Delete the oldest report files to stay under the size cap (closes #54)

When a report would take the report files past DATA_DIR_MAX_BYTES,
reportbuf now deletes the oldest report files until it fits, and does
the same at start when files left by an earlier run are already past
it. A file joins the files that may be deleted only once it is
completely written, so a file still being written is never deleted. A
report is refused with 507 only when the reports waiting to be written
fill the cap on their own, and then no file is deleted. The reports of
a failed write stop counting, and the part of its file written is
removed. A file whose deletion fails keeps counting; one already
deleted by hand counts as freed.

Model: opus-5-5

.dockerignore

@@ -4,3 +4,6 @@ tmp
 .DS_Store
 *.log
 .claude
+
+# .git is sent so the build can stamp the version, without its config.
+.git/config

Dockerfile

@@ -20,7 +20,10 @@ RUN make lint
 # golang:1.25-alpine (2026-02-27)
 FROM golang:1.25-alpine@sha256:f6751d823c26342f9506c03797d2527668d095b0a15f1862cddb4d927a7a4ced AS builder
 
-RUN apk add --no-cache make
+# gcc and musl-dev are for make test: its race detector needs cgo, which
+# Go turns on by itself once a C compiler is present. make build still
+# sets CGO_ENABLED=0, so the binary stays static.
+RUN apk add --no-cache gcc git make musl-dev
 
 WORKDIR /src
 
@@ -37,11 +40,24 @@ RUN make test
 
 # make build is a shim around backend/script/build, the one definition
 # of the build command:
-# CGO_ENABLED=0 go build -trimpath -ldflags "-s -w -X main.Version=... -X main.Buildarch=..."
+# CGO_ENABLED=0 go build -trimpath -ldflags "-s -w -X main.Version=..."
 # That script reads VERSION from the environment, so it is handed over
 # there rather than as a make variable.
-ARG VERSION=dev
+#
-RUN VERSION="${VERSION}" make build
+# The version is the VERSION build argument when one is given, otherwise
+# `git describe --tags --always` of the repo's .git: the tag on a tagged
+# commit, tag-N-gHASH on a commit after one, the short commit when no
+# tag is reachable. A version that still comes out empty, dev or unknown
+# fails the build. .git goes to /git, not /src/.git, where go build would
+# find it and record VCS details of a work tree holding only backend/.
+COPY .git /git
+ARG VERSION
+RUN version="${VERSION:-$(git --git-dir=/git describe --tags --always)}"; \
+    case "$version" in ""|dev|unknown) \
+        echo "version is '$version' although .git is present" >&2; \
+        exit 1 ;; \
+    esac; \
+    VERSION="$version" make build
 
 # Frontend stage
 # node:22-alpine as of 2026-02-22
@@ -52,8 +68,9 @@ RUN yarn install --frozen-lockfile
 RUN apk add --no-cache git make
 COPY . .
 # make frontend-check is the frontend half of make check (test + lint +
-# fmt-check); its test step is the production yarn build, so this both
+# fmt-check); its test step runs the unit tests, then the production
-# produces dist/ and gates the image on lint/fmt-check/test regressions.
+# yarn build, so this both produces dist/ and gates the image on
+# lint/fmt-check/test regressions.
 # This node stage has neither Go nor Docker; the lint and builder stages
 # above gate the backend half.
 RUN make frontend-check

README.md

@@ -39,21 +39,23 @@ halves, so the root `make check` fails if either one is broken. We provide:
 - `script/bootstrap` — install all dependencies (the pinned node via nvm unless
   one new enough for the frontend's dependencies is installed, yarn via
   corepack, `yarn install --frozen-lockfile`, the pinned Go unless one at least
-  as new as `backend/go.mod` asks for is installed, and the Go modules), linking
+  as new as `backend/go.mod` asks for is installed, the Go modules, and gcc with
-  what it installs itself into `~/.local/bin`, which has to be on `PATH`. It
+  the C library headers unless gcc is installed, for the race detector in
-  installs no Go linter and not Docker: `make lint` runs the linter in Docker
+  `make test`), linking what it installs itself into `~/.local/bin`, which has
+  to be on `PATH`. It installs no Go linter and not Docker: `make lint` runs the
+  linter in Docker
 - `script/setup` — make a fresh clone ready for development: bootstrap plus the
   git pre-commit hook
 - `script/projectname` — print the project name (used for the Docker image tag)
-- `script/test` — run `script/frontend-test`, then the backend's Go tests, both
+- `script/test` — run `script/frontend-test`, then the backend's Go tests, each
-  within one 30-second timeout
+  under its own 30-second timeout
 - `script/lint` — run `script/frontend-lint`, then golangci-lint in Docker, by
   building the lint stage of `Dockerfile` without the cache
 - `script/fmt` — format all files (writes): prettier, then gofmt over `backend/`
 - `script/fmt-check` — check formatting (read-only): prettier, then gofmt
 - `script/check` — run test, lint, and fmt-check
-- `script/frontend-test` — run the production build as the frontend's test (no
+- `script/frontend-test` — run the unit tests in `test/unit/` with Node's
-  unit tests yet)
+  built-in test runner, then the production build
 - `script/frontend-lint` — run prettier in check mode
 - `script/frontend-fmt` — format everything prettier understands (writes)
 - `script/frontend-fmt-check` — check prettier formatting (read-only)
@@ -136,8 +138,14 @@ Local hosts are tracked separately from WAN stats.
 ### Latency measurement
 
 HEAD requests with `mode: 'no-cors'` and `cache: 'no-store'`, timed with
-`performance.now()`. 1-second timeout; anything over 1000ms is clamped to
+`performance.now()`. Each check times out after 80% of the refresh interval (24
-unreachable. IPv4 only.
+seconds at 30 seconds) and is then recorded as a timeout, so a round's checks
+have all finished before the next round is due. When no WAN host answers, a
+recovery probe checks 4 random WAN hosts every half second, giving up the checks
+it started half a second before. As soon as one answers, a new round starts at
+once, as it does after an interval change. A round started early gives up the
+last round's checks if they are still waiting, and that round records nothing,
+so rounds never overlap. IPv4 only.
 
 ### Color coding
 
@@ -212,7 +220,7 @@ What the [upaas](https://git.eeqj.de/sneak/upaas) app for netwatch needs:
     - `REPORTS_PER_MINUTE`, default `60`: reports each client address may send a
       minute
     - `DATA_DIR_MAX_BYTES`, default `1073741824` (1 GiB): the most room the
-      report files may take
+      report files may take; the oldest are deleted to stay under it
     - `CORS_ALLOWED_ORIGINS`, default empty: other origins whose pages may call
       the API
     - `DEBUG`, default `false`: debug logging

TODO.md

@@ -23,6 +23,34 @@ latest run passes.
 
