feat: Sentry SDK integration for in-process PII scrubbing

[sidmohan0]

Summary

Adds a third enforcement-point adapter alongside the Claude Code hook and LiteLLM guardrail: datafog.integrations.sentry, which scrubs PII from Sentry telemetry before it leaves the process.

Sentry's built-in client-side EventScrubber matches sensitive key names only and never inspects values — an email in an exception message, an SSN in a stack-frame local variable, or a card number in a breadcrumb goes straight through. Server-side scrubbing inspects content, but only after the data has already left the machine. This integration fills that gap with datafog's content-based regex engine.

Usage

import sentry_sdk
from datafog.integrations.sentry import DataFogSentryIntegration

sentry_sdk.init(dsn="...", integrations=[DataFogSentryIntegration()])

or manually: sentry_sdk.init(dsn="...", before_send=scrub_event).

Design

• Integration class + global event processor (the pattern used by sentry-dramatiq / pytest-sentry): covers error events and transactions (span descriptions/data), and leaves the single before_send slot free for the application. Config is resolved from the active client per event, so clients without the integration pass through untouched.
• Logs, trace metrics, and streamed spans bypass event processors, so setup_once_with_options composes scrubbing wrappers around the user's before_send_log / before_send_metric / _experiments["before_send_span"] callbacks: scrub first, then chain — the application's callback only ever sees scrubbed telemetry. Wrapping is idempotent and follows the SDK's own top-level-then-_experiments precedence.
• AI agent monitoring coverage: gen_ai.* prompt/completion span attributes (which Sentry documents as not covered by default server-side scrubbing) are scrubbed in transaction events, whether attached as nested message lists or JSON strings. Users can keep include_prompts on without shipping user PII.
• Same config surface as the other adapters: entity_types (default EMAIL, PHONE, CREDIT_CARD, SSN; noisy types opt-in), allowlist / allowlist_patterns, fail_policy (open = send unscanned on engine error so a scrubber bug never costs crash reports; closed = drop the payload).
• Never echoes matched values — logs and errors carry entity-type counts only.

Safety properties (each locked in by a regression test)

• Text beyond the scan caps is replaced with [UNSCANNED_N_CHARS] markers, never emitted raw — a PII value straddling a cap boundary would otherwise fail to match in the truncated chunk and reassemble unredacted.
• Tuples/sets are rebuilt as lists and scanned (walker parity with the Claude Code hook).
• Aliased containers are scanned once (no budget double-spend); cyclic structures terminate.
• The integration lookup in the global processor and telemetry wrappers fails open, so delivery never breaks.
• The _experiments copy of a wrapped callback is overwritten too, so no raw unscrubbed user callback stays reachable.
• Top-level machine identifiers (event_id, release, environment, sdk, ...) are never touched.

Packaging

pip install datafog[sentry] (any sentry-sdk 2.x; logs/metrics wrapping engages on SDK versions that support them). The module is not imported by datafog core; apps that already ship sentry-sdk need no extra.

Test plan

• 39 tests in tests/test_sentry_integration.py, 100% line coverage of the new module
• End-to-end through a real sentry_sdk.init client: capture_message, capture_exception with local-variable capture, allowlist, pass-through when the integration is absent, and Sentry Logs (enable_logs=True + sentry_sdk.logger) with before_send_log chaining
• Fail-policy tests: open returns payload on engine error, closed drops it, error logs never contain event content
• Telemetry wrapper tests: precedence (top-level shadows _experiments, no double invocation), idempotent re-wrap, lookup-failure pass-through, metric attribute scrubbing, experimental span wrapping
• gen_ai span attributes scrubbed as message lists and JSON strings
• Full suite: 660 passed; only pre-existing spacy-extra failures in test_text_service_integration.py (run in dedicated CI jobs)
• black / isort / flake8 clean
• CI matrix (install profiles, Python versions) — needs this PR's checks to run

Follow-ups (separate PRs)

• Attachment policy via hint["attachments"]
• Advanced Data Scrubbing rule-pack generator (server-side safety net for non-Python SDKs)
• Benchmark entry (µs per event) in benchmarks/