Shared telemetry and node inventory

latent-telemetry implements one bounded export pipeline per node, a structured local sink, and activation observation. latent-node supplies lifecycle hooks and bounded inventory collection. These Rust APIs are available to the standalone node composition and management adapters; they do not create a public listener.

Ownership and lifecycle observations

LocalActivationServices::observer attaches an ActivationObserver directly to the existing activation owner. The observer receives payload-free ActivationObservationContext and ActivationObservation records. It does not wrap the manager, allocate another cancellation owner, or finalize another ledger.

Receipt follows successful identity/journal reservation and is observed before the handle is returned. Phase observations follow committed journal transitions. Resolution supplies the pinned release, revision, and route generation. Admission supplies the actual granted budget. Queue-wait and phase durations come from the manager's monotonic elapsed time, including interruption while queued.

Terminal observation follows resource disposition, budget finalization, selection of the cancellation/completion winner, and terminal journal publication. It contains the exact resulting terminal state and consumption. Guest success, declared guest error, and platform failure have distinct classifications. Handle abandonment and contained poll/destructor panics use that same terminal path. Cleanup observations distinguish confirmed release/quarantine, reclamation before execution, abandonment, failure, and cases without an acquired cell. Standalone transport disconnects keep that same observation lifetime while the bounded cleanup supervisor drives the original owner. A raw disconnect emits no accepted explicit-cancellation observation; its terminal result does not by itself prove that the cell was safely reused.

Each lifecycle stage produces correlated logs/events. Terminal observation emits one complete activation span using the supplied span identity; intermediate stages do not create additional spans with that same ID.

Callbacks run outside journal and cancellation locks. Observer panics cannot replace a committed activation result. The built-in observer uses a short mutex critical section for bounded correlation bookkeeping; it does no exporter work under that lock. It never waits for exporter capacity.

Correlation is keyed by a manager/sequence token, independently of the caller's activation ID. Overlapping incarnations can coexist after journal eviction or across managers. Completing an older token cannot remove a newer one. Guest logs carry an activation ID: when that ID maps to multiple active tokens, the observer drops the ambiguous log instead of choosing a tenant or invocation.

Bounds and redaction

SharedActivationObserverConfig bounds active correlation count, each identity string, aggregate context bytes, borrowed guest-log input, and exported guest fields/body. Borrowed input is validated before retention or redaction copies. An entry-count cap also bounds boxed correlation records and their token/ID indexes. Capacity pressure rejects additional correlation entries and counts the drop; it never evicts a live entry. Terminal processing removes its matching token even when export submission fails. Retained allowlist vector/string capacities are validated at construction. Configuration arithmetic includes context strings, the secondary identifier index, and conservative bookkeeping. Context replacement does not reuse spare capacity from earlier field values.

Correlation includes bounded activation/root/parent IDs, tenant, service, contract/function, trace ID/span ID/flags, and pinned release/revision/generation. It carries no invocation payload, principal subject/claims, caller metadata, baggage, error payload/message, or cancellation reason. Exported identifiers may be truncated to the configured attribute bound; internal matching uses complete validated identifiers and tokens.

Guest message bodies are [REDACTED] by default. Unknown guest field names are dropped entirely, including their names. Only exact names in allowed_guest_field_names may export bounded values, under a guest. prefix. The field-name bound includes that prefix. Case-insensitive latent. names and credential-shaped field names are forbidden in the allowlist; recognized credential/backtrace-shaped values remain redacted. Trusted correlation comes only from the registered lifecycle context. Opting into body or field export requires selecting appropriate content; substring filtering is not a general secret classifier.

The Wasmtime structured-log bridge submits each accepted host log immediately. It does not replay a captured-log snapshot after execution. Default queue drops remain successful host-log acceptance, with normal budget charging and visible telemetry drop counters. Explicit fail_on_drop is a test option and may return host-log unavailability. The existing bounded runtime capture remains a local diagnostic facility, separate from redacted export.

Metric dimensions come from fixed enumerations: lifecycle stage, outcome class, platform error code, resource, severity, cancellation/disposition result, and configured cell class. Activation/tenant/service/release identifiers and guest values are never metric labels. Metrics include lifecycle/outcome counts, monotonic latency and queue wait, granted/final consumption, observed resource limits reached, cancellation and cleanup. Floating-point metrics are approximate; terminal logs preserve exact integer consumption. Reaching a grant is separate from the terminal resource-exhaustion error classification.