 # Completed Steps
 
+- 2026-10-03: `DATA_DIR_MAX_BYTES` is now how much of the report files is kept
+  (issue #54): when a report would take them past it, the oldest report files
+  are deleted to make room, each deletion logged, and at start files already
+  past it are deleted the same way. A file still being written is never deleted.
+  A report is refused with 507 only when the reports waiting to be written fill
+  the cap on their own, and then no file is deleted. The reports of a failed
+  write stop counting, and the part of its file written is removed. A file that
+  cannot be deleted still counts until the next start; one already deleted by
+  hand counts as freed
+- 2026-10-03: the Go tests run with the race detector and coverage (issue #88):
+  `backend/script/test` runs `go test -timeout 30s -race -cover ./...` and, if
+  that fails, runs it again with `-v` and fails. Go's `-timeout` bounds the
+  tests, not their compile; the root `script/test` no longer puts one 30-second
+  timeout around both halves, which a cold Go build cache could use up on
+  compiling alone. The race detector needs a C compiler: the builder stage of
+  `Dockerfile` has gcc and musl-dev, and `script/bootstrap` installs gcc, with
+  the C library headers on apt and apk, when gcc is missing; the binary is still
+  built with `CGO_ENABLED=0`. New tests cover the health check's answer, a valid
+  report's answer, a report file's exact contents, and the flush when the buffer
+  reaches 10 MiB; the handlers' `TestImport` stub is gone
+- 2026-10-03: each target check times out after 80% of the refresh interval
+  (issue #78), 24 seconds at 30 seconds, where it was capped at 3 seconds. A
+  round started early, after an interval change or when the recovery probe finds
+  a target answering, gives up the last round's checks if they are still
+  waiting, so rounds never overlap; the recovery probe gives up its own checks
+  after half a second. The frontend has its first unit tests, run by
+  `script/frontend-test` with Node's built-in test runner; for them,
+  `index.html` now links `src/styles.css`, which `src/main.js` used to import
 - 2026-09-29: the container sets up its own data directory (issue #75):
   `bin/entrypoint.sh`, still as root, creates `DATA_DIR` if missing and gives it
   and `/data` to the `netwatch` user with mode 750 before starting the backend

backend/README.md

@@ -32,10 +32,13 @@ pattern as the repo root: the targets in `backend/Makefile` are thin shims over
 `test`, `fmt` and `fmt-check`:
 
 - `script/build` — compile the static `netwatch-server` binary with its version
-  and architecture stamped in. The version is `VERSION` from the environment;
+  stamped in. The version is `VERSION` from the environment;
   when that is unset or empty, it falls back to `git describe` inside a git
   checkout, then to `dev`
-- `script/test` — run the Go tests under a 30-second timeout
+- `script/test` — run the Go tests with the race detector and coverage. Go's
+  `-timeout 30s` bounds the tests, not their compile. If they fail, they run
+  again with `-v` for the details, and the script fails. The race detector needs
+  a C compiler
 - `script/lint` — check `.golangci.yml` against its pinned sha256, then run
   golangci-lint. It runs inside the golangci-lint image of the lint stage of
   the root `Dockerfile`; from a checkout, run `make lint` at the repo root,
@@ -80,7 +83,7 @@ Internal packages in `internal/` follow standard Go project layout:
 | `BIND_ADDRESS`         | empty                | IP address to listen on; empty listens on every interface                                                |
 | `PORT`                 | `8080`               | HTTP listen port                                                                                         |
 | `DATA_DIR`             | `./data/reports`     | Directory for compressed reports                                                                         |
-| `DATA_DIR_MAX_BYTES`   | `1073741824` (1 GiB) | Largest total size of the report files in `DATA_DIR`; see [Report limits](#report-limits)                |
+| `DATA_DIR_MAX_BYTES`   | `1073741824` (1 GiB) | Most bytes of report files kept in `DATA_DIR`, oldest deleted first; see [Report limits](#report-limits) |
 | `DEBUG`                | `false`              | Enable debug logging                                                                                     |
 | `TRUSTED_PROXIES`      | loopback + RFC1918   | Comma-separated CIDRs whose `X-Forwarded-For` / `X-Real-IP` headers are trusted for client IP resolution |
 | `REPORTS_PER_MINUTE`   | `60`                 | Reports each client address may send a minute; see [Report limits](#report-limits)                       |
@@ -125,10 +128,11 @@ Reports are written as `reports-<timestamp>-<number>.jsonl.zst` files in
 `DATA_DIR`. The timestamp is in UTC to the millisecond, so the names sort by
 time. The number starts at 1 when the server starts and goes up by one for each
 file the server starts to write, so two files written in the same millisecond
-still get different names. A failed write uses up its number, leaving a gap in
+still get different names. A failed write uses up its number and leaves a gap in
-the numbers if the file could not be created and otherwise a file under that
+the numbers: its file, if it was created, is removed. The file stays, counted
-number that may be incomplete. Each file contains one JSON object per line,
+toward `DATA_DIR_MAX_BYTES` from the next start, only if removing it fails too.
-compressed with zstd. Files are created with `O_EXCL` to prevent overwrites.
+Each file contains one JSON object per line, compressed with zstd. Files are
+created with `O_EXCL` to prevent overwrites.
 
 ### Report limits
 
@@ -148,11 +152,17 @@ credentials, so it is bounded instead. Both refusals below answer with the same
   `X-RateLimit-Reset` headers.
 - **Size cap.** The report files in `DATA_DIR` may total at most
   `DATA_DIR_MAX_BYTES`, counting the files already there at start. Reports
-  waiting in memory count at their uncompressed size until they are written, so
+  waiting in memory count at their uncompressed size until they are written;
-  a report that would take the total past the cap is refused with 507, and
+  those lost to a failed write stop counting, and the part of its file written
-  nothing of it is stored. Deleting report files frees room only at the next
+  is removed. When a report would take the total past the cap, the oldest report
-  start, when the files are counted again. The default of 1 GiB is small enough
+  files are deleted to make room, and each deletion is logged with the file's
-  for any host; set it to the space you can give `DATA_DIR`.
+  name and size; a file still being written is never deleted. A report is
+  refused with 507, and nothing of it is stored, only when the reports waiting
+  to be written fill the cap on their own, and then no file is deleted. At
+  start, report files past the cap, as after lowering it, are deleted the same
+  way. So the cap is how much of the newest reports is kept: the default of 1
+  GiB is small enough for any host; set it to the space you can give
+  `DATA_DIR`.
 
