During coalesced link refresh, the server may return links for chunks
not yet in the provider's map (newly-discovered chunks beyond
highestKnownChunkIndex). Previously these were silently skipped.

Now: create new chunks from refresh response links, update
highestKnownChunkIndex, and set endOfStreamReached from the response's
hasMore flag. Follows the per-chunk state-machine reconciliation from
the bounded SEA API spec.

Co-authored-by: Isaac
Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

Fixes 3 gaps found by comparing with the legacy ChunkDownloadTask:

1. Outer catch(Throwable) + exception chaining in finally: uncaught
   exceptions were lost — the finally block created a generic exception
   without the original cause. Now captures uncaughtException and chains
   it, matching ChunkDownloadTask's pattern.

2. Thread context propagation: download threads had no connection context
   or statementId for telemetry/logging. Now captures caller's context
   via DatabricksThreadContextHolder and clears in finally.

3. Download timing: added task-level timing log (totalMs, retries)
   matching ChunkDownloadTask's diagnostics.

Also includes the RuntimeException catch (parity with PR databricks#1302).

Co-authored-by: Isaac
Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

P0-1: Remove redundant chunk.setStatus(DOWNLOAD_FAILED) in inner catch
— defer entirely to finally block. Fixes StreamingChunkDownloadTaskTest.

P0-2: Add NEXT_CHANGELOG.md entry under ### Added for UseBoundedSeaApi.

P1-1: Call triggerDownloads() after reconciliation creates new chunks
from refresh response — prevents newly-discovered chunks sitting PENDING.

P1-2/P1-3: Un-gated changes (new chunk creation, EOS from refresh,
triggerDownloads) are intentional parity fixes for all
EnableStreamingChunkProvider=1 users. EnableStreamingChunkProvider
defaults to off, so default users are unaffected.

P1-4: Revert RuntimeException from inner catch — DatabricksError is
caught by outer catch(Throwable) and fails immediately (no retry),
matching ChunkDownloadTask behavior exactly. NPE/ISE won't be retried.

P2-1: Always send row_offset (even 0 for chunk 0) when bounded SEA
enabled — explicit is safer than relying on server default.

P2-3: Update nextLinkFetchIndex after reconciliation to avoid prefetch
thread re-fetching chunks already discovered via refresh.

P2-5: Add "Requires server support" to connection property help text.

Co-authored-by: Isaac
Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

…on updates

P1-1: Use putIfAbsent in createChunkFromLink to prevent double row-count
and chunk replacement when called concurrently from prefetch and download
threads. Without this, a race could leave a CompletableFuture that is
never completed, causing consumer hangs.

P1-2: Bundle nextLinkFetchIndex + nextRowOffsetToFetch into an immutable
FetchPosition holder updated via volatile reference. This ensures the
prefetch thread always reads a consistent (index, rowOffset) pair, which
matters for bounded SEA API where row_offset is used by the server.

P2-1: Bump reconciliation failure log from DEBUG to WARN for production
visibility.

Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

Co-authored-by: Isaac
Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

…ownload=0)

When UseBoundedSeaApi=1 AND EnableQueryResultDownload=0:

1. Session conf: sends can_cloud_download=false to tell server to inline
   results instead of using cloud storage.

2. Execute request: uses ARROW_STREAM + INLINE_OR_EXTERNAL_LINKS (instead
   of JSON_ARRAY + INLINE). This gets Arrow IPC data in the attachment
   field rather than JSON in data_array.

3. New SeaInlineArrowChunkProvider: lazily fetches inline Arrow chunks via
   GetResultData with row_offset. Similar to Thrift's
   LazyThriftInlineArrowResult but using SEA's ArrowIPC format in the
   attachment field. Decompresses via DecompressionUtil, creates
   ArrowResultChunk per chunk. End-of-stream when nextChunkIndex is null.

4. getResultChunksData overload: accepts rowOffset parameter, appends
   ?row_offset=N when bounded SEA is enabled.

Without UseBoundedSeaApi=1, EnableQueryResultDownload=0 continues to use
the existing JSON_ARRAY inline format (unchanged).

Co-authored-by: Isaac
Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

…oring ThriftStreamingProvider

Replace synchronous lazy-fetch design with background prefetch thread
using ReentrantLock + Condition signaling, ConcurrentHashMap for indexed
chunk storage, and explicit chunksInMemory tracking for backpressure.
This ensures consistency with ThriftStreamingProvider's design patterns.

Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

Co-authored-by: Isaac
Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

Document why nextFetchChunkIndex and nextFetchRowOffset are safe as
separate AtomicLongs (single prefetch thread writer) unlike
StreamingChunkProvider which needs a bundled FetchPosition holder.

Signed-off-by: Gopal Lal <gopal.lal@databricks.com>

Co-authored-by: Isaac
Signed-off-by: Gopal Lal <gopal.lal@databricks.com>