught.

This PR keeps a single guard in the only branch that can 0-row, and is written
against the current main shape of _execute_update_action (pool-based,
with the preflight + FOR SHARE liveness checks that #3094's base predated).

Validation

Mock-level regression tests in tests/test_integrity_violation_not_retried.py:

• test_update_action_bails_when_observation_row_missing — UPDATE returns
  "UPDATE 0"; asserts the function returns None and _append_observation_history
  is not called (the FK-violating INSERT is skipped).
• test_update_action_writes_history_when_row_present — positive control:
  "UPDATE 1"; asserts history is appended and the embedding is returned.

Mutation-checked: reverting the guard turns the first test red (history append
fires) while the positive control stays green. ruff, ruff format, and ty
all pass.