 ### CORS
 
@@ -170,7 +180,6 @@ starting, with an error naming `CORS_ALLOWED_ORIGINS`.
 - Add integration test that POSTs a report and verifies the compressed output
 - Add report decompression/query endpoint
 - Add metrics (Prometheus) for buffer size, flush count, report count
-- Add retention policy to prune old report files
 
 ## License
 

backend/cmd/netwatch-server/main.go

@@ -21,7 +21,6 @@ import (
 var (
 	Appname = "netwatch-server"
 	Version string
-	Buildarch string
 )
 
 func main() {
@@ -40,7 +39,6 @@ func main() {
 
 	globals.Appname = Appname
 	globals.Version = Version
-	globals.Buildarch = Buildarch
 
 	fx.New(
 		fx.Provide(

backend/internal/globals/globals.go

@@ -10,22 +10,18 @@ var (
 	Appname string
 	// Version is the git version tag.
 	Version string
-	// Buildarch is the build architecture.
-	Buildarch string
 )
 
 // Globals holds build-time metadata for the application.
 type Globals struct {
 	Appname string
 	Version string
-	Buildarch string
 }
 
 // New creates a Globals instance from package-level variables.
 func New(_ fx.Lifecycle) (*Globals, error) {
 	return &Globals{
 		Appname: Appname,
-		Buildarch: Buildarch,
 		Version: Version,
 	}, nil
 }

backend/internal/handlers/handlers_test.go

@@ -1,13 +0,0 @@
-package handlers_test
-
-import (
-	"testing"
-
-	_ "sneak.berlin/go/netwatch/internal/handlers"
-)
-
-func TestImport(t *testing.T) {
-	t.Parallel()
-	// Compilation check — verifies the package parses
-	// and all imports resolve.
-}

backend/internal/handlers/healthcheck_test.go

@@ -0,0 +1,120 @@
+package handlers_test
+
+import (
+	"encoding/json"
+	"maps"
+	"net/http"
+	"net/http/httptest"
+	"slices"
+	"testing"
+	"time"
+
+	"sneak.berlin/go/netwatch/internal/globals"
+	"sneak.berlin/go/netwatch/internal/handlers"
+	"sneak.berlin/go/netwatch/internal/healthcheck"
+	"sneak.berlin/go/netwatch/internal/logger"
+
+	"go.uber.org/fx/fxtest"
+)
+
+// newStartedHandlers builds Handlers with a real health check for the
+// server named in g, and starts them, which records the time the
+// uptime counts from.
+func newStartedHandlers(t *testing.T, g *globals.Globals) *handlers.Handlers {
+	t.Helper()
+
+	lc := fxtest.NewLifecycle(t)
+
+	log, err := logger.New(lc, logger.Params{Globals: g})
+	if err != nil {
+		t.Fatalf("logger: %v", err)
+	}
+
+	hc, err := healthcheck.New(lc,
+		healthcheck.Params{Globals: g, Logger: log})
+	if err != nil {
+		t.Fatalf("health check: %v", err)
+	}
+
+	h, err := handlers.New(lc,
+		handlers.Params{Globals: g, Healthcheck: hc, Logger: log})
+	if err != nil {
+		t.Fatalf("handlers: %v", err)
+	}
+
+	lc.RequireStart()
+	t.Cleanup(lc.RequireStop)
+
+	return h
+}
+
+// TestHandleHealthCheck checks the health check's answer: 200, a JSON
+// content type, and a JSON object with exactly the fields of
+// healthcheck.Response, carrying this server's name and version and
+// an uptime counted from its start.
+func TestHandleHealthCheck(t *testing.T) {
+	t.Parallel()
+
+	g := &globals.Globals{Appname: "netwatch-server", Version: "v1.2.3"}
+	h := newStartedHandlers(t, g)
+
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequestWithContext(t.Context(),
+		http.MethodGet, "/.well-known/healthcheck", http.NoBody)
+
+	h.HandleHealthCheck().ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusOK)
+	}
+
+	contentType := rec.Header().Get("Content-Type")
+	if contentType != "application/json; charset=utf-8" {
+		t.Errorf("Content-Type = %q, want %q",
+			contentType, "application/json; charset=utf-8")
+	}
+
+	var body map[string]any
+
+	err := json.Unmarshal(rec.Body.Bytes(), &body)
+	if err != nil {
+		t.Fatalf("body not a JSON object: %v (%q)", err, rec.Body.String())
+	}
+
+	fields := []string{
+		"appname", "now", "status", "uptimeHuman", "uptimeSeconds", "version",
+	}
+	if got := slices.Sorted(maps.Keys(body)); !slices.Equal(got, fields) {
+		t.Fatalf("fields = %v, want %v", got, fields)
+	}
+
+	for field, want := range map[string]string{
+		"appname": g.Appname, "status": "ok", "version": g.Version,
+	} {
+		if body[field] != want {
+			t.Errorf("%s = %v, want %q", field, body[field], want)
+		}
+	}
+
+	now, _ := body["now"].(string)
+
+	at, err := time.Parse(time.RFC3339Nano, now)
+	if err != nil || time.Since(at).Abs() > time.Minute {
+		t.Errorf("now = %q, want the current time in RFC 3339 (%v)", now, err)
+	}
+
+	// Started just now, so the uptime is well under a minute.
+	human, _ := body["uptimeHuman"].(string)
+
+	uptime, err := time.ParseDuration(human)
+	if err != nil || uptime > time.Minute {
+		t.Errorf("uptimeHuman = %q, want a duration under a minute (%v)",
+			human, err)
+	}
+
+	seconds, ok := body["uptimeSeconds"].(float64)
+	if !ok || seconds < 0 || seconds > time.Minute.Seconds() {
+		t.Errorf("uptimeSeconds = %v, want a number of seconds under a minute",
+			body["uptimeSeconds"])
+	}
+}

backend/internal/handlers/report.go

@@ -86,11 +86,12 @@ func (s *Handlers) decodeErrorStatus(err error) int {
 }
 
 // appendErrorStatus logs a failure to store a report and returns
-// the status to send: 507 when the report files are at their size
+// the status to send: 507 when the reports waiting to be written fill
-// cap, otherwise 500.
+// the size cap, otherwise 500.
 func (s *Handlers) appendErrorStatus(err error) int {
 	if errors.Is(err, reportbuf.ErrFull) {
-		s.log.Warn("report refused: report files at their size cap")
+		s.log.Warn("report refused: " +
+			"reports waiting to be written fill the size cap")
 
 		return http.StatusInsufficientStorage
 	}

backend/internal/handlers/report_test.go

@@ -46,6 +46,63 @@ func decodeStatus(t *testing.T, body []byte) string {
 	return resp.Status
 }
 
