fix(transport)+test: close deferred-work register rows 58, 60, 65 (Low-tail sweep)

Row 65 — trigger FIRED, fixed. BundleUnlockAttempt rows were pruned only by
nodes actively serving unlocks, so a cluster that stopped importing entirely
left its last rows in place forever. The row's own trigger was "Transport gains
a central scheduler a prune tick could ride" — it has one now:
BundleSessionEvictionService, the central-only 1-minute session sweep added by
the transport-session fixes. It now also ticks the new
IBundleUnlockAttemptWindow.PruneAsync. Both drivers share the ONE 15-minute CAS
cadence claim, so ticking on a 1-minute sweep costs at most one bounded DELETE
per quarter hour and the two can never prune concurrently. PruneAsync is
contractually non-throwing (unlike TryRegisterAttemptAsync, which throws and
lets the limiter own the fail-open decision) — a background sweep has no caller
to report to and must survive a database blip. Both Transport docs updated;
three tests in FleetWideUnlockThrottleTests, incl. the retention sweep running
with no unlock traffic anywhere in the test, which is the whole point.

Row 58 — closed with a pin, and the OLD pin was found blind. The existing
Deployments_Dispose_UnsubscribesFromNotifier is behavioural: it is satisfied by
the _disposed guard alone, so deleting the `-=` from Dispose leaves it green
while the singleton notifier keeps a delegate pinning the circuit-scoped
component. Verified by mutation — with the `-=` removed the old test still
passed and only the two new pins failed. Added:
  * a structural retention pin asserting the notifier's invocation list actually
    shrinks to 0 on Dispose;
  * a source scan pinning the row's stated TRIGGER ("a second page subscribes,
    or any subscriber lands without a matching Dispose unsubscribe") — every
    CentralUI file that subscribes must also unsubscribe and be IDisposable,
    with a non-empty guard so the scan cannot pass vacuously.

Row 60 — closed with a pin, and the register's premise corrected.
GlobalArtifactSnapshot does NOT hold EF-materialised entities: it is projected
into pure Commons DTOs before the fan-out. The EF state that genuinely crosses
into the parallel send is the tracked Site entities the lambda closes over, and
that is safe only while lazy loading is off. Pinned in ConfigurationDatabase
.Tests two ways — no proxies extension on the built DbContextOptions, and no
EntityFrameworkCore.Proxies package reference anywhere in the repo (the
structurally stronger half: UseLazyLoadingProxies() is not a compilable call
without it). The real invariant is now spelled out at the fan-out itself.

Row 59 — no product change; added the missing tripwire remark stating the
constraint the safety argument rests on (serial by design, one scoped DbContext,
do not Task.WhenAll; if it gets slow the fix is a scope PER SITE, not a shared
scope under fan-out). Line numbers in the register have drifted to :528-545.

Rows 36, 37, 64: verification only, no change — see the sweep report.

docs/components/Transport.md

@@ -138,7 +138,7 @@ The two levels interact in a specific order, and it is load-bearing (register ro
 
 - **It fails OPEN.** A store error logs a Warning naming the client and the cap, and *permits* the attempt. This is a best-effort rate cap layered over a surface that is already authenticated and Administrator-gated, so the availability of legitimate imports outranks the cap — a database blip must not make the import wizard unusable. Falling back to the old process-local window instead was considered and rejected: it would still block a legitimate operator's 11th import during an outage, which is precisely the outcome fail-open exists to prevent. The Warning is the operator's signal that the fleet cap is not currently being enforced.
 - **Accuracy is deliberately approximate.** The count-then-insert pair is not atomic across the two nodes, so genuinely simultaneous attempts can overshoot the cap by up to the number in flight. A conditional single-statement insert would close that at the cost of provider-specific raw SQL; the overshoot is bounded and self-corrects within the window.
-- **Rows are pruned opportunistically, not by a background service.** Transport owns no central scheduler, so the prune rides the unlock path: claimed by CAS at most once per 15 minutes per process, bounded to 1000 rows per pass, retaining `2 ×` the window so a prune racing a count can never delete a row the count still needs, and wrapped in a try/catch that logs and swallows — retention maintenance must never fail the operator's actual unlock. This is the same shape `DeploymentService` uses for terminal deployment records.
+- **Rows are pruned by two drivers sharing one cadence claim (register row 65).** The prune is claimed by CAS at most once per 15 minutes per process, bounded to 1000 rows per pass, retaining `2 ×` the window so a prune racing a count can never delete a row the count still needs, and wrapped in a try/catch that logs and swallows — retention maintenance must never fail the operator's actual unlock. This is the same shape `DeploymentService` uses for terminal deployment records. It originally rode the **unlock path only**, on the reasoning that Transport owned no central scheduler; the consequence was that a cluster which stopped importing entirely left its last rows in place indefinitely, since only a node actively serving unlocks ever pruned. Transport has since gained exactly such a scheduler — `BundleSessionEvictionService`, the central-only 1-minute session sweep — so it now also ticks `IBundleUnlockAttemptWindow.PruneAsync`, and the table converges whether or not anyone is importing. Because both drivers take the **same** claim, ticking on the 1-minute sweep costs at most one bounded DELETE per quarter hour and the two can never prune concurrently. `PruneAsync` is contractually non-throwing (unlike `TryRegisterAttemptAsync`, which throws and lets the limiter own the fail-open decision) — a background sweep has no caller to report to and must survive a database blip.
 
