scadalink-design/code-reviews/StoreAndForward/findings.md

# Code Review — StoreAndForward

| Field | Value |
|-------|-------|
| Module | `src/ScadaLink.StoreAndForward` |
| Design doc | `docs/requirements/Component-StoreAndForward.md` |
| Status | Reviewed |
| Last reviewed | 2026-05-16 |
| Reviewer | claude-agent |
| Commit reviewed | `9c60592` |
| Open findings | 11 |

## Summary

The Store-and-Forward module is small and readable, with a clean SQLite persistence
layer, a sensible service API, and reasonable test coverage of the storage and service
happy paths. However the review surfaced two issues that undermine the module's core
purpose. First, the active delivery path never invokes the `ReplicationService` —
`ReplicateEnqueue/Remove/Park` have no callers anywhere in the codebase, so buffered
messages are not replicated to the standby node and the design's failover-durability
guarantee (Component doc "Persistence", CLAUDE.md "Store-and-Forward") is not met.
Second, there is an off-by-one in retry accounting: the immediate-failure path stores a
buffered message with `RetryCount = 1`, so a message configured with `MaxRetries = N`
is actually attempted `N` times in total rather than one immediate attempt plus `N`
retries, and a per-source `MaxRetries` of 1 produces zero retry attempts. Additional
themes: SQLite connection-per-call with no transactional grouping of multi-statement
operations, no concurrency guard against a parked message being retried while the
sweep is mid-flight, an unused enum member (`InFlight`) that drifts from the documented
status set, and untested critical paths (retry-due timing, replication-from-active,
the actor bridge). None of the findings are blockers for compilation, but the
replication and retry-count issues are functional defects against the design.

## Checklist coverage

| # | Category | Examined | Notes |
|---|----------|----------|-------|
| 1 | Correctness & logic bugs | ☑ | Off-by-one in retry counting (003); parked-message retry timing (010). |
| 2 | Akka.NET conventions | ☑ | `ContinueWith` used instead of `PipeTo`-friendly continuations; default supervision; see 007. |
| 3 | Concurrency & thread safety | ☑ | Sweep guarded by `Interlocked`, but no guard against retry-vs-manage races (005); `OnActivity` event not thread-safe (009). |
| 4 | Error handling & resilience | ☑ | Replication never invoked from active path (001); no-handler messages buffered then stuck (002). |
| 5 | Security | ☑ | No issues found — parameterised SQL throughout; no secrets handled directly; payload JSON treated opaquely. |
| 6 | Performance & resource management | ☑ | New SQLite connection per call; multi-statement operations not wrapped in a transaction (006, 008). |
| 7 | Design-document adherence | ☑ | Replication gap (001); `InFlight` status undocumented/unused (011); "retrying" status from design doc not modelled. |
| 8 | Code organization & conventions | ☑ | `StoreAndForwardMessage` is an entity-like POCO living in the component, not Commons (012). |
| 9 | Testing coverage | ☑ | Retry-due timing, replication-from-active, and `ParkedMessageHandlerActor` are untested (013). |
| 10 | Documentation & comments | ☑ | XML doc on `RegisterDeliveryHandler` contract is inconsistent with code (004). |

## Findings

### StoreAndForward-001 — Replication to standby is never triggered by the active node

| | |
|--|--|
| Severity | Critical |
| Category | Error handling & resilience |
| Status | Resolved |
| Location | `src/ScadaLink.StoreAndForward/ReplicationService.cs:40`, `:53`, `:66`; `src/ScadaLink.StoreAndForward/StoreAndForwardService.cs:155`, `:212`, `:222`, `:236` |

**Description**

`ReplicationService` exposes `ReplicateEnqueue`, `ReplicateRemove` and `ReplicatePark`
to forward buffer operations to the standby node, but a codebase-wide search shows these
methods have no callers. `StoreAndForwardService` — which performs every add (`EnqueueAsync`
line 155 / 163), remove (`RemoveMessageAsync` call at line 212) and park
(`UpdateMessageAsync` calls at lines 222/236) — holds no reference to `ReplicationService`
and never invokes it. Only the receiving half is wired (`SetReplicationHandler` and
`ApplyReplicatedOperationAsync` are used by `SiteReplicationActor`). The Component design
doc ("Persistence") and CLAUDE.md ("Store-and-Forward") require the active node to
forward every buffer operation to the standby so that, on failover, the new active node
"has a near-complete copy of the buffer." As written, the standby's S&F SQLite database
stays empty and a failover loses the entire buffer — a data-loss defect against a core
requirement.