+// TestHandleReportAcceptsValidReports checks the answer to a valid
+// report, one with no hosts and one shaped as the frontend sends them:
+// 200 and {"status":"ok"} as JSON.
+func TestHandleReportAcceptsValidReports(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name string
+		body string
+	}{
+		{
+			name: "no hosts",
+			body: `{"clientId":"c1","geo":null,"hosts":[],` +
+				`"timestamp":"2026-10-03T12:00:00.000Z"}`,
+		},
+		{
+			name: "a host with a latency and an error sample",
+			body: `{"clientId":"c1","geo":null,"hosts":[{` +
+				`"name":"Example","url":"https://example.com/",` +
+				`"status":"error","history":[` +
+				`{"t":1790000000000,"latency":42,"error":null},` +
+				`{"t":1790000003000,"latency":null,"error":"timeout"}]}],` +
+				`"timestamp":"2026-10-03T12:00:00.000Z"}`,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			h := newTestHandlers(stubAppender{}, io.Discard)
+
+			rec := httptest.NewRecorder()
+			req := httptest.NewRequestWithContext(t.Context(),
+				http.MethodPost, "/api/v1/reports",
+				strings.NewReader(tt.body),
+			)
+
+			h.HandleReport().ServeHTTP(rec, req)
+
+			if rec.Code != http.StatusOK {
+				t.Fatalf("status = %d, want %d", rec.Code, http.StatusOK)
+			}
+
+			contentType := rec.Header().Get("Content-Type")
+			if contentType != "application/json; charset=utf-8" {
+				t.Errorf("Content-Type = %q, want %q",
+					contentType, "application/json; charset=utf-8")
+			}
+
+			if got := rec.Body.String(); got != "{\"status\":\"ok\"}\n" {
+				t.Errorf("body = %q, want %q", got, "{\"status\":\"ok\"}\n")
+			}
+		})
+	}
+}
+
 func TestHandleReportStorageFailureIsNon2xx(t *testing.T) {
 	t.Parallel()
 
@@ -68,9 +125,9 @@ func TestHandleReportStorageFailureIsNon2xx(t *testing.T) {
 	}
 }
 
-// TestHandleReportFullIs507 checks the answer when the report files
+// TestHandleReportFullIs507 checks the answer when the reports waiting
-// are at their size cap: 507 and the usual error body, which tells
+// to be written fill the size cap: 507 and the usual error body, which
-// the client nothing more.
+// tells the client nothing more.
 func TestHandleReportFullIs507(t *testing.T) {
 	t.Parallel()
 

backend/internal/logger/logger.go

@@ -5,6 +5,7 @@ package logger
 import (
 	"log/slog"
 	"os"
+	"runtime"
 
 	"sneak.berlin/go/netwatch/internal/globals"
 
@@ -95,6 +96,6 @@ func (l *Logger) Identify() {
 	l.log.Info("starting",
 		"appname", l.params.Globals.Appname,
 		"version", l.params.Globals.Version,
-		"buildarch", l.params.Globals.Buildarch,
+		"arch", runtime.GOARCH,
 	)
 }

backend/internal/reportbuf/export_test.go

@@ -1,6 +1,13 @@
 package reportbuf
 
-import "time"
+import (
+	"os"
+	"time"
+)
+
+// FlushSizeThreshold exposes the buffer size at which Append starts
+// writing a report file to the external tests.
+const FlushSizeThreshold = flushSizeThreshold
 
 // Flush writes the buffered reports to a file now, as the periodic
 // flush does, so tests need not wait a minute for it.
@@ -13,3 +20,10 @@ func (b *Buffer) Flush() error {
 func (b *Buffer) StopClock(at time.Time) {
 	b.now = func() time.Time { return at }
 }
+
+// OnFileCreated makes the buffer call fn with each report file it
+// writes from now on, once the file is created and before anything is
+// written to it.
+func (b *Buffer) OnFileCreated(fn func(f *os.File)) {
+	b.fileCreated = fn
+}

backend/internal/reportbuf/reportbuf.go

@@ -1,5 +1,6 @@
 // Package reportbuf accumulates telemetry reports in memory
-// and periodically flushes them to zstd-compressed JSONL files.
+// and periodically flushes them to zstd-compressed JSONL files,
+// deleting the oldest files to keep them under a size cap.
 package reportbuf
 
 import (
@@ -37,9 +38,10 @@ const (
 	fileSuffix = ".jsonl.zst"
 )
 
-// ErrFull is returned by Append when storing the report would
+// ErrFull is returned by Append when the reports waiting to be
-// take the report files past the configured maximum size.
+// written leave no room for the report under the configured maximum
-var ErrFull = errors.New("report files at their size cap")
+// size, however many report files are deleted.
+var ErrFull = errors.New("reports waiting to be written fill the size cap")
 
 // Params defines the dependencies for Buffer.
 type Params struct {
@@ -49,12 +51,29 @@ type Params struct {
 	Logger *logger.Logger
 }
 
+// reportFile is a report file that may be deleted to make room, with
+// the size it counts for in usedBytes.
+type reportFile struct {
+	name string
+	size int64
+}
+
 // Buffer accumulates JSON lines in memory and flushes them
 // to zstd-compressed files on disk.
 type Buffer struct {
 	buf     bytes.Buffer
 	dataDir string
 	done    chan struct{}
+	// fileCreated is called with each report file once it is
+	// created, before anything is written to it: it does nothing,
+	// except in tests that hold the write open or make it fail.
+	fileCreated func(f *os.File)
+	// files are the report files that may be deleted to make room,
+	// oldest first: those in dataDir at start, then each one this
+	// buffer writes, once it is complete. A file still being written
+	// is not among them. filesBytes is their total size.
+	files      []reportFile
+	filesBytes int64
 	log        *slog.Logger
 	maxBytes   int64
 	mu         sync.Mutex
@@ -66,8 +85,8 @@ type Buffer struct {
 	seq      atomic.Uint64
 	stopOnce sync.Once
 	// usedBytes is what Append checks against maxBytes: the size
-	// of the report files in dataDir, plus the reports not yet
+	// of the report files in dataDir, plus the reports waiting to
-	// written to one at their uncompressed size.
+	// be written to one at their uncompressed size.
 	usedBytes int64
 }
 
@@ -85,6 +104,7 @@ func New(
 	b := &Buffer{
 		dataDir:     dir,
 		done:        make(chan struct{}),
+		fileCreated: func(*os.File) {},
 		log:         params.Logger.Get(),
 		maxBytes:    params.Config.DataDirMaxBytes,
 		now:         time.Now,
@@ -97,12 +117,28 @@ func New(
 				return fmt.Errorf("create data dir: %w", err)
 			}
 
-			// Report files left by earlier runs count too.
+			// Report files left by earlier runs count too, and are
-			b.usedBytes, err = reportFilesSize(b.dataDir)
+			// the first to be deleted to make room.
+			files, err := reportFiles(b.dataDir)
 			if err != nil {
 				return err
 			}
 
+			b.mu.Lock()
+
+			b.files = files
+			for _, f := range files {
+				b.filesBytes += f.size
+			}
+
+			b.usedBytes = b.filesBytes
+
+			// The files may be past the cap, if it was lowered since
+			// the last run.
+			b.deleteOldestFiles(0)
+
+			b.mu.Unlock()
+
 			go b.flushLoop()
 
 			return nil
@@ -128,9 +164,10 @@ func New(
 }
 
 // Append marshals v as a single JSON line and appends it to
-// the buffer. It stores nothing and returns ErrFull if the line
+// the buffer. If the line would take usedBytes past maxBytes, the
-// would take usedBytes past maxBytes. If the buffer reaches the
+// oldest report files are deleted to make room; it stores nothing
-// size threshold, it is drained and written to disk
+// and returns ErrFull if that cannot make room. If the buffer
+// reaches the size threshold, it is drained and written to disk
 // asynchronously.
 func (b *Buffer) Append(v any) error {
 	line, err := json.Marshal(v)
@@ -142,6 +179,8 @@ func (b *Buffer) Append(v any) error {
 
 	b.mu.Lock()
 