 Behaviour the operator sees is otherwise unchanged: same config key, same semantics ("attempts, not failures"), same ordering relative to the per-bundle lockout, same `<unattributed>` bucket.
 

docs/requirements/Component-Transport.md

@@ -338,7 +338,7 @@ Imports are all-or-nothing per bundle. A bundle either applies fully or not at a
 - **Passphrase never persisted.** It lives only inside the export/import service call path and is discarded after use. No environment variable, no log line.
 - **Failed-unlock rate limit:** per-bundle 3-strike lockout (keyed by `ContentHash`, server-side); per-IP-per-hour cap (`MaxUnlockAttemptsPerIpPerHour`, default 10, configurable) to deter brute force against a stolen bundle. Each failed attempt produces a `BundleImportUnlockFailed` audit row. The two are evaluated in order — bundle lockout, then per-IP throttle, then decrypt — so hammering an already-locked bundle spends no hourly budget, and a throttle denial (`BundleUnlockThrottledException`, fail-closed with a curated message on both the UI and CLI surfaces) burns **no** per-bundle strike, since it never reaches the decryptor. Both typed exceptions live in `Commons/Types/Transport/` — `ManagementService` references Commons but not Transport, so a type declared beside the importer cannot be curated by `MapFault` and reaches the CLI as a generic internal-error message (register rows 44 and 51). Successful unlocks count against the hourly budget too: it caps attempts, not failures. The client key is the caller's remote IP — from `CircuitClientAddressProvider` (snapshotted at Blazor circuit establishment, since `HttpContext` is unavailable mid-circuit) for the Central UI, and from `ManagementEnvelope.ClientIp` for the CLI; a caller with no attributable address shares a single stricter bucket rather than bypassing the throttle. Per-client isolation is only real behind a ForwardedHeaders-trusted proxy, the same caveat the login throttle carries in `Component-Security.md`. (Register row 44: the limiter existed, was configured and unit-tested, but was never invoked until this was wired.)
 
-**The per-IP window is fleet-wide and restart-durable (register row 52).** It is persisted in the central MS SQL `BundleUnlockAttempt` table — the HA store both central nodes already share — rather than in each node's process memory, so the documented cap is a fleet cap and not N × the cap, and a node restart no longer clears it. `BundleUnlockRateLimiter` still owns the policy; `SqlBundleUnlockAttemptWindow` is the store behind the `IBundleUnlockAttemptWindow` seam, and `IBundleUnlockAttemptRepository` (Commons interface, ConfigurationDatabase implementation) is the persistence. Three deliberate trades: **the throttle fails OPEN** — a store error logs a Warning and permits the attempt, because this is a best-effort rate cap over an already-authenticated Administrator-gated surface and the availability of legitimate imports outranks the cap; **accuracy is approximate** — count-then-insert is not atomic across nodes, so simultaneous attempts can overshoot by the number in flight, which is bounded and self-corrects within the window; and **rows are pruned opportunistically on the unlock path** (CAS-claimed at most every 15 minutes per process, bounded to 1000 rows, retaining twice the window, best-effort) rather than by a new background service, since Transport owns no central scheduler.
+**The per-IP window is fleet-wide and restart-durable (register row 52).** It is persisted in the central MS SQL `BundleUnlockAttempt` table — the HA store both central nodes already share — rather than in each node's process memory, so the documented cap is a fleet cap and not N × the cap, and a node restart no longer clears it. `BundleUnlockRateLimiter` still owns the policy; `SqlBundleUnlockAttemptWindow` is the store behind the `IBundleUnlockAttemptWindow` seam, and `IBundleUnlockAttemptRepository` (Commons interface, ConfigurationDatabase implementation) is the persistence. Three deliberate trades: **the throttle fails OPEN** — a store error logs a Warning and permits the attempt, because this is a best-effort rate cap over an already-authenticated Administrator-gated surface and the availability of legitimate imports outranks the cap; **accuracy is approximate** — count-then-insert is not atomic across nodes, so simultaneous attempts can overshoot by the number in flight, which is bounded and self-corrects within the window; and **rows are pruned best-effort** (CAS-claimed at most every 15 minutes per process, bounded to 1000 rows, retaining twice the window). That prune originally rode the unlock path alone, since Transport owned no central scheduler — which meant a cluster that stopped importing entirely left its last rows in place indefinitely. **Register row 65 closed that**: Transport's central-only `BundleSessionEvictionService` sweep now also ticks `IBundleUnlockAttemptWindow.PruneAsync`, sharing the one 15-minute cadence claim with the unlock path, so the table converges with no import traffic at all. `PruneAsync` never throws (the fail-open policy stays with the limiter on the attempt path).
 - **Bundle size cap** on upload (default 100 MB, configurable) to bound memory.
 - **In-transit:** existing HTTPS to the Central UI; no new channel.
 - **Audit trail is the chain of custody.** Every export, every import (including aborted ones at validation), and every unlock failure is audit-logged with source env, content hash, encrypted yes/no, and artifact summary.