**Recommendation**

Inject `ReplicationService` into `StoreAndForwardService` and call `ReplicateEnqueue`
after a successful `_storage.EnqueueAsync`, `ReplicateRemove` after `RemoveMessageAsync`,
and `ReplicatePark` after a park-causing `UpdateMessageAsync`. Update
`ServiceCollectionExtensions.AddStoreAndForward` to pass the dependency. Add a test that
asserts the replication handler observes each operation type.

**Resolution**

Resolved 2026-05-16. `ReplicationService` is now injected into `StoreAndForwardService`
(wired in `AddStoreAndForward`), and every buffer operation is forwarded to the standby:
a new `BufferAsync` helper calls `ReplicateEnqueue` after each persist, `ReplicateRemove`
runs after a successful retry removes a message, and `ReplicatePark` runs on both park
paths. Replication stays fire-and-forget and is a no-op when `ReplicationEnabled` is
false or no handler is wired. Regression tests `StoreAndForwardReplicationTests` assert
the replication handler observes the Add, Remove and Park operations. Fixed by the
commit whose message references `StoreAndForward-001`.

### StoreAndForward-002 — Messages enqueued with no registered handler are buffered but never deliverable

| | |
|--|--|
| Severity | ~~High~~ → Low (re-triaged) |
| Category | Error handling & resilience |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardService.cs:162`, `:201` |

**Description**

`EnqueueAsync` falls through to "No handler registered — buffer for later" (line 162)
when no delivery handler is registered for the category. The retry sweep
(`RetryMessageAsync`, line 201) then logs "No delivery handler for category" and
`return`s without touching the message. No caller in the codebase ever calls
`RegisterDeliveryHandler` (the External System Gateway, Notification Service and
Database Gateway only call `EnqueueAsync`), so in the current wiring **every** buffered
message lands in this dead state: it is persisted, counts toward buffer depth, but can
never be retried, delivered or parked. It will sit Pending forever. Either the handler
registration is missing from Host/gateway startup, or the "buffer for later" path is a
silent trap. Either way the engine cannot deliver anything.

**Recommendation**

Decide the intended contract. If handlers are expected to be registered before
`EnqueueAsync` is reachable, make `EnqueueAsync` reject (or log an error) when no
handler exists rather than silently buffering an undeliverable message, and wire
`RegisterDeliveryHandler` calls in Host startup for all three categories. If late
registration is intended, the retry sweep should treat a still-missing handler as a
transient condition with bounded logging rather than a permanent no-op.

**Re-triage note (2026-05-16)**

The finding's central factual claim — *"No caller in the codebase ever calls
`RegisterDeliveryHandler`"* and therefore *"every buffered message lands in this dead
state"* — is **no longer true at the reviewed code**. `ScadaLink.Host`
(`AkkaHostedService.RegisterSiteActors`, `AkkaHostedService.cs:353-379`) registers all
three delivery handlers (`ExternalSystem`, `CachedDbWrite`, `Notification`) at site
startup, immediately after `StoreAndForwardService.StartAsync()`. The finding was
written against commit `9c60592` before that wiring existed; the High-severity
"engine cannot deliver anything" outcome no longer occurs.

The remaining residual risk is narrow: a message enqueued for a category that genuinely
has no handler (e.g. an enqueue racing ahead of `RegisterDeliveryHandler`, or a future
category added without a handler) is still buffered and then skipped by the sweep
forever. That is a real but minor robustness gap, hence the **downgrade to Low**.

It is left **Open** rather than fixed in this pass because the finding's recommended
fix — making `EnqueueAsync` reject when no handler is registered — is a behavioural
contract change, not a localised bug fix: the "buffer with no handler yet" path is
exercised by `StoreAndForwardReplicationTests` and by three NotificationService and
ExternalSystemGateway tests (`Send_TransientError_WithStoreAndForward_BuffersMessage`,
`Send_Smtp4xxCommandException_ClassifiedTransientAndBuffered`,
`Send_SmtpProtocolException_ClassifiedTransient`) which construct a real
`StoreAndForwardService` without registering a handler and assert `WasBuffered`.
Changing the contract requires deciding whether late handler registration is supported
and updating tests in modules outside this review's edit scope — a design decision that
should be made deliberately rather than forced here.

