feat: DTail fork — server/client feature development

Squashed development of the snonux/dtail fork's product code (internal/, cmd/)
since diverging from mimecast/dtail. Major areas:

- Read/output path: the former "turbo" channel-less path is now the single,
  default server-side read/output path for cat/grep/tail and MapReduce; the old
  channel-based path and its config/env toggles were removed.
- MapReduce: single aggregate implementation (server + serverless) fed directly
  by a processor pipeline, with input-exhausted finalization via the shutdown
  coordinator; high-concurrency and data-race fixes.
- Journal source reads (journal:unit.service) via journalctl, Linux-gated behind
  a journal-v1 capability.
- Auth-key fast reconnect: in-memory per-user public-key cache with TTL/max-keys,
  registered over an authenticated session (AUTHKEY), checked before
  authorized_keys.
- Interactive query reload (--interactive-query) with SESSION START/UPDATE
  generation boundaries and capability negotiation.
- Client-side deadlines: --timeout / --shutdownAfter as context deadlines;
  follow shutdown handling.
- Client logging: diagnostics-only daily log by default, opt-in payload tee via
  --log-payload.
- Numerous correctness fixes (buffer-pool double-recycle races, EOF-sentinel
  leaks, glob-expansion cap, TOCTOU in CSV parsing) with accompanying unit tests.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

1 files changed, 54 insertions, 0 deletions

diff --git a/internal/server/handlers/line_writer_alloc_test.go b/internal/server/handlers/line_writer_alloc_test.go
new file mode 100644
index 0000000..fc82e81
--- /dev/null
+++ b/internal/server/handlers/line_writer_alloc_test.go
@@ -0,0 +1,54 @@
+package handlers
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/mimecast/dtail/internal/io/dlog"
+	"github.com/mimecast/dtail/internal/io/pool"
+)
+
+// nopLineWriter is a zero-allocation LineWriter used to isolate the
+// per-line trace-guard overhead in the alloc test below.
+type nopLineWriter struct{}
+
+func (nopLineWriter) WriteLineData(lineContent []byte, lineNum uint64, sourceID string) error {
+	return nil
+}
+func (nopLineWriter) WriteServerMessage(message string) error { return nil }
+func (nopLineWriter) Flush() error                            { return nil }
+
+// TestDirectLineProcessorProcessLineNoAllocWhenTraceOff locks in task 1t0: the
+// per-line hot path must not allocate when trace logging is off (the default
+// production/output level). Before the fix, ProcessLine unconditionally built a
+// []interface{} and boxed lineCount/lineNum (runtime.convT64) and sourceID
+// (convTstring) on every line even though dlog.Trace early-returns — that call
+// site was ~98% of all allocated objects in the output serverless dcat profile.
+// The dlog.Server.TraceEnabled() guard elides all of it.
+func TestDirectLineProcessorProcessLineNoAllocWhenTraceOff(t *testing.T) {
+	orig := dlog.Server
+	// Zero-value DLog has maxLevel == None, i.e. trace disabled.
+	dlog.Server = &dlog.DLog{}
+	t.Cleanup(func() { dlog.Server = orig })
+
+	if dlog.Server.TraceEnabled() {
+		t.Fatal("precondition failed: trace must be disabled for this test")
+	}
+
+	p := NewDirectLineProcessor(nopLineWriter{}, "globID")
+
+	allocs := testing.AllocsPerRun(1000, func() {
+		// Mirror the real hot path: a pooled buffer that ProcessLine recycles.
+		buf := pool.BytesBuffer.Get().(*bytes.Buffer)
+		buf.Reset()
+		buf.WriteString("some representative log line content")
+		if err := p.ProcessLine(buf, 42, "sourceID"); err != nil {
+			t.Fatalf("ProcessLine: %v", err)
+		}
+	})
+
+	if allocs != 0 {
+		t.Fatalf("ProcessLine allocated %v objects/run with trace off; want 0 "+
+			"(the per-line trace guard should elide all interface boxing)", allocs)
+	}
+}