Closes the last open code-review finding. The unreachable
IngestCachedTelemetryAsync path now carries production cached-call
lifecycle traffic, delivering the design's "AuditLog + SiteCalls in one
MS SQL transaction" guarantee. Before this commit, the SiteCalls
operational half had NO production transport at all — central's
SiteCallAuditActor.OnUpsertAsync had zero producers, so cached-call
operational state never reached the central mirror.
Site-side partition (so neither path double-emits):
- ISiteAuditQueue.ReadPendingCachedTelemetryAsync — new method returning
rows where Kind ∈ {CachedSubmit, ApiCallCached, DbWriteCached,
CachedResolve} AND ForwardState = Pending.
- ISiteAuditQueue.ReadPendingAsync — XML doc updated, SQLite impl now
filters Kind NOT IN the cached set so cached rows no longer ride the
audit-only drain.
New cached-drain in SiteAuditTelemetryActor:
- Optional IOperationTrackingStore? ctor param (null on central
composition roots — the cached scheduler is never armed there).
- Independent CachedDrain message + scheduler tick parallel to the
existing Drain — a stall on one path can't block the other; shared
lifecycle CTS gates both.
- OnCachedDrainAsync: reads cached audit rows, joins each with its
matching SiteCallOperational snapshot via CorrelationId →
TrackedOperationId from the tracking store, builds CachedTelemetryBatch,
pushes via IngestCachedTelemetryAsync, marks ack'd rows Forwarded.
- Orphan rows (no tracking snapshot, thrown tracking-store call,
missing CorrelationId) logged at Warning + skipped — they stay
Pending so reconciliation/retry picks them up later. Best-effort
contract preserved.
Central side: AuditLogIngestActor.OnCachedTelemetryAsync was already
implemented (M3 Bundle G dead code today, alive after this commit) —
performs InsertIfNotExists for AuditLog + UpsertAsync for SiteCalls
inside a BeginTransactionAsync. The handler is idempotent on EventId,
so any duplicate arrivals from concurrent push + reconciliation are
silent no-ops.
Composition root: AkkaHostedService now resolves IOperationTrackingStore
via GetService<>() (site-only) and threads it through the actor's
Props.Create.
Tests added (+3 in SiteAuditTelemetryActorTests):
- Cached rows route through the new transport, not the audit-only drain.
- Orphan cached row (no tracking match) is logged + skipped, drain
doesn't crash.
- Ordinary audit rows still flow through the audit-only drain unchanged.
- ParentExecutionIdCorrelationTests now unions both queues to assert
all expected Kinds remain covered after the partition.
Build clean; AuditLog.Tests 250/251 (the 1 fail is the pre-existing
date-sensitive PartitionPurgeTests integration flake explicitly accepted
across the session); SiteRuntime.Tests 302/302.
README regenerated: 0 pending of 481 total.
Session-final totals: 136 of 136 originally-open Theme findings closed
across 11 commits (10 themed batches + this architectural close).
Site: site emitters of SiteCallOperational (ExternalSystemClient, the script-API
cached call path in ScriptRuntimeContext, CachedCallLifecycleBridge) inject
INodeIdentityProvider and stamp SourceNode = NodeName at construction.
OperationTrackingStore call site in CachedCallTelemetryForwarder now stamps
SourceNode too.
Central: SiteCallAuditRepository.UpsertAsync INSERT includes SourceNode in the
column list; conditional monotonic UPDATE uses
COALESCE(@SourceNode, SourceNode) so later packets cannot blank a previously-
stamped value. After this commit every SiteCalls row carries node-a/node-b in
SourceNode (subject to monotonic preservation).
The store-and-forward retry loop emits the per-attempt and terminal cached
audit rows (ApiCallCached/DbWriteCached Attempted, CachedResolve) via
CachedCallLifecycleBridge from a CachedCallAttemptContext, not from the
script context. The ExecutionId rollout (Task 4) already threaded ExecutionId
and SourceScript through this path; ParentExecutionId — the spawning
inbound-API request's ExecutionId — was not, so those retry-loop rows had
ParentExecutionId = null even for an inbound-API-routed run.
Thread it additively as a sibling at every carry point ExecutionId passes
through:
- StoreAndForwardMessage gains ParentExecutionId (Guid?).
- StoreAndForwardStorage adds a nullable parent_execution_id column via the
same idempotent PRAGMA-probed ALTER TABLE migration; rows persisted by an
older build read back null (back-compat). The defensive Guid.TryParse read
helper (ParseExecutionId) is renamed ParseGuidColumn and reused for both
columns so a corrupt value cannot abort the retry sweep.
- StoreAndForwardService.EnqueueAsync gains an optional parentExecutionId
param, stamped onto the buffered message and surfaced on the
CachedCallAttemptContext built in the retry loop.
- CachedCallAttemptContext gains ParentExecutionId.
- CachedCallLifecycleBridge.BuildPacket sets AuditEvent.ParentExecutionId
from the context, beside the existing ExecutionId.
- IExternalSystemClient.CachedCallAsync / IDatabaseGateway.CachedWriteAsync
gain an optional parentExecutionId param; ScriptRuntimeContext's CachedCall
/ CachedWrite helpers pass _parentExecutionId.
