Implements control over which cluster node runs a trigger, for use cases
like in-memory caching between job runs.

Core feature:
- TriggerBuilder.WithPreferredNode() API for pinning triggers to nodes
- Manual pin (specific instance id) and auto-pin ("*" sentinel)
- SQL-level acquisition filter with checkin-time-aware liveness check
- INextVersion* internal interface pattern — no breaking changes on 3.x
- Optional PREFERRED_NODE column on QRTZ_TRIGGERS, probed at startup

Auto-pin prefix ("auto:"):
- Auto-pinned triggers stored as "auto:nodeA" to distinguish from
  explicit pins ("nodeA") in the PREFERRED_NODE column
- Steal-on-failover: a node firing a trigger whose raw pin is "*" or
  "auto:<other-node>" claims it as "auto:<self>" — acquisition only
  releases a foreign auto-pin via the liveness fallback, so the claim
  converges to a live, eligible node. Also fixes a write-back race
  where a fire acquired before ClusterRecover's reset would persist
  the stale "auto:<deadNode>" back after the state row was deleted,
  permanently pinning the trigger to a nonexistent node
- Dirty-flag persistence: UpdateTrigger only writes PREFERRED_NODE
  when the value was changed on that instance (setter, builder,
  deserialization, auto-pin claim) — never when merely loaded at
  acquire time. Prevents routine fires from reverting concurrent
  UpdateTriggerDetails re-pins or ClusterRecover resets
- ClusterRecover resets auto-pins to "*" sentinel so any eligible node
  can claim — correctly handles execution group limits
- Explicit pins preserved through failover; original node reclaims
  when it returns
- GetTriggerBuilder() strips prefix, produces explicit pin for
  clone/reschedule flows
- Validation rejects "auto:" anywhere in user-supplied values
- Clustered schedulers throw at startup if InstanceId contains "auto:"
  (gated on HasPreferredNodeColumn)
- SetPreferredNodeRaw() internal bypass for Quartz's own writes

Additional:
- PreferredNode/ExecutionGroup added to JSON scheduling configuration
- Cached INextVersionDelegate field for hot-path performance
- Missing-column log level changed from DEBUG to WARN
- WARN when UpdateTriggerDetails cannot persist a preferred-node change
  because the column is missing
- MySQL acquisition SQL uses the IDX_{1} unqualified-prefix FORCE INDEX
  pattern from #3086 (schema-qualified table prefixes)

Tests:
- Unit tests incl. RAMJobStore metadata-only behavior
- SQLite: old-schema graceful degradation, ghost-node pin still fires,
  stale-auto-pin steal regression, misfire preserves pin,
  acquire-without-fire does not claim
- PostgreSQL cluster (2-3 real nodes): exclusion, failover, write-back
  race regression, execution-group-saturated survivor, runtime re-pin
  redirect; base class unbinds nodes from SchedulerRepository (name-only
  non-proxy lookup previously collapsed concurrent nodes to one),
  cleans DB state per test, and uses a short idleWaitTime so nodes
  notice remote changes quickly

Co-Authored-By: Claude Fable 5 <[email protected]>

Implements control over which cluster node runs a trigger, for use cases
like in-memory caching between job runs.

Core feature:
- TriggerBuilder.WithPreferredNode() API for pinning triggers to nodes
- Manual pin (specific instance id) and auto-pin ("*" sentinel)
- SQL-level acquisition filter with checkin-time-aware liveness check
- INextVersion* internal interface pattern — no breaking changes on 3.x
- Optional PREFERRED_NODE column on QRTZ_TRIGGERS, probed at startup

Auto-pin prefix ("auto:"):
- Auto-pinned triggers stored as "auto:nodeA" to distinguish from
  explicit pins ("nodeA") in the PREFERRED_NODE column
- Steal-on-failover with compare-and-swap: a node firing a trigger whose
  raw pin is "*" or "auto:<other-node>" claims it as "auto:<self>" via a
  conditional UPDATE against the acquire-time value, so concurrent
  re-pins/clears (UpdateTriggerDetails, ClusterRecover reset) win over
  the claim instead of being clobbered
- Dirty-flag persistence: UpdateTrigger writes PREFERRED_NODE only when
  the value was changed on that instance (setter, builder,
  deserialization) — never when merely loaded at acquire time; loading
  clears the flag so blob triggers do not stay permanently dirty. The
  flag is serialized ([OptionalField]) so explicit changes survive
  net472 remoting by-value marshaling
- Builder-built triggers fully define the pin: a definition without
  WithPreferredNode clears a stored pin on replace (matches
  EXECUTION_GROUP semantics)
- ClusterRecover resets auto-pins to "*" sentinel so any eligible node
  can claim — correctly handles execution group limits
- Explicit pins preserved through failover; original node reclaims
  when it returns
- GetTriggerBuilder() strips prefix, produces explicit pin for
  clone/reschedule flows
- Validation rejects "auto:" anywhere in user-supplied values
- Clustered schedulers throw at startup if InstanceId contains "auto:"
  (gated on HasPreferredNodeColumn)
- UpdateTriggerDetails throws when a pin change cannot be persisted
  because the column is missing (instead of reporting success)

Robustness (deep-review round):
- Optional-column probing tolerates connectivity failures: Initialize
  no longer requires a live database and no longer latches transient
  errors as "column missing" — the probe retries on the first database
  operation and at Start()
- Blob-stored ExecutionGroup/PreferredNode values are preserved (not
  scrubbed) when the optional column is absent, preventing permanent
  erasure on blob re-serialization
- Positional-provider safety: named parameters appear exactly once per
  statement (liveness subquery correlates on t.SCHED_NAME) and new
  statements add parameters in SQL token order
- preferredNode is [OptionalField] so pre-upgrade binary blobs still
  deserialize
- One-time WARN when a custom delegate's SelectTriggerToAcquire
  override is bypassed by the preferred-node-aware overload