+	b.deleteOldestFiles(lineBytes)
+
 	if b.usedBytes+lineBytes > b.maxBytes {
 		b.mu.Unlock()
 
@@ -171,6 +210,37 @@ func (b *Buffer) Append(v any) error {
 	return nil
 }
 
+// deleteOldestFiles deletes report files, oldest first, until n more
+// bytes fit under maxBytes. It deletes none when the reports waiting
+// to be written leave no room for n even with every file gone, since
+// that would lose the files for nothing. The caller must hold b.mu.
+func (b *Buffer) deleteOldestFiles(n int64) {
+	for b.usedBytes+n > b.maxBytes && len(b.files) > 0 {
+		if b.usedBytes-b.filesBytes+n > b.maxBytes {
+			return
+		}
+
+		f := b.files[0]
+		b.files = b.files[1:]
+		b.filesBytes -= f.size
+
+		// A file already gone, deleted by hand, has freed its room too.
+		err := os.Remove(filepath.Join(b.dataDir, f.name))
+		if err != nil && !errors.Is(err, fs.ErrNotExist) {
+			// The file is still there, so it still counts. It is
+			// not tried again until the next start.
+			b.log.Error("delete report file failed",
+				"file", f.name, "error", err)
+
+			continue
+		}
+
+		b.usedBytes -= f.size
+		b.log.Info("deleted report file to make room",
+			"file", f.name, "bytes", f.size)
+	}
+}
+
 // flushLoop runs a ticker that periodically flushes buffered
 // data to disk until the done channel is closed.
 func (b *Buffer) flushLoop() {
@@ -217,15 +287,41 @@ func (b *Buffer) drainBuf() []byte {
 	return data
 }
 
-// writeFile creates a timestamped zstd-compressed JSONL file
+// writeFile writes data, reports drained from the buffer, to a new
-// in the data directory.
+// timestamped zstd-compressed JSONL file in the data directory.
 func (b *Buffer) writeFile(data []byte) error {
 	// The timestamp comes first, so the names sort by time; the number
 	// after it tells apart files named in the same millisecond.
 	ts := b.now().UTC().Format("2006-01-02T15-04-05.000Z")
 	name := fmt.Sprintf("%s%s-%d%s", filePrefix, ts, b.seq.Add(1), fileSuffix)
-	path := filepath.Join(b.dataDir, name)
 
+	size, err := b.createFile(filepath.Join(b.dataDir, name), data)
+
+	// The reports no longer wait to be written, so they stop counting
+	// at their uncompressed size. If the write failed they are lost;
+	// otherwise they count as the file, which from here on may be
+	// deleted to make room.
+	b.mu.Lock()
+	defer b.mu.Unlock()
+
+	b.usedBytes -= int64(len(data))
+
+	if err != nil {
+		return err
+	}
+
+	b.usedBytes += size
+	b.files = append(b.files, reportFile{name: name, size: size})
+	b.filesBytes += size
+
+	return nil
+}
+
+// createFile creates the file at path holding data compressed with
+// zstd, and returns its size. If the write fails once the file is
+// created, it removes the file, so that a failed write leaves nothing
+// behind to take room.
+func (b *Buffer) createFile(path string, data []byte) (int64, error) {
 	// path is built from the operator-supplied dataDir plus a
 	// generated timestamp and number, so it carries no external input.
 	f, err := os.OpenFile( //nolint:gosec // see comment above
@@ -234,61 +330,65 @@ func (b *Buffer) writeFile(data []byte) error {
 		filePerms,
 	)
 	if err != nil {
-		return fmt.Errorf("create report file: %w", err)
+		return 0, fmt.Errorf("create report file: %w", err)
 	}
 
-	// Closes the file on the early returns below. The success
+	b.fileCreated(f)
-	// path closes it explicitly to check the error; closing it
-	// a second time here is harmless.
-	defer func() { _ = f.Close() }()
 
+	size, err := writeCompressed(f, data)
+	if err != nil {
+		_ = f.Close()
+
+		return 0, errors.Join(err, os.Remove(path))
+	}
+
+	err = f.Close()
+	if err != nil {
+		err = fmt.Errorf("close report file: %w", err)
+
+		return 0, errors.Join(err, os.Remove(path))
+	}
+
+	return size, nil
+}
+
+// writeCompressed writes data to f compressed with zstd, and returns
+// the size of f.
+func writeCompressed(f *os.File, data []byte) (int64, error) {
 	enc, err := zstd.NewWriter(f)
 	if err != nil {
-		return fmt.Errorf("create zstd encoder: %w", err)
+		return 0, fmt.Errorf("create zstd encoder: %w", err)
 	}
 
 	_, err = enc.Write(data)
 	if err != nil {
 		_ = enc.Close()
 
-		return fmt.Errorf("write compressed data: %w", err)
+		return 0, fmt.Errorf("write compressed data: %w", err)
 	}
 
 	err = enc.Close()
 	if err != nil {
-		return fmt.Errorf("close zstd encoder: %w", err)
+		return 0, fmt.Errorf("close zstd encoder: %w", err)
 	}
 
 	info, err := f.Stat()
 	if err != nil {
-		return fmt.Errorf("stat report file: %w", err)
+		return 0, fmt.Errorf("stat report file: %w", err)
 	}
 
-	err = f.Close()
+	return info.Size(), nil
-	if err != nil {
-		return fmt.Errorf("close report file: %w", err)
-	}
-
-	// The reports counted at their uncompressed size while they
-	// waited; now they count as the file. After a failed write they
-	// stay counted as they were, which errs toward refusing reports
-	// early rather than letting the files pass the cap.
-	b.mu.Lock()
-	b.usedBytes += info.Size() - int64(len(data))
-	b.mu.Unlock()
-
-	return nil
 }
 
-// reportFilesSize returns the total size of the report files in
+// reportFiles returns the report files in dir, oldest first:
-// dir.
+// os.ReadDir sorts them by name, and the names sort by time.
-func reportFilesSize(dir string) (int64, error) {
+func reportFiles(dir string) ([]reportFile, error) {
 	entries, err := os.ReadDir(dir)
 	if err != nil {
-		return 0, fmt.Errorf("read data dir: %w", err)
+		return nil, fmt.Errorf("read data dir: %w", err)
 	}
 
-	var total int64
+	files := make([]reportFile, 0, len(entries))
 
 	for _, entry := range entries {
 		name := entry.Name()
@@ -299,11 +399,11 @@ func reportFilesSize(dir string) (int64, error) {
 
 		info, err := entry.Info()
 		if err != nil {
-			return 0, fmt.Errorf("stat report file: %w", err)
+			return nil, fmt.Errorf("stat report file: %w", err)
 		}
 
