feat: JFR events for open, HTTP range, segment decode/encode, chunk filtering and writer back-pressure

Start by reading the reader and writer entry points named in the issue: VortexReader.open, VortexHttpReader.open, SerializedArrayDecoder.decode, ScanIterator, and VortexWriter. Add the JFR event classes and wire them into those paths as specified, then use the named tests to check event fields, filtering counts, and back-pressure duration. Done when the event tests pass and the reference guide, how-to guide, changelog, and module map are updated.

Goal

Emit JDK Flight Recorder events from the reader and writer, so a JFR recording shows where a read or write spends its time (file open/mapping, HTTP range fetches, per-segment decode/encode, chunk pruning, writer back-pressure) with no extra dependency and near-zero cost when recording is off.

Events

New package core/src/main/java/io/github/dfa1/vortex/core/jfr/ (with package-info.java). Each event is a public class extending jdk.jfr.Event, annotated @Name("io.github.dfa1.vortex.X"), @Label, @Category({"Vortex", …}), @Description and @StackTrace(false). Public fields get /// docs (doclint is all); byte sizes get @DataAmount.

Reader wiring

Pattern: event.begin(), set fields, commit().

• VortexReader.open(path, registry, layoutRegistry): FileOpened spans the whole open and is committed after parse; FileMapping wraps channel.map. A failed open emits no FileOpened.
• VortexHttpReader.open(uri, registry, layoutRegistry, client): FileOpened begins before fetchTail and is committed after the reader is built.
  • New private sendRange(req, client, uri, offset, length) wraps client.send in try/finally and always commits HttpRange, with status = 0 on failure. fetchTail calls it with offset -1 and TAIL_SIZE.bytes(); fetchRange with from and to - from + 1.
  • New package-private static String describe(URI) returning scheme://host[:port] + raw path. It drops user info, query and fragment, because presigned URLs carry credentials. Used for every source field.
• SerializedArrayDecoder.decode: begin at the top; after registry.decode(ctx) set encoding = node.encodingId().id(), rowCount, segmentSize = segLen, then commit. Covers both readers.
• ScanIterator:
  • New counters chunksPruned, chunksWithoutMatch, chunksReturned.
  • In hasNext(), split canPruneChunk(...) || !selectMatchingRows(...) into two branches that each count, chunkIndex++; continue.
  • Increment chunksReturned in next().
  • close() calls emitChunkFilterEvent(), a no-op when filter == null || chunks == null.

Writer wiring (VortexWriter)

• encodeSegment: begin at the top; before returning, if (event.shouldCommit()) fill encoding from result.rootNode().encodingId().id(), bufferCount, encodedSize, then commit. Failures emit nothing.
• drainSegments: replace pending.poll().join() with a static awaitHead(head, pending.size() + 1). It joins directly if head.isDone(); otherwise begins a SegmentWaitEvent, joins in a try, commits in the finally.
• No change to the writer's default behavior. (The default executor runs tasks on the caller, so SegmentWait never appears there.)

Tests

Given/When/Then throughout. For events committed on other threads, assert presence and field values, never exact counts.

• core/src/test/.../core/testing/JfrRecorderTest.java: abstract base; a RecordingStream with setReuse(false) starts in @BeforeEach; offers awaitEvents() and events(name).
• reader VortexReaderJfrTest (fixture primitives.vortex): FileOpened and FileMapping fields; a too-small file emits no FileOpened.
• reader VortexHttpReaderJfrTest (mocked HttpClient): tail HttpRange (offset -1, status 206); source has no credentials or query; status 0 when send throws IOException.
• writer JfrEventsTest (uses LegacyZoneMaps.OPTIONS, same package):
  • SegmentEncoded present with non-negative sizes (a vortex.sequence segment has 0 buffers).
  • Default executor: no SegmentWait.
  • Executor sleeping 200 ms before each task: SegmentWait duration > 50 ms and segmentsInFlight > 0.
  • Chunks 1..50, 51..100, 101..150 with eq(id, 75) → totals 3/2/0/1 and 1 row.
  • Single chunk {1, 100} with eq(id, 50) → 1/0/1/0 and 0 rows.
  • Unfiltered scan emits no ChunkFilter.

Docs (same commit)

• docs/reference.md: "JFR events" section (event table) + TOC entry.
• docs/how-to.md: "Find where a read or write spends its time (JFR)".
• CHANGELOG.md: ## [Unreleased] → Added.
• AGENTS.md: core.jfr line in the module map.

Verify

• ./mvnw test -pl core,reader,writer -am
• ./mvnw verify -DskipTests (Javadoc lint + checkstyle)