Additional:
- PreferredNode/ExecutionGroup added to JSON scheduling configuration
- Cached INextVersionDelegate field for hot-path performance
- Missing-column log level changed from DEBUG to WARN
- MySQL acquisition SQL uses the IDX_{1} unqualified-prefix FORCE INDEX
  pattern from #3086 (schema-qualified table prefixes)

Tests:
- Unit tests incl. RAMJobStore metadata-only behavior
- SQLite: old-schema graceful degradation, ghost-node pin still fires,
  stale-auto-pin steal regression, misfire preserves pin,
  acquire-without-fire does not claim
- PostgreSQL cluster (2-3 real nodes): exclusion, failover, write-back
  race regression, execution-group-saturated survivor, runtime re-pin
  redirect; base class unbinds nodes from SchedulerRepository (name-only
  non-proxy lookup previously collapsed concurrent nodes to one),
  cleans DB state per test, and uses a short idleWaitTime so nodes
  notice remote changes quickly

Co-Authored-By: Claude Fable 5 <[email protected]>

Implements control over which cluster node runs a trigger, for use cases
like in-memory caching between job runs.

Core feature:
- TriggerBuilder.WithPreferredNode() API for pinning triggers to nodes
- Manual pin (specific instance id) and auto-pin ("*" sentinel)
- SQL-level acquisition filter with checkin-time-aware liveness check
- INextVersion* internal interface pattern — no breaking changes on 3.x
- Optional PREFERRED_NODE column on QRTZ_TRIGGERS, probed at startup

Auto-pin prefix ("auto:"):
- Auto-pinned triggers stored as "auto:nodeA" to distinguish from
  explicit pins ("nodeA") in the PREFERRED_NODE column
- Steal-on-failover with compare-and-swap: a node firing a trigger whose
  raw pin is "*" or "auto:<other-node>" claims it as "auto:<self>" via a
  conditional UPDATE against the acquire-time value, so concurrent
  re-pins/clears (UpdateTriggerDetails, ClusterRecover reset) win over
  the claim instead of being clobbered
- Dirty-flag persistence: UpdateTrigger writes PREFERRED_NODE only when
  the value was changed on that instance (setter, builder,
  deserialization) — never when merely loaded at acquire time; loading
  clears the flag so blob triggers do not stay permanently dirty. The
  flag is serialized ([OptionalField]) so explicit changes survive
  net472 remoting by-value marshaling
- ClusterRecover resets auto-pins to "*" sentinel so any eligible node
  can claim — correctly handles execution group limits
- Explicit pins preserved through failover; original node reclaims
  when it returns
- GetTriggerBuilder() strips prefix, produces explicit pin for
  clone/reschedule flows
- Validation of the reserved "auto:" token shared across TriggerBuilder,
  AbstractTrigger, and TriggerDetailsUpdate (single helper)
- Clustered schedulers throw at startup if InstanceId contains "auto:"
  (gated on HasPreferredNodeColumn)

Robustness (deep-review rounds):
- Optional-column probing tolerates connectivity failures: a canary
  query against an always-present column distinguishes "table
  unreachable" (transient — retried, not latched) from "column missing"
  (latched). Initialize no longer requires a live database; the probe
  retries on the first database operation and at Start(), serialized by
  a semaphore so concurrent early operations don't double-probe
- Custom delegates that override the legacy acquisition SQL but not the
  preferred-node overload keep acquisition on the legacy path (their
  override still runs; preferred-node filtering not applied, warned)
  instead of being silently bypassed
- UpdateTriggerDetails degrades gracefully (warns, applies other
  changes) when the column is missing, instead of throwing and losing
  bundled description/priority updates
- Blob-stored ExecutionGroup/PreferredNode values are preserved (not
  scrubbed) when the optional column is absent
- Positional-provider safety: named parameters appear exactly once per
  statement (liveness subquery correlates on t.SCHED_NAME) and new
  statements add parameters in SQL token order
- preferredNode is [OptionalField] so pre-upgrade binary blobs still
  deserialize
- Acquisition SQL no longer projects/materializes the unread
  PREFERRED_NODE column (filtered in WHERE only)

Additional:
- PreferredNode/ExecutionGroup added to JSON scheduling configuration
- Cached INextVersionDelegate field for hot-path performance
- Missing-column log level changed from DEBUG to WARN
- MySQL acquisition SQL uses the IDX_{1} unqualified-prefix FORCE INDEX
  pattern from #3086 (schema-qualified table prefixes)

Tests:
- Unit tests incl. RAMJobStore metadata-only behavior
- SQLite: old-schema graceful degradation, ghost-node pin still fires,
  stale-auto-pin steal regression, misfire preserves pin,
  acquire-without-fire does not claim; CreateScheduler unbinds from
  SchedulerRepository so a failing test cannot leak a scheduler
- PostgreSQL cluster (2-3 real nodes): exclusion, failover, write-back
  race regression, execution-group-saturated survivor, runtime re-pin
  redirect; base class unbinds nodes from SchedulerRepository (name-only
  non-proxy lookup previously collapsed concurrent nodes to one),
  cleans DB state per test, uses a short idleWaitTime so nodes notice
  remote changes quickly

Co-Authored-By: Claude Fable 5 <[email protected]>