-		total += info.Size()
+		files = append(files, reportFile{name: name, size: info.Size()})
 	}
 
-	return total, nil
+	return files, nil
 }

backend/internal/reportbuf/reportbuf_test.go

@@ -139,11 +139,19 @@ func lineBytes(t *testing.T, report any) int {
 	return len(line) + 1
 }
 
+// TestAppendPastCapIsRefused fills the cap with a report not yet
+// written. The next report is refused, and the report file already in
+// DATA_DIR is kept: it is smaller than a report, so deleting it could
+// not make room.
 func TestAppendPastCapIsRefused(t *testing.T) {
 	report := map[string]string{"id": "cap"}
+	dir := t.TempDir()
+	earlier := reportFilePath(dir, 1)
 
-	t.Setenv("DATA_DIR", t.TempDir())
+	writeBytes(t, earlier, 1)
-	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(lineBytes(t, report)))
+
+	t.Setenv("DATA_DIR", dir)
+	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(1+lineBytes(t, report)))
 
 	buf := startBuffer(t)
 
@@ -156,24 +164,38 @@ func TestAppendPastCapIsRefused(t *testing.T) {
 	if !errors.Is(err, reportbuf.ErrFull) {
 		t.Fatalf("report past the cap: error = %v, want ErrFull", err)
 	}
+
+	if !exists(t, earlier) {
+		t.Fatal("report file deleted, though that could not make room")
+	}
 }
 
-// TestCapCountsReportFilesAlreadyInDataDir starts on a data
+// TestOldestReportFileDeletedFirst starts on a data directory holding
-// directory holding a report file from an earlier run, and a file
+// report files from an earlier run, and a file that is not a report,
-// that is not a report, which must not count.
+// which neither counts nor is ever deleted. Nothing is deleted while
-func TestCapCountsReportFilesAlreadyInDataDir(t *testing.T) {
+// there is room; then only the oldest report file is.
-	const earlierBytes = 100
+func TestOldestReportFileDeletedFirst(t *testing.T) {
+	const fileBytes = 100
 
-	report := map[string]string{"id": "cap"}
+	report := map[string]string{"id": "oldest"}
 	dir := t.TempDir()
+	oldest := reportFilePath(dir, 1)
+	kept := []string{
+		reportFilePath(dir, 2),
+		reportFilePath(dir, 3),
+		filepath.Join(dir, "notes.txt"),
+	}
 
-	writeBytes(t, filepath.Join(dir, "reports-2026-01-01T00-00-00.000Z.jsonl.zst"),
+	writeBytes(t, oldest, fileBytes)
-		earlierBytes)
+
-	writeBytes(t, filepath.Join(dir, "notes.txt"), 10*earlierBytes)
+	for _, path := range kept {
+		writeBytes(t, path, fileBytes)
+	}
 
 	t.Setenv("DATA_DIR", dir)
+	// Room for the three report files and one report.
 	t.Setenv("DATA_DIR_MAX_BYTES",
-		strconv.Itoa(earlierBytes+lineBytes(t, report)))
+		strconv.Itoa(3*fileBytes+lineBytes(t, report)))
 
 	buf := startBuffer(t)
 
@@ -182,9 +204,55 @@ func TestCapCountsReportFilesAlreadyInDataDir(t *testing.T) {
 		t.Fatalf("report that fills the cap exactly: %v", err)
 	}
 
+	if !exists(t, oldest) {
+		t.Fatal("oldest report file deleted while there was room")
+	}
+
 	err = buf.Append(report)
-	if !errors.Is(err, reportbuf.ErrFull) {
+	if err != nil {
-		t.Fatalf("report past the cap: error = %v, want ErrFull", err)
+		t.Fatalf("report past the cap: %v", err)
+	}
+
+	if exists(t, oldest) {
+		t.Fatal("oldest report file kept when room was needed")
+	}
+
+	for _, path := range kept {
+		if !exists(t, path) {
+			t.Fatalf("%s deleted; only the oldest report file should be", path)
+		}
+	}
+}
+
+// TestStartDeletesFilesPastCap starts on report files past the cap,
+// as after the cap is lowered: the oldest are deleted until the rest
+// fit.
+func TestStartDeletesFilesPastCap(t *testing.T) {
+	const fileBytes = 100
+
+	dir := t.TempDir()
+	oldest := reportFilePath(dir, 1)
+	kept := []string{reportFilePath(dir, 2), reportFilePath(dir, 3)}
+
+	writeBytes(t, oldest, fileBytes)
+
+	for _, path := range kept {
+		writeBytes(t, path, fileBytes)
+	}
+
+	t.Setenv("DATA_DIR", dir)
+	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(len(kept)*fileBytes))
+
+	startBuffer(t)
+
+	if exists(t, oldest) {
+		t.Fatal("oldest report file kept, though the files were past the cap")
+	}
+
+	for _, path := range kept {
+		if !exists(t, path) {
+			t.Fatalf("%s deleted, though the rest fit without it", path)
+		}
 	}
 }
 
@@ -195,8 +263,9 @@ func TestWrittenReportsCountAtFileSize(t *testing.T) {
 	// Repetitive, so its file is far smaller than its JSON.
 	report := map[string]string{"id": strings.Repeat("a", 1000)}
 	size := lineBytes(t, report)
+	dir := t.TempDir()
 
-	t.Setenv("DATA_DIR", t.TempDir())
+	t.Setenv("DATA_DIR", dir)
 	// Room for the report twice over only if the first one counts
 	// at its file's size by the time the second arrives.
 	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(2*size-1))
@@ -217,13 +286,22 @@ func TestWrittenReportsCountAtFileSize(t *testing.T) {
 	if err != nil {
 		t.Fatalf("second report, after the first was written: %v", err)
 	}
+
+	if !hasReportFile(t, dir) {
+		t.Fatal("first report's file deleted, though the second fit beside it")
+	}
 }
 
-// TestWrittenReportsKeepCounting writes one report file after another
+// TestWrittenFilesDeletedToMakeRoom writes one report file after
-// under a small cap: each report must be taken while the files on disk
+// another under a small cap. Every report must be taken; files are
-// leave room for it, and refused once they do not.
+// deleted only when the report would not fit beside them, and the
-func TestWrittenReportsKeepCounting(t *testing.T) {
+// files kept leave room for it.
-	const maxBytes = 200
+func TestWrittenFilesDeletedToMakeRoom(t *testing.T) {
+	const (
+		maxBytes = 200
+		// One file each, which take far more than maxBytes together.
+		reports = 50
+	)
 
 	report := map[string]string{"id": "written"}
 	size := int64(lineBytes(t, report))
@@ -234,23 +312,24 @@ func TestWrittenReportsKeepCounting(t *testing.T) {
 
 	buf := startBuffer(t)
 
