feat(redaction): add event-level redact_event hook

[naji247]

Summary

Redaction decisions often need more context than a single string. The existing redact_sensitive_information hook is applied blindly to every string field in an event, so consumers can't tell which tool produced a value, what type of event is being published, or which field a string belongs to.

This PR adds a new event-level redaction hook, redact_event, alongside the existing string-level hook. It receives the full Event object — resource_name, event_type, parameters, response, user_intent, and the rest — and returns a modified event, or None to drop the event entirely.

from agentcat import AgentCatOptions

def redact_event(event):
    # Drop events from tools that handle secrets entirely
    if event.resource_name == "get_credentials":
        return None
    # Strip response payloads from a specific tool
    if event.resource_name == "export_report":
        modified = event.model_copy()
        modified.response = None
        return modified
    return event

agentcat.track(server, "proj_0000000", AgentCatOptions(redact_event=redact_event))

Behavior

• Signature: Callable[[Event], Event | None | Awaitable[Event | None]] — sync or async, exported as EventRedactionFunction. Event is now also exported from the package root.
• Ordering: redact_event runs first in the queue pipeline and sees raw, unredacted values with full event context. redact_sensitive_information (if configured), sanitization, and truncation all run on its output, so the string-level scrubber remains the last line of defense.
• Drop semantics: returning None drops the event. If the hook raises, the event is dropped and the error is logged to ~/agentcat.log — the same fail-closed semantics as the existing string-level hook.
• System-managed fields: id, session_id, project_id, event_type, and timestamp are restored if the hook changes or removes them, protecting session and project attribution. Everything else is fully consumer-controlled, and fields the hook clears stay cleared.

Behavior note worth knowing

The string-level redact_sensitive_information hook currently never fires on the live publish path — the recursive walker doesn't descend into Pydantic event models (tracked by the xfail suite in tests/e2e/official/test_redaction_http.py). The new event-level hook operates on the Pydantic model directly and works on the live path today, so it also serves as a working redaction mechanism until that bug is fixed. The structural ordering (event hook first, string hook second) is preserved for when it is.

Test plan

• Unit tests for the new apply_event_redaction helper: sync/async hooks, None drops, cleared-field honoring, system-managed field restoration, internal function fields hidden from the hook with redaction_fn preserved, error propagation, same-object mutation for pipeline observers (tests/test_redaction.py)
• Queue-level tests: hook attachment from track() options via publish_event, hook rewrite published, drop-on-None, drop-on-raise, ordering relative to string redaction (tests/test_event_queue.py)
• End-to-end tests over a real Streamable-HTTP transport (not xfail — the event hook works on the live path): metadata visibility and field rewrite by resource_name, drop-on-None with other events still publishing, and a raising hook dropping only the affected event while the server keeps serving (tests/e2e/official/test_redact_event_http.py)
• Full suite green: 432 passed, 96 skipped, 3 xfailed; ruff and mypy --strict at their pre-existing baselines with no new findings