- ProbeRepository.add_failed now takes (probe_id, message) to store
  error message strings, matching original ProbeManager behavior
- ProbeRepository methods return the actual hash (not .dup()) for
  backward compatibility with existing test assertions
- ProbeRepository RBS updated to reflect String values in failed_probes
- Remove argument from probe.disable!() call since the method doesn't
  accept parameters
- Update probe_notifier_worker tests to properly mock instance_double
  methods and use .send() interception

Co-Authored-By: Claude Opus 4.5 <[email protected]>

The failed_probes hash stores error message strings (not Probe objects),
so the RBS return type should be Hash[String, String].

Co-Authored-By: Claude Opus 4.5 <[email protected]>

The code is self-explanatory for disabling the probe and reporting status.

Co-Authored-By: Claude Opus 4.5 <[email protected]>

When a probe fails to install (e.g., due to instrumentation error or
already-installed error), report ERROR status to the agent so it shows
up in the UI. Previously, failures were only logged and tracked in
telemetry.

This addresses the TODO comment that was in the code.

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Address review comments:
- Remove "avoiding circular dependencies" from comments
- Document batch parameter in handle_snapshot_serialization_failure
- Clarify find_failed looks up by probe ID
- Clarify failed_probes stores error messages, not probe objects
- Explain when remove_failed is called
- Add yield/lock safety analysis for clear_all method
- Improve class-level documentation describing the three collections

- Fixed in lib/datadog/di/probe_repository.rb (class docs, find_failed, add_failed, remove_failed, clear_all)
- Fixed in lib/datadog/di/probe_notifier_worker.rb (handle_snapshot_serialization_failure docs)

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Address review comments:
- Clarify that yield only happens when block is given
- Move yield outside the lock to avoid blocking other operations

The probes are copied and cleared atomically under the lock, then
the cleanup callback is invoked outside the lock. This prevents
blocking if unhook is slow.

- Fixed in lib/datadog/di/probe_repository.rb:clear_all

Co-Authored-By: Claude Opus 4.5 <[email protected]>

The flush method blocks until queues are empty and IO is complete.
Error handling happens synchronously in do_send_snapshot, so flush
is sufficient - no additional sleep needed.

Verified tests pass 5x in a row without sleep (deterministic).

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Since DI is @api private, no backward compatibility is needed.
Remove installed_probes, pending_probes, failed_probes delegation
methods from ProbeManager - callers should access probe_repository
directly.

- Removed 3 delegation methods from lib/datadog/di/probe_manager.rb
- Updated RBS file
- Updated tests to use manager.probe_repository.* instead

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Fix integration tests to use probe_repository directly after removing
installed_probes, pending_probes, and failed_probes delegation methods
from ProbeManager.

- Updated spec/datadog/di/integration/instrumentation_spec.rb
- Updated spec/datadog/di/integration/everything_from_remote_config_spec.rb

Changed all occurrences:
- probe_manager.installed_probes → probe_manager.probe_repository.installed_probes
- probe_manager.pending_probes → probe_manager.probe_repository.pending_probes

Co-Authored-By: Claude Sonnet 4.5 <[email protected]>

…dling

Previously, when a batch of snapshots contained one that failed JSON
encoding, all probes in the batch were disabled. This was incorrect
because healthy probes were being disabled due to failures in unrelated
probes.

This refactors the design to:
1. Transport serializes each snapshot individually, isolating failures
2. Transport calls on_serialization_error callback when serialization fails
3. ProbeNotifierWorker provides callback that disables the affected probe
   and sends ERROR status to the backend
4. ProbeManager sends Hash objects (not JSON strings), maintaining clean
   separation of concerns

The callback-based design keeps serialization in the transport layer where
it belongs, while allowing the caller to handle failures appropriately.

- Transport.send_input now accepts on_serialization_error callback
- ProbeNotifierWorker.handle_serialization_error disables probe and sends
  ERROR status when serialization fails
- Component passes probe_repository and probe_notification_builder to
  worker for error handling
- ProbeManager simplified to just pass Hash payloads

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Add the on_serialization_error keyword parameter to the send_input
method signature in the RBS type definitions.

Co-Authored-By: Claude Opus 4.5 <[email protected]>

When binary data (ASCII-8BIT) is included in snapshot data, JSON.dump
raises Encoding::UndefinedConversionError rather than JSON::GeneratorError.
This can happen when custom serializers return binary strings.

Add Encoding::UndefinedConversionError to the rescue clause so that
encoding failures from binary data are properly caught and reported
via the on_serialization_error callback.

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Binary data in JSON encoding can raise either JSON::GeneratorError or
Encoding::UndefinedConversionError depending on Ruby version and how
JSON.dump handles the encoding. Update test expectations to accept
either exception type.

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Use a blanket rescue clause to catch all possible serialization errors,
not just JSON::GeneratorError and Encoding::UndefinedConversionError.
This handles edge cases like:
- Encoding::InvalidByteSequenceError
- TypeError (non-serializable objects like Proc, IO)
- SystemStackError (infinite recursion)
- Errors from custom to_json implementations

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Address review feedback for PR #5448:

- Make probe_repository and probe_notification_builder required keyword
  args in ProbeNotifierWorker (production always provides them)
- Remove nil guard clause in handle_serialization_error
- Add @return/@param/@raise YARD tags to ProbeRepository, ProbeManager,
  and ProbeNotifierWorker public methods
- Replace bare allow(logger).to receive(:debug) with expect_lazy_log
  that verifies message content (error details, probe IDs)
- Add dedicated test for serialization error log message
- Remove test for nil probe_repository (no longer applicable)
- Update RBS signatures to match non-optional types
- Reorder keyword params per StandardRB (required before optional)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Wrap the on_serialization_error callback invocation in a rescue block
so that exceptions from the callback do not break iteration over
remaining snapshots in the batch. This ensures that a failure in the
error-handling path (e.g., probe disabling or status reporting) does
not prevent other healthy snapshots from being sent.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Move build_status above the private keyword so it can be called
directly instead of via send(:build_status, ...). This method is
the general-purpose status builder used by all other public build_*
methods, and is now also called directly by ProbeNotifierWorker for
serialization error reporting.

- Add YARD documentation to build_status
- Update ProbeNotifierWorker to call build_status directly
- Update test expectations from receive(:send) to receive(:build_status)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Add synchronize method to ProbeRepository exposing the Monitor lock
for compound operations. Wrap add_probe, install_pending_method_probes,
and install_pending_line_probes in probe_repository.synchronize to
ensure atomicity of check-then-act sequences.

This restores the thread-safety properties of the old code where the
entire add_probe flow (check installed → check failed → hook → store)
was under a single lock. The Monitor is reentrant, so individual
ProbeRepository methods called within the synchronized block safely
re-acquire the lock.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

The comment claimed the rate limiter "allows an initial burst of
notifications, then refills tokens over time." In reality, the
TokenBucket is initialized with rate=1, max_tokens=1, which is a
simple 1-per-second limit with no burst capacity. Simplified to
accurately describe the behavior.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Update probe_notifier_worker integration spec to provide the
now-required probe_repository and probe_notification_builder
keyword arguments.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Resolve conflict in spec/datadog/di/serializer_spec.rb: keep both
DISerializerBinaryTestClass (from this branch) and
DISerializerStackOverflowTestClass/DISerializerOutOfMemoryTestClass
(from master).

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>