-	// Every file takes at least a byte, so they fill the cap within
+	for range reports {
-	// maxBytes rounds.
+		before := reportFilesBytes(t, dir)
-	for range maxBytes {
-		used := reportFilesBytes(t, dir)
 
 		err := buf.Append(report)
-		if used+size > maxBytes {
-			if !errors.Is(err, reportbuf.ErrFull) {
-				t.Fatalf("with %d bytes of report files: error = %v, "+
-					"want ErrFull", used, err)
-			}
-
-			return
-		}
-
 		if err != nil {
-			t.Fatalf("with %d bytes of report files: %v", used, err)
+			t.Fatalf("with %d bytes of report files: %v", before, err)
+		}
+
+		after := reportFilesBytes(t, dir)
+
+		if before+size <= maxBytes && after != before {
+			t.Fatalf("files deleted, though the report fit beside "+
+				"their %d bytes", before)
+		}
+
+		if after+size > maxBytes {
+			t.Fatalf("%d bytes of report files kept, leaving no room "+
+				"for the report", after)
 		}
 
 		err = buf.Flush()
@@ -258,8 +337,217 @@ func TestWrittenReportsKeepCounting(t *testing.T) {
 			t.Fatalf("flush: %v", err)
 		}
 	}
+}
 
-	t.Fatal("the report files never filled the cap")
+// TestFailedDeletionStillCounts makes deleting the oldest report file
+// fail. It is still there, so it still takes room, and the next oldest
+// is deleted in its place.
+func TestFailedDeletionStillCounts(t *testing.T) {
+	const fileBytes = 100
+
+	report := map[string]string{"id": "stuck"}
+	dir := t.TempDir()
+	stuck := reportFilePath(dir, 1)
+	next := reportFilePath(dir, 2)
+	newest := reportFilePath(dir, 3)
+
+	for _, path := range []string{stuck, next, newest} {
+		writeBytes(t, path, fileBytes)
+	}
+
+	t.Setenv("DATA_DIR", dir)
+	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(3*fileBytes))
+
+	buf := startBuffer(t)
+
+	// A directory that is not empty cannot be deleted, even by root,
+	// which the tests run as in the backend image.
+	err := os.Remove(stuck)
+	if err != nil {
+		t.Fatalf("remove %s: %v", stuck, err)
+	}
+
+	err = os.Mkdir(stuck, 0o750)
+	if err != nil {
+		t.Fatalf("make directory %s: %v", stuck, err)
+	}
+
+	writeBytes(t, filepath.Join(stuck, "file"), 1)
+
+	err = buf.Append(report)
+	if err != nil {
+		t.Fatalf("report past the cap: %v", err)
+	}
+
+	if exists(t, next) {
+		t.Fatal("next oldest report file kept: the failed deletion " +
+			"counted as making room")
+	}
+
+	if !exists(t, newest) {
+		t.Fatal("newest report file deleted, though deleting one made room")
+	}
+}
+
+// TestFileDeletedByHandFreesRoom deletes the oldest report file by
+// hand after start. When room is needed, its room counts as freed, so
+// no other file is deleted.
+func TestFileDeletedByHandFreesRoom(t *testing.T) {
+	const fileBytes = 100
+
+	report := map[string]string{"id": "by-hand"}
+	dir := t.TempDir()
+	gone := reportFilePath(dir, 1)
+	kept := reportFilePath(dir, 2)
+
+	writeBytes(t, gone, fileBytes)
+	writeBytes(t, kept, fileBytes)
+
+	t.Setenv("DATA_DIR", dir)
+	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(2*fileBytes))
+
+	buf := startBuffer(t)
+
+	err := os.Remove(gone)
+	if err != nil {
+		t.Fatalf("remove %s: %v", gone, err)
+	}
+
+	err = buf.Append(report)
+	if err != nil {
+		t.Fatalf("report past the cap: %v", err)
+	}
+
+	if !exists(t, kept) {
+		t.Fatal("report file deleted, though the one deleted by hand " +
+			"had made room")
+	}
+}
+
+// TestFileBeingWrittenIsNeverDeleted holds the write of one report file
+// open while a second write completes, then sends a report that needs
+// room. Deleting either file would make it, and the one being written is
+// the older, but only the complete one may be deleted. Once the first
+// write is complete, its file is deleted when room is needed.
+func TestFileBeingWrittenIsNeverDeleted(t *testing.T) {
+	report := map[string]string{"id": "writing"}
+
+	t.Setenv("DATA_DIR", t.TempDir())
+	// Room for two reports waiting to be written, but not for two
+	// beside a report file.
+	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(2*lineBytes(t, report)))
+
+	buf := startBuffer(t)
+
+	created := make(chan string)
+	release := make(chan struct{})
+
+	buf.OnFileCreated(func(f *os.File) {
+		created <- f.Name()
+
+		<-release
+	})
+
+	err := buf.Append(report)
+	if err != nil {
+		t.Fatalf("first report: %v", err)
+	}
+
+	flushed := make(chan error)
+
+	go func() { flushed <- buf.Flush() }()
+
+	writing := <-created
+
+	// Only the first write is held; the second goes through, and
+	// so do the writes after it, the final one at stop included.
+	var complete string
+
+	buf.OnFileCreated(func(f *os.File) { complete = f.Name() })
+
+	// Errorf, not Fatalf, until the first write is released, so that a
+	// failure here does not leave it held.
+	err = buf.Append(report)
+	if err != nil {
+		t.Errorf("second report: %v", err)
+	}
+
+	err = buf.Flush()
+	if err != nil {
+		t.Errorf("flush of the second report: %v", err)
+	}
+
+	err = buf.Append(report)
+	if err != nil {
+		t.Errorf("report that needs room: %v", err)
+	}
+
+	if !exists(t, writing) {
+		t.Error("report file deleted while it was being written")
+	}
+
+	if exists(t, complete) {
+		t.Error("complete report file kept, though room was needed")
+	}
+
+	close(release)
+
+	err = <-flushed
+	if err != nil {
+		t.Fatalf("flush of the first report: %v", err)
+	}
+
+	err = buf.Append(report)
+	if err != nil {
+		t.Fatalf("report after the first write was complete: %v", err)
+	}
+
+	if exists(t, writing) {
+		t.Fatal("complete report file kept when room was needed")
+	}
+}
+
+// TestFailedWriteStopsCounting makes a write fail once its file is
+// created. Its reports are lost, so they stop counting, and the part of
+// the file written is removed, so it takes no room.
+func TestFailedWriteStopsCounting(t *testing.T) {
+	report := map[string]string{"id": "failed"}
+
+	t.Setenv("DATA_DIR", t.TempDir())
+	t.Setenv("DATA_DIR_MAX_BYTES", strconv.Itoa(lineBytes(t, report)))
+
+	buf := startBuffer(t)
+
+	var failed string
+
+	// Closing the file under the write makes the write fail.
+	buf.OnFileCreated(func(f *os.File) {
+		failed = f.Name()
+
+		_ = f.Close()
+	})
+
+	err := buf.Append(report)
+	if err != nil {
+		t.Fatalf("report that fills the cap exactly: %v", err)
+	}
+
+	err = buf.Flush()
+	if err == nil {
+		t.Fatal("flush succeeded, though its file was closed under it")
+	}
+
+	if exists(t, failed) {
+		t.Fatal("file of the failed write kept")
+	}
+
+	// Writes from here on, the final one at stop included, succeed.
+	buf.OnFileCreated(func(*os.File) {})
+
+	err = buf.Append(report)
+	if err != nil {
+		t.Fatalf("report after the failed write: %v", err)
+	}
 }
 
