Phase 2 canary outcome observation

latent-telemetry::BoundedPhase2CanaryOutcomeWindow collects bounded activation outcomes for control-registered rollout cohorts. The activation manager now owns capture through terminal publication. The rollout coordinator and canary evaluator use this primitive with explicit policies. There is no public outcome-submission RPC. The implementation evolves the outcome classes, finite labels, bounds and tests contributed in PR176.

Registration and exact attribution

A trusted host creates the shared owner, calls register(&CanaryWindowSpec) and retains the returned CanaryWindow. Its immutable identity contains tenant, service, deployment, rollout ID, coordinator step and route generation. Up to eight revision bindings contain the exact revision and canonical component ReleaseDigest, plus the actual optional PackageDigest. Package associations come from the trusted control owner; capture does not infer them from arbitrary route attributes or independently verify publication authority. Trusted-local content without a package uses None.

The owner assigns a non-reused process-local epoch to each registration. At most one open window observes a tenant/service pair. Registration validates and makes fresh bounded copies of strings and revision rows, dropping caller spare capacity. An expired or explicitly closed window may coexist with a new window, but their epochs and sample owners remain separate. A CanaryWindow grants only access to its own readout. The rollout adapter authenticates tenant administrators; supplying tenant text is never an authorization grant. Rollout windows also bind an opaque control digest covering the catalog owner, exact rollout revision, policy and compiled cohort; diagnostic windows may omit it.

Install Some(owner.capture_handle()) in LocalActivationServices.canary before constructing the activation manager. This is optional and independent of its structured telemetry observer. The standalone node supplies the same repository and coordinator hub when optional rollouts.canary settings are enabled. Its clock is shared with activation capture. Otherwise the node leaves capture off. The maintained integration fixture shows the complete host setup in canary.rs.

One accepted activation, one sample owner

The manager attempts capture once after accepting the activation into its journal, using the existing unique manager/sequence observation token. The returned CanarySample is affine: it cannot be cloned or deserialized. The existing lifecycle binds it to the checked selected revision, records actual admission, and consumes it only after accounting and the terminal journal transaction finalize. Repeated phase refreshes cannot change the selection. Completion never resolves the current route again.

The canary-only path does not clone the structured observer context or retain payloads, principal claims, trace baggage, arbitrary diagnostics or cancellation reasons. Its Invoke hooks use fixed counters, existing immutable slots, Arc ownership and nonblocking try_lock; they do not insert allocating maps, wait for a query/exporter, perform I/O or start tasks. Control registration allocates the bounded slots and metadata. This adds no blocking acquisition to Invoke and does not remove unrelated pre-existing observer/execution locks.

The five fixed outcome classes remain success, declared domain error, other platform error, deadline exceeded and cancelled. The finalized terminal result is authoritative: a cancellation request that loses to committed success still counts as success. The collector does not decide whether a domain error is healthy. Dropping a sample without terminal publication records loss before releasing its live allowance; it never refunds execution budgets or changes the activation result.

Window membership and denominators

Registration begins a half-open monotonic interval of at most one hour. A successful capture inside that interval remains in its original cohort even when it finishes after the interval closes. close() stops new memberships without waiting for existing samples. Snapshot/registration use the same bounded registry gate; terminal bookkeeping becomes visible before the live count falls. A retained sample cannot inject into a replacement window or release its slot prematurely. Dropping the control owner retires the window and prevents its retained readouts from becoming complete evidence.

window.snapshot(required_samples) returns a bounded owned readout with:

• exact immutable identity, epoch and revision/package/component bindings;
• total starts, selected/admitted starts, terminal and live counts;
• unattributed terminal and abandoned counts;
• per-revision selected/admitted/admitted-terminal counts and outcome classes;
• nine fixed latency buckets from acceptance-relative monotonic elapsed time.

Latency bucket upper bounds, expressed in microseconds, are 100, 1,000, 5,000, 10,000, 50,000, 100,000, 1,000,000 and 10,000,000; the ninth is overflow. Latency includes all selected terminal outcomes, including selected admission failures. Comparison uses the exact duration: an edge plus one nanosecond is above that inclusive bound, without rounding it down to whole microseconds. Wall-clock sample timestamps are not used to choose cohort membership.