**Resolution**

_Open — re-triaged to Low. Premise (no handler registration anywhere) is stale: Host
now wires all three handlers. Residual gap is minor and the prescribed fix is a
cross-module contract change needing a design decision._

### StoreAndForward-003 — Off-by-one in retry accounting: immediate failure pre-counts as retry 1

| | |
|--|--|
| Severity | High |
| Category | Correctness & logic bugs |
| Status | Resolved |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardService.cs:153`, `:229`, `:233` |

**Description**

On a transient immediate-delivery failure, `EnqueueAsync` buffers the message with
`message.RetryCount = 1` (line 153). The retry sweep then increments `RetryCount` before
the max check (`RetryCount++` at line 229; `RetryCount >= MaxRetries` at line 233).
Consequences: (1) a message configured with `MaxRetries = 1` is parked on the *first*
retry sweep without ever being retried, because after the immediate attempt `RetryCount`
is already 1 and the first sweep makes it 2 ≥ 1 — zero actual retries occur, contradicting
the design intent that the immediate attempt and the retry budget are distinct;
(2) the design doc's `Retry Count` field is "Number of attempts so far," but here it is
seeded to 1 before any *retry* has happened, making the parked-message `AttemptCount`
shown to operators off by one relative to configured `MaxRetries`. The
`EnqueueAsync_TransientFailure_BuffersForRetry` test even asserts `RetryCount == 1`,
locking in the ambiguity.

**Recommendation**

Choose one consistent meaning for `RetryCount` (recommended: total delivery attempts,
including the immediate one) and apply it uniformly. If `MaxRetries` is meant to bound
*retries* after the immediate attempt, buffer with `RetryCount = 0` and treat the
immediate failure as attempt 0; if it bounds *total attempts*, document that and adjust
the comparison. Update the affected test to match the chosen semantics.

**Resolution**

Resolved 2026-05-16 (commit `<pending>`). `RetryCount` now consistently means "number
of background retry-sweep attempts so far"; the initial immediate (or caller-made)
delivery attempt is attempt 0 and is not counted, and `MaxRetries` bounds retry-sweep
attempts after that initial attempt. `EnqueueAsync` no longer seeds `RetryCount = 1` on
either the transient-immediate-failure path or the `attemptImmediateDelivery: false`
path — a freshly buffered message has `RetryCount = 0`. `RetryMessageAsync` already
increments before the `>= MaxRetries` check, which is now correct, so a message with
`MaxRetries = 1` gets exactly one real retry before parking (previously zero). The
`StoreAndForwardMessage.RetryCount` XML doc was corrected to match. Regression test
`RetryPendingMessagesAsync_MaxRetriesOne_PerformsExactlyOneRetryBeforeParking` asserts
the immediate attempt plus exactly one retry occur before parking; the affected
existing tests (`EnqueueAsync_TransientFailure_BuffersForRetry`,
`EnqueueAsync_AttemptImmediateDeliveryFalse_BuffersWithoutInvokingHandler`,
`RetryPendingMessagesAsync_MaxRetriesReached_ParksMessage`) were updated to the
corrected semantics.

### StoreAndForward-004 — `RegisterDeliveryHandler` XML doc contradicts the implemented contract

| | |
|--|--|
| Severity | Medium |
| Category | Documentation & comments |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardService.cs:38`, `:60` |

**Description**

The XML comment on the handler delegate (lines 37–40) says "Returns true on success,
throws on transient failure. Permanent failures should return false (message will NOT
be buffered)." That last clause is wrong for the retry path: in `RetryMessageAsync`,
a handler returning `false` does not "not buffer" — the message is already buffered, and
the code *parks* it immediately (lines 218–224). The comment describes only the
`EnqueueAsync` immediate path and misleads anyone implementing a handler about what
`false` means once a message is in the retry loop.

**Recommendation**

Reword the contract to cover both paths explicitly: `true` = delivered (remove from
buffer); `false` = permanent failure (not buffered on immediate attempt, parked on a
retry); exception = transient failure (buffer / increment retry).

