feat: opt-in last_exception redaction hook for the DLQ (F3-01)

[lesnik512]

What

PR 4 of the tail program. Closes F3-01: the DLQ stores repr(exc) verbatim, and for a poison message that repr can embed the request body / rejected row / credentials — durably, in the audit table.

Adds an opt-in last_exception_renderer: Callable[[BaseException], str | None] to OutboxBroker (and, for parity, the FastAPI OutboxRouter):

# Redact: store only the class name.
OutboxBroker(engine, outbox_table=t, dlq_table=dlq,
             last_exception_renderer=lambda exc: type(exc).__name__)

# Drop entirely:
OutboxBroker(..., last_exception_renderer=lambda exc: None)

• Default (None) keeps the full repr for forensic detail.
• The renderer's output is still length-capped at 8 KiB.
• Rendering centralized in _render_last_exception (replacing _truncate_exception), read from OutboxBrokerConfig.last_exception_renderer.

Backward-compatible: with no renderer configured, every existing path is byte-for-byte identical (the existing truncation tests still pass unchanged).

Tests

• test_fake_dlq_redacts_last_exception_with_renderer — class-name-only renderer; asserts the secret-bearing message is absent.
• test_fake_dlq_drops_last_exception_when_renderer_returns_none — None → stored None.
• Existing test_flush_terminal_dlq_payload_{truncates,short} cover the default repr path.
• just test → 516 passed, 100% coverage; just lint clean.

Docs

CLAUDE.md (DLQ invariants) and docs/usage/dlq.md (new "Redacting last_exception" section) updated.

🤖 Generated with Claude Code