fix(retain): retry and fail loudly on schema-drifted fact extraction (#3708) (#3760)

A model without strict-schema support can return well-formed JSON whose fact
objects have the wrong shape (no `what`/`factual_core`/`text`). Every such fact
was skipped with only a log line, and — unlike the two sibling skip paths — the
`missing 'what'` path did not set `has_malformed_facts`, so the re-prompt loop
never ran. A response whose facts were all drifted returned `[]` on the first
attempt, the document was committed with 0 memory units, and the operation
reported `completed`: silent data loss.

Streaming path: count an absent text key as malformed so the existing re-prompt
runs, and raise once retries are exhausted and nothing usable came back — the
same rule the non-dict response already follows (#1833), so the worker retries
and ultimately fails the operation instead of committing an empty document.

Batch path: it cannot re-prompt a single request, so record the drop in
`extraction_errors` (surfaced in the operation's result_metadata, and
escalatable via HINDSIGHT_API_FAIL_ON_EXTRACTION_ERRORS) instead of dropping it
silently.

A `what` that is *present* but empty or "N/A" is the model saying "nothing to
extract here" — that stays a quiet skip in both paths, as does `"facts": []`,
so content that genuinely holds no facts does not become a hard failure.