**Resolution**

_Unresolved._

### StoreAndForward-005 — Parked-message retry/discard can race with the in-progress retry sweep

| | |
|--|--|
| Severity | Medium |
| Category | Concurrency & thread safety |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardService.cs:184`, `:266`, `:280` |

**Description**

`RetryPendingMessagesAsync` loads a snapshot of due messages (line 179) and then
processes them one by one (line 184), `await`-ing delivery for each. Meanwhile
`RetryParkedMessageAsync` / `DiscardParkedMessageAsync` (operator actions arriving via
`ParkedMessageHandlerActor`) run on unrelated threads and mutate the same rows. Because
each operation opens its own SQLite connection and there is no row-level coordination,
an operator can `DiscardParkedMessageAsync` a message that the sweep is concurrently
delivering: the sweep's later `RemoveMessageAsync`/`UpdateMessageAsync` operates on a
now-deleted row (harmless) — but if an operator `RetryParkedMessageAsync` resets a row
to Pending while the sweep simultaneously parks the same in-flight message, the operator
intent is silently overwritten. The `Interlocked` guard only prevents *overlapping
sweeps*, not sweep-vs-management races.

**Recommendation**

Funnel all message-state mutations through a single serialization point — e.g. perform
all S&F state changes inside the `ParkedMessageHandlerActor` (or a dedicated S&F actor)
so the actor mailbox serialises them, or make status transitions conditional in SQL
(e.g. `UPDATE ... WHERE id = @id AND status = @expected`) and re-check the affected
row count.

**Resolution**

_Unresolved._

### StoreAndForward-006 — `GetParkedMessagesAsync` count and page run without a transaction

| | |
|--|--|
| Severity | Low |
| Category | Performance & resource management |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardStorage.cs:166`, `:175` |

**Description**

`GetParkedMessagesAsync` issues a `COUNT(*)` and then a separate paged `SELECT` on two
commands on the same connection with no surrounding transaction. A concurrent
enqueue/park/discard between the two statements yields a `TotalCount` inconsistent with
the returned page (e.g. total reported as 51 while only 50 distinct parked rows now
exist, or a row visible in the page but excluded from the count). For a paginated UI
this produces flickering totals and occasional off-by-one page math.

**Recommendation**

Wrap both reads in a single transaction (`BeginTransaction`) so they see a consistent
snapshot, or accept the staleness and document it. A transaction is cheap here and
removes the inconsistency.

**Resolution**

_Unresolved._

### StoreAndForward-007 — Async work in `ParkedMessageHandlerActor` uses `ContinueWith` without scheduler/affinity guarantees

| | |
|--|--|
| Severity | Low |
| Category | Akka.NET conventions |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/ParkedMessageHandlerActor.cs:34`, `:68`, `:87` |

**Description**

The three handlers call a `Task`-returning service method and chain `.ContinueWith(...)
.PipeTo(sender)`. `Sender` is correctly captured into a local first, so the closure is
safe. However `ContinueWith` without an explicit `TaskScheduler` runs the continuation
on a thread-pool thread and the captured continuation builds the response objects there
— acceptable since it only touches locals, but it bypasses the idiomatic
`PipeTo`-with-success/failure-projection pattern and is fragile if someone later adds a
line touching actor state inside the continuation. There is also no `TaskContinuationOptions`,
so a faulted antecedent still runs the continuation (handled here via `IsCompletedSuccessfully`,
but only by convention).

**Recommendation**

Replace `ContinueWith(...).PipeTo(sender)` with `PipeTo(sender, success: result => ...,
failure: ex => ...)`, which is the documented Akka pattern, keeps response construction
off the actor thread safely, and makes the success/failure branches explicit.

**Resolution**

_Unresolved._

### StoreAndForward-008 — A SQLite connection is opened and torn down on every storage call

| | |
|--|--|
| Severity | Low |
| Category | Performance & resource management |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardStorage.cs:28`, `:61`, `:93`, `:117`, `:144`, `:162`, `:199`, `:221`, `:237`, `:267`, `:285`, `:305`, `:319` |

**Description**

Every method in `StoreAndForwardStorage` constructs a fresh `SqliteConnection` and calls
`OpenAsync`. Microsoft.Data.Sqlite pools connections, so this is not a correctness bug,
but a retry sweep over a large buffer performs one open per `UpdateMessageAsync`/
`RemoveMessageAsync` call inside the loop (`RetryMessageAsync`), multiplying connection
churn under load. With no max buffer size (by design) the buffer can grow large, so the
per-message connection acquisition is a measurable overhead on the hot retry path.