The required count is for the whole cohort, not an implicit minimum for each candidate. Consumers have separate revision counts to choose their denominators. Resolution failure and pre-resolution abandonment are explicitly unattributed; they cannot be fabricated as candidate successes or failures.

CanaryCoverage distinguishes Open, Draining, NoSamples, Insufficient, Incomplete and CompleteData. Complete data requires closure, zero live owners, complete selected terminal accounting, no known loss and the requested count. It is not permission to promote. Early explicit closure can produce complete diagnostic accounting, but cannot produce a full-duration sealed window. A drained closed window with no calls remains NoSamples rather than becoming a zero-error success rate.

Loss, ownership and bounds

Sample/window/live capacity failures mark the affected window incomplete. A failed registry try-acquisition cannot identify the affected window reliably, so it advances a shared loss epoch. Every retained window spanning that epoch is conservatively incomplete. A retained snapshot's coverage() may therefore downgrade after it was taken; its counters remain the historical copy. Epoch saturation or competing loss updates that cannot be recorded in one atomic attempt set a permanent unknown flag for that shared owner. Creating a fresh owner establishes a new observation domain, never restores old healthy evidence.

Terminal gate contention, missing correlation, unexpected selection changes and unfinished Drop likewise never disappear as healthy missing samples. Counts of unknown dropped calls are intentionally not guessed. Structured exporter loss is independent: the affine capture path neither depends on exporter delivery nor creates a durable audit record per Invoke. Rollout decision audit records belong to the control owner and the separate Phase 2 audit journal.

Resource	Default	Hard ceiling
Retained windows (maximum_series)	16	64
Revision bindings per window	8	8
Captured starts per window	10,000	1,000,000
Starts across retained windows	100,000	16,000,000
Bytes per arbitrary identity string	256	1,024
Live sample owners	4,096	65,536
Concurrent owned snapshots	4	16

The finite window/revision/string product also bounds retained metadata; hashes have fixed canonical length. Fixed arrays bound counter/histogram storage. Slots and captured-start charges remain retained while samples or snapshots own a retired window. A snapshot cannot be cloned and retains a read allowance until Drop. Control reads return bounded errors on contention or pressure. Query and capture do not evict earlier samples to manufacture coverage. Public diagnostic counter total() saturates for arbitrary literals; validated retained counters cannot approach u64 overflow under these limits.

There is no worker, connection, timer, execution cell or guest Store per window, and no unbounded sample log. After restart the data is absent and a new epoch is required. Counters, identities and snapshots are host observation data, not trust, eligibility or promotion capabilities. Only fixed outcome/coverage labels are suitable metric dimensions; never export identifiers or digests as labels.

Sealed observation and validation

window.try_seal() returns a private-constructed, non-cloneable SealedCanaryWindow only after the complete interval, closed membership, no live samples and no loss. An atomic live-attempt guard covers registry acquisition through loss publication. The seal requires that frontier to be quiescent, so a failed capture paused before publishing loss cannot disappear from the decision. Registration captures its loss baseline before the interval starts.

A sealed window owns the same actual window slot and a snapshot allowance. Its frozen counters do not change when later, unrelated captures lose data. The catalog still verifies exact configured owner/control binding and policy itself; possession of copied counters or a diagnostic assessment cannot mint this input. Each attempted seal has a finite immediate result and creates no waiting task.

The small unit tests retain PR176's limits/fresh-ownership/label cases and add success-to-cap followed by rejected failure, exact selection, live retirement, snapshot ownership, gate contention, shared loss, missing terminals, epoch separation, half-open membership and late terminal accounting. Actual activation manager tests cover selected generation changes, finalized outcome categories, cancellation, pre-poll abandonment and unchanged resource accounting without running a guest compiler or load benchmark.

cargo test -p latent-telemetry --locked
cargo test -p latent-node --test activation_lifecycle --locked

The evaluator uses candidate-specific minimums and exact integer rate criteria, and promotion rechecks catalog state and release eligibility at commit. Tests also cover early closure, the delayed-loss frontier, registration loss ordering, nanosecond threshold boundaries, real retained proof allowances and restart without old positive evidence. Promotion remains operator-triggered.