Export pipeline and local sink

TelemetryRuntime::spawn(config, sink) returns a TelemetryHandle and its worker owner. Submit metrics/logs/spans through try_emit_*; producers never await the sink. The default queue holds at most 1024 records, each charged at most 64 KiB. Charges include owned string capacities and conservative record/map bookkeeping. Channel slots and the single currently exporting record are separately bounded. The local sink defaults to 4096 records and 8 MiB, evicting its oldest records within both bounds. Configure observer attribute limits consistently with the pipeline's record and attribute limits.

Snapshots expose accepted/exported records, queue depth, full/closed/invalid drops, sink failures/timeouts, and flush/shutdown failures. Observer snapshots separately expose retained correlations and rejected or failed observation attempts. Sink failures do not change the production activation outcome.

Export, flush, and shutdown have explicit deadlines. Exporters must cooperate with async polling: a deadline cannot interrupt blocking code inside a callback. flush() is a barrier over preceding export attempts, including failed/timed-out attempts; inspect the counters to determine delivery. Dropping the runtime owner closes producers and aborts the worker. Queued/export futures are reclaimed when the runtime polls that abort. No exporter task belongs to a dormant service.

An external exporter implements TelemetrySink and is installed once at node composition. It should use bounded transport buffers and cooperate with its deadline. OTLP/network exporter implementations remain an integration choice; the maintained local sink supports tests and development without infrastructure.

Inventory

StandaloneInventoryReporter::new(config, node, sources) provides synchronous snapshot_now() and the InventoryReporter::snapshot() port. Collection reads the configured cell classes, scheduler counters, local quota totals, route generation, bounded cache/topology sources, and load observations. It creates no worker and never enumerates a release, deployment, tenant, or activation catalog. NodeInventory::metric_points() produces a bounded set of fixed-dimension metrics.

The result includes capacity/availability/active/quarantined cells, queue bounds and wait totals, route generation, cache costs, quotas, load age/availability, health/readiness, and topology. Cache source/metadata/compiled-image and in-flight preparation costs remain separate from process RSS and activation pins. retained_bytes charges the returned snapshot, including spare capacity and collection bookkeeping, rather than source-owned state.

Topology explicitly distinguishes NodeFixed, ActivationScoped, and ServiceResident resources. Configured counts and observed active counts are separate: None means unmeasured, and source availability/completeness remain explicit. Cache occupancy cannot establish process memory pressure. Busy cells can remain ready; unavailable/stale load or no usable cells makes the node unready. A scheduler that has stopped accepting work is unready even while its configured capacity remains visible; deliberate shutdown can remain healthy. Optional cache/topology failures degrade diagnostics without inventing observations or changing guest execution.

The standalone wasmtime-compiler topology row is a NodeFixed thread resource. Its configured count comes from the compiler pool's maximum workers, and its active count is the pool's observed live workers, including idle workers. Job counts and joined workers do not stand in for live threads. The row uses the existing bounded topology writer; a selection that omits rows remains incomplete.

Phase 2 audit and canary observations

The optional durable audit owner records supply-chain verification, protected native-cache lookups, and administrative mutations. Its bounded journal and authorized query surface have separate retention and failure semantics from the lossy telemetry exporter. Audit records retain validated identities and fixed reason codes; they do not persist credentials, invocation payloads, or arbitrary external errors. Critical control operations reserve their audit capacity before mutation. The activation path does no audit-journal I/O.

LocalActivationServices::canary attaches a bounded capture handle to the same activation owner, independently of structured telemetry configuration. It binds the checked revision/component/generation and consumes the actual terminal winner. Canary observation windows retain fixed outcome counts and latency buckets, with explicit open, draining, missing-sample, and incomplete coverage. Lost or unattributable outcomes cannot establish a healthy rollout. Window identities belong to bounded records, not metric labels; the metric dimension rules above remain unchanged.

Validation

cargo test -p latent-telemetry --locked
cargo test -p latent-node --locked
tools/validate_contracts.sh

Tiny fixtures cover redaction, distinct outcomes, pinned correlation, monotonic timing, overlapping tokens, bounded queues/sinks, failure isolation, and inventory bounds. Lifecycle and real-component integrations exercise observation through cleanup. The separate Phase 1 completion report records the completed scaling/reclamation gate; the extension report records later performance results and their limits.