**Recommendation**

Consider a batched retry API that opens one connection (and one transaction) per sweep,
or pass an open connection into the per-message update calls. At minimum, document that
the design relies on the Sqlite connection pool for acceptable performance.

**Resolution**

_Unresolved._

### StoreAndForward-009 — `OnActivity` event invocation is not thread-safe against concurrent subscribe/unsubscribe

| | |
|--|--|
| Severity | Low |
| Category | Concurrency & thread safety |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardService.cs:46`, `:309` |

**Description**

`OnActivity` is a public `event Action<...>` raised via `OnActivity?.Invoke(...)` in
`RaiseActivity` (line 309). `RaiseActivity` is called from both `EnqueueAsync` (caller
thread) and `RetryMessageAsync` (timer thread). The `?.Invoke` null-conditional captures
the delegate once so it will not NRE, but there is no synchronisation around the event
field itself; a subscriber added/removed concurrently with a raise has no defined
ordering. More importantly, subscriber callbacks run synchronously on the timer thread,
so a slow or throwing subscriber stalls or aborts the retry sweep (an exception in a
subscriber propagates out of `RaiseActivity` into `RetryMessageAsync`'s `try` and is
swallowed as a "transient failure," wrongly incrementing the message's retry count).

**Recommendation**

Snapshot the delegate (already done) and additionally wrap subscriber invocation in a
`try/catch` so a faulting logging subscriber cannot be misclassified as a delivery
failure. Document that handlers must be fast and non-throwing, or dispatch activity
notifications asynchronously.

**Resolution**

_Unresolved._

### StoreAndForward-010 — Retry of a parked message does not reset `LastAttemptAt`, so its retry timing is unspecified

| | |
|--|--|
| Severity | Medium |
| Category | Correctness & logic bugs |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardStorage.cs:203`, `:101` |

**Description**

`RetryParkedMessageAsync` sets `status = Pending, retry_count = 0, last_error = NULL`
but leaves `last_attempt_at` unchanged (line 203–206). The retry-due query
(`GetMessagesForRetryAsync`, line 101–105) selects Pending rows where
`last_attempt_at IS NULL OR ... elapsed >= retry_interval_ms`. A message parked after
exhausting retries has an old `last_attempt_at`; once re-queued, the elapsed time since
that stale timestamp is almost certainly already greater than the retry interval, so the
operator-retried message is attempted on the very next sweep regardless of the
configured interval. That is probably the desired behaviour (operator wants it tried
now), but it is unspecified and inconsistent — if `retry_interval_ms` were very large the
behaviour would instead be "try immediately" by accident rather than by design.

**Recommendation**

Explicitly decide and encode the intent: either set `last_attempt_at = NULL` on
re-queue so the message is unambiguously due now, or set it to "now" so it waits one
interval. Document the chosen behaviour in the method's XML comment.

**Resolution**

_Unresolved._

### StoreAndForward-011 — `StoreAndForwardMessageStatus.InFlight` is unused and the doc's "retrying" status is unmodelled

| | |
|--|--|
| Severity | Low |
| Category | Design-document adherence |
| Status | Open |
| Location | `src/ScadaLink.Commons/Types/Enums/StoreAndForwardMessageStatus.cs:9`; `src/ScadaLink.StoreAndForward/StoreAndForwardService.cs:219`, `:235` |

**Description**

The enum defines `Pending, InFlight, Parked, Delivered`. The module only ever uses
`Pending` and `Parked` — `InFlight` and `Delivered` are never assigned (delivered
messages are deleted, not marked `Delivered`). Meanwhile the Component design doc
("Message Format" -> Status) specifies the set "Pending, retrying, or parked." So the
code's enum drifts from the doc in two directions: it carries dead members the doc does
not mention (`InFlight`, `Delivered`) and omits the doc's `retrying` state. A message
mid-retry is indistinguishable from one that has never been attempted.

**Recommendation**

Reconcile the enum with the design. Either drop the unused members and update the doc,
or implement the documented `retrying` state and use `InFlight` to mark a message the
sweep is actively delivering (which would also help with finding 005).

