Keep JSONRPC output backpressure outside configuration and deduplication locks

[karthiknadig]

Tracking plan: #528
Priority: P2. Evidence: lock scopes confirmed by source inspection; reproduce impact with controlled slow-consumer tests.

Problem

The deduplicating reporter invokes its downstream reporter while holding the manager/environment map write lock. The generation-guarded reporter holds the configuration read lock through notification reporting. The downstream JSONRPC writer synchronously writes/flushes stdout.

A slow or disconnected consumer can therefore couple transport I/O to discovery deduplication and configuration publication. The audit does not claim a measured production deadlock; the first task is to prove the coupling and required ordering with deterministic tests.

Sources: deduplicating reporter, generation reporter, framed output.

Scope

Separate deduplication and generation decisions from blocking output. Evaluate a bounded single-writer queue or equivalent narrow transport boundary; do not replace lock contention with an unbounded memory queue. Define exactly where generation checks occur and how notification/reply ordering is guaranteed when configure races with queued output.

Make serialization/write/flush failures explicit and preserve stdout as framed protocol output only. Rename CacheReporter around deduplication if the type is changed, and retain only the data its behavior needs.

Acceptance criteria

• A controllably blocked output sink cannot hold the configuration or deduplication locks for the duration of the write.
• Tests cover slow consumers, closed output, queue saturation, and shutdown with pending notifications.
• Stale-generation notifications are suppressed at the documented publication boundary, including configure racing with queued output.
• Environment/manager deduplication remains atomic; a refresh reply cannot overtake the notifications it is defined to complete.
• Concurrent replies/notifications remain correctly framed, with bounded buffering and explicit error propagation.
• Gate client-observed refresh latency and define accurate first-environment timing #531/Benchmark long-lived PET sessions, real concurrent resolves, and inventory scaling #533 show no regression in complete-inventory correctness, first-result latency, or resource bounds.

Dependencies

Depends on #536 for generation/snapshot semantics and #533 for slow-consumer/resource validation. Build on the transport seams from #529/#532. Coordinate with bounded scheduling #539 rather than introducing a second conflicting queue policy. The complete sequence is in #528.