All threading is additive — ParentExecutionId is Guid? everywhere, null for
non-routed runs, and old buffered S&F rows still deserialize with the new
field null.
The store-and-forward retry loop emits the per-attempt and terminal cached
audit rows (ApiCallCached/DbWriteCached Attempted, CachedResolve) via
CachedCallLifecycleBridge from a CachedCallAttemptContext, not from the
script context. ExecutionId (and SourceScript) were not threaded through the
S&F buffer, so those rows had ExecutionId = null and SourceScript = null.
Thread both, additively, from the cached-call enqueue path:
- StoreAndForwardMessage gains ExecutionId (Guid?) / SourceScript (string?).
- StoreAndForwardStorage adds nullable execution_id / source_script columns
via an idempotent PRAGMA-probed ALTER TABLE migration; rows persisted by
an older build read back null (back-compat).
- StoreAndForwardService.EnqueueAsync gains optional executionId /
sourceScript params, stamped onto the buffered message and surfaced on the
CachedCallAttemptContext built in the retry loop.
- CachedCallAttemptContext gains ExecutionId / SourceScript.
- CachedCallLifecycleBridge.BuildPacket sets AuditEvent.ExecutionId and
AuditEvent.SourceScript from the context (replacing the hard-coded
SourceScript = null and its now-stale comment).
- IExternalSystemClient.CachedCallAsync / IDatabaseGateway.CachedWriteAsync
gain optional executionId / sourceScript params; ScriptRuntimeContext's
CachedCall / CachedWrite helpers pass _executionId / _sourceScript.
Script-side cached rows (CachedSubmit, immediate Attempted+Resolve) are
unchanged. All threading is additive — old buffered S&F rows still
deserialize and process with the new fields null.
Wire the M3 cached-call audit pipeline end-to-end for the database
channel and close the loop between the S&F lifecycle observer and the
site-side dual emitter.
* DatabaseCachedWriteEmissionTests covers Database.CachedWrite (set up
in Bundle E3): mints a TrackedOperationId, emits one CachedSubmit
packet on DbOutbound, threads the id into IDatabaseGateway, and is
best-effort on a thrown forwarder. Mirrors ExternalSystem.CachedCall
coverage from E3.
* CachedCallLifecycleBridge (new) implements ICachedCallLifecycleObserver
and lives alongside CachedCallTelemetryForwarder. The bridge ingests
per-attempt notifications from the S&F retry loop and fans them out
to the forwarder:
- TransientFailure -> 1 Attempted row
- Delivered -> Attempted + CachedResolve(Delivered)
- PermanentFailure -> Attempted + CachedResolve(Parked)
- ParkedMaxRetries -> Attempted + CachedResolve(Parked)
Channel string -> AuditKind mapping (ApiOutbound->ApiCallCached,
DbOutbound->DbWriteCached). Best-effort top-level catch swallows any
unexpected throw so the S&F retry bookkeeping is never disturbed.
* Bridge tests (7) cover all four outcomes, channel mapping, provenance
propagation, and the no-throw-on-forwarder-failure contract.
Bundle F (Host registration) will instantiate the bridge and inject it
into StoreAndForwardService.cachedCallObserver, closing the wiring path
end-to-end.
Bundle E task E6.
Sister to SiteAuditTelemetryActor: takes a combined CachedCallTelemetry
packet and fans it out to the two site-local stores.
* AuditEvent half writes through IAuditWriter (the M2 FallbackAuditWriter
+ SqliteAuditWriter chain — same site SQLite hot-path as sync calls).
* SiteCallOperational half maps Audit.Kind to the matching
IOperationTrackingStore method:
- CachedSubmit -> RecordEnqueueAsync (insert-if-not-exists)
- ApiCallCached / DbWriteCached -> RecordAttemptAsync (monotonic)
- CachedResolve -> RecordTerminalAsync (first-write-wins)
Best-effort contract (alog.md §7): independent try/catch per half so a
thrown writer cannot starve the tracking row (and vice-versa); both
failures are logged at warning level and swallowed — the calling script
never sees them.
Wire push deferred to M6 — the NoOp ISiteStreamAuditClient binding stays
in effect; the forwarder writes only to the local stores in M3. The
existing SiteAuditTelemetryActor drain loop will sweep the audit rows
once a real gRPC client lands.
Bundle E task E2.
Add the second site→central RPC seam alongside the existing M2
IngestAuditEventsAsync. The Bundle D proto already lit up
IngestCachedTelemetry (CachedTelemetryBatch / IngestAck) so this commit
just plumbs the client-side abstraction:
* ISiteStreamAuditClient gains IngestCachedTelemetryAsync(batch, ct).
* NoOpSiteStreamAuditClient implements it returning an empty IngestAck
(same shape as M2 — production gRPC client lands in M6).
* SyncCallEmissionEndToEndTests' DirectActorSiteStreamAuditClient stub
throws NotSupportedException from the new method so a regression that
accidentally routes a cached packet through the sync stub fails loudly.
* New NoOpSiteStreamAuditClientTests cover the null-guard + empty-ack
contract for both batch shapes.
Bundle E task E1.
Joseph Doherty