**Resolution**

_Unresolved._

### StoreAndForward-012 — `StoreAndForwardMessage` is a persistence entity but lives in the component, not Commons

| | |
|--|--|
| Severity | Low |
| Category | Code organization & conventions |
| Status | Open |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardMessage.cs:9` |

**Description**

`StoreAndForwardMessage` is a persistence-ignorant POCO that maps directly to the
`sf_messages` table and is also carried across the network inside `ReplicationOperation`
(replicated to the standby node over Akka remoting). CLAUDE.md "Code Organization" states
that entity classes are persistence-ignorant POCOs in Commons and that message contracts
follow additive-only evolution. Because this type doubles as a replication wire contract
but lives in the component assembly, it is not co-located with the other Commons
entities and its evolution is not governed by the additive-only message-contract rule.
This is a borderline case (the type is site-local), but the cross-node use via
`ReplicationOperation` makes it a de-facto message contract.

**Recommendation**

Either move `StoreAndForwardMessage` (and `ReplicationOperation`) into the Commons
`Entities`/`Messages` hierarchy so they are governed by the contract-evolution rules, or
introduce a separate DTO for replication and keep `StoreAndForwardMessage` purely as the
local persistence model. Document the decision.

**Resolution**

_Unresolved._

### StoreAndForward-013 — Critical paths lack test coverage: retry-due timing, replication-from-active, and the actor bridge

| | |
|--|--|
| Severity | Medium |
| Category | Testing coverage |
| Status | Open |
| Location | `tests/ScadaLink.StoreAndForward.Tests/` (whole directory); `src/ScadaLink.StoreAndForward/StoreAndForwardStorage.cs:101`; `src/ScadaLink.StoreAndForward/ParkedMessageHandlerActor.cs` |

**Description**

The existing tests cover storage CRUD and the service happy/failure paths well, but
three important behaviours are untested: (1) the retry-due time filter in
`GetMessagesForRetryAsync` — every service test sets `DefaultRetryInterval = TimeSpan.Zero`,
so the `julianday` elapsed-time comparison (the most error-prone SQL in the module) is
never exercised with a non-zero interval; a message that is *not yet due* should be
skipped, and that is never verified. (2) Replication from the active side — no test
asserts that an enqueue/remove/park causes a `Replicate*` call (this is exactly the gap
behind finding 001; a test would have caught it). (3) `ParkedMessageHandlerActor` has no
test at all — the Query/Retry/Discard request-to-response mapping and the
`ExtractMethodName` JSON parsing are unverified, including the malformed-JSON branch.

**Recommendation**

Add tests for: a non-zero retry interval where a recently-attempted message is excluded
and an older one is included; active-side replication invocation per operation type
(once finding 001 is fixed); and `ParkedMessageHandlerActor` using `Akka.TestKit`,
including `ExtractMethodName` for `MethodName`, `Subject`, missing-property and
invalid-JSON payloads.

**Resolution**

_Unresolved._

### StoreAndForward-014 — Storage does not create its SQLite database directory

| | |
|--|--|
| Severity | Medium |
| Category | Error handling & resilience |
| Status | Resolved |
| Location | `src/ScadaLink.StoreAndForward/StoreAndForwardStorage.cs:26` |

**Found 2026-05-16** while verifying the store-and-forward fixes — this defect was
not part of the original baseline review.

**Description**

`StoreAndForwardStorage.InitializeAsync` opened a `SqliteConnection` against the
configured `SqliteDbPath` (default `./data/store-and-forward.db`) without ensuring the
parent directory exists. SQLite creates the database *file* on demand but not its
*directory*, so when `data/` does not already exist the connection fails to open with
`SQLite Error 14: 'unable to open database file'`. Every site-host boot therefore failed
in any environment whose working directory has no `data/` folder — the cause of the six
failing `SiteActorPathTests` (the host's `RegisterSiteActors` aborts at
`StoreAndForwardService.StartAsync`). Production masked it because `data/` is created by
the Docker image / deployment.

**Recommendation**

Create the parent directory of a file-backed SQLite database before opening it.

**Resolution**

Resolved 2026-05-16. `InitializeAsync` now calls a new `EnsureDatabaseDirectoryExists`
helper that parses the connection string with `SqliteConnectionStringBuilder` and, for a
file-backed database, creates the parent directory if it is missing (in-memory databases
and bare filenames are skipped). Regression test
`InitializeAsync_FileInMissingDirectory_CreatesDirectory` fails against the pre-fix code;
all six `SiteActorPathTests` now pass. Fixed by the commit whose message references
`StoreAndForward-014`.