 // TestConcurrentAppendsStopAtCap appends from many goroutines at once
@@ -334,7 +622,11 @@ func TestTwoFlushesInOneMillisecond(t *testing.T) {
 		}
 	}
 
-	files := readReportFiles(t, dir)
+	files, err := readReportFiles(dir)
+	if err != nil {
+		t.Fatalf("read report files: %v", err)
+	}
+
 	if len(files) != flushes {
 		t.Fatalf("%d report files after %d flushes", len(files), flushes)
 	}
@@ -347,6 +639,88 @@ func TestTwoFlushesInOneMillisecond(t *testing.T) {
 	}
 }
 
+// TestReportFileHoldsTheLinesAppended flushes three reports and reads
+// their file back: it must decompress to exactly their JSON lines, in
+// the order they were appended.
+func TestReportFileHoldsTheLinesAppended(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("DATA_DIR", dir)
+
+	buf := startBuffer(t)
+
+	for _, id := range []int{1, 2, 3} {
+		err := buf.Append(map[string]int{"id": id})
+		if err != nil {
+			t.Fatalf("append report %d: %v", id, err)
+		}
+	}
+
+	err := buf.Flush()
+	if err != nil {
+		t.Fatalf("flush: %v", err)
+	}
+
+	files, err := readReportFiles(dir)
+	if err != nil {
+		t.Fatalf("read report files: %v", err)
+	}
+
+	want := `{"id":1}` + "\n" + `{"id":2}` + "\n" + `{"id":3}` + "\n"
+	if len(files) != 1 || files[0] != want {
+		t.Fatalf("report files = %q, want one holding %q", files, want)
+	}
+}
+
+// TestFlushAtSizeThreshold appends reports until the buffer holds
+// FlushSizeThreshold bytes. The append that gets it there must write
+// them all to one report file, with no call to Flush and the periodic
+// flush a minute away, and no earlier append may write one.
+func TestFlushAtSizeThreshold(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("DATA_DIR", dir)
+
+	buf := startBuffer(t)
+
+	// Large reports, so the threshold takes a few hundred appends.
+	pad := strings.Repeat("a", 64<<10)
+
+	var appended strings.Builder
+
+	for id := 0; appended.Len() < reportbuf.FlushSizeThreshold; id++ {
+		report := map[string]any{"id": id, "pad": pad}
+
+		err := buf.Append(report)
+		if err != nil {
+			t.Fatalf("append report %d: %v", id, err)
+		}
+
+		line, err := json.Marshal(report)
+		if err != nil {
+			t.Fatalf("marshal report %d: %v", id, err)
+		}
+
+		appended.Write(line)
+		appended.WriteByte('\n')
+	}
+
+	// Append writes the file in the background, so wait for it.
+	deadline := time.Now().Add(10 * time.Second)
+
+	for {
+		files, err := readReportFiles(dir)
+		if err == nil && len(files) == 1 && files[0] == appended.String() {
+			return
+		}
+
+		if time.Now().After(deadline) {
+			t.Fatalf("%d report files (error: %v), want one holding the "+
+				"%d bytes appended", len(files), err, appended.Len())
+		}
+
+		time.Sleep(10 * time.Millisecond)
+	}
+}
+
 // reportFilesBytes returns the total size of the report files in dir.
 func reportFilesBytes(t *testing.T, dir string) int64 {
 	t.Helper()
@@ -371,20 +745,19 @@ func reportFilesBytes(t *testing.T, dir string) int64 {
 }
 
 // readReportFiles returns the decompressed contents of each report
-// file in dir.
+// file in dir. A file still being written does not decompress, so it
-func readReportFiles(t *testing.T, dir string) []string {
+// gives an error.
-	t.Helper()
+func readReportFiles(dir string) ([]string, error) {
-
 	files := os.DirFS(dir)
 
 	names, err := fs.Glob(files, "reports-*.jsonl.zst")
 	if err != nil {
-		t.Fatalf("list report files: %v", err)
+		return nil, fmt.Errorf("list report files: %w", err)
 	}
 
 	dec, err := zstd.NewReader(nil)
 	if err != nil {
-		t.Fatalf("create zstd decoder: %v", err)
+		return nil, fmt.Errorf("create zstd decoder: %w", err)
 	}
 	defer dec.Close()
 
@@ -393,18 +766,26 @@ func readReportFiles(t *testing.T, dir string) []string {
 	for _, name := range names {
 		compressed, readErr := fs.ReadFile(files, name)
 		if readErr != nil {
-			t.Fatalf("read %s: %v", name, readErr)
+			return nil, fmt.Errorf("read %s: %w", name, readErr)
 		}
 
 		data, decErr := dec.DecodeAll(compressed, nil)
 		if decErr != nil {
-			t.Fatalf("decompress %s: %v", name, decErr)
+			return nil, fmt.Errorf("decompress %s: %w", name, decErr)
 		}
 
 		contents = append(contents, string(data))
 	}
 
-	return contents
+	return contents, nil
+}
+
+// reportFilePath returns the path in dir of a report file named as
+// written on the given day of January 2026, so that a lower day sorts
+// as older.
+func reportFilePath(dir string, day int) string {
+	return filepath.Join(dir,
+		fmt.Sprintf("reports-2026-01-%02dT00-00-00.000Z-1.jsonl.zst", day))
 }
 
 func writeBytes(t *testing.T, path string, n int) {
@@ -416,6 +797,21 @@ func writeBytes(t *testing.T, path string, n int) {
 	}
 }
 
+func exists(t *testing.T, path string) bool {
+	t.Helper()
+
+	_, err := os.Stat(path)
+	if errors.Is(err, fs.ErrNotExist) {
+		return false
+	}
+
+	if err != nil {
+		t.Fatalf("stat %s: %v", path, err)
+	}
+
+	return true
+}
+
 func hasReportFile(t *testing.T, dir string) bool {
 	t.Helper()
 

backend/internal/server/http.go

@@ -4,6 +4,7 @@ import (
 	"errors"
 	"net"
 	"net/http"
+	"runtime"
 	"strconv"
 	"time"
 
@@ -53,7 +54,7 @@ func (s *Server) listenAndServe() {
 	s.log.Info("http begin listen",
 		"listenaddr", s.httpServer.Addr,
 		"version", s.params.Globals.Version,
-		"buildarch", s.params.Globals.Buildarch,
+		"arch", runtime.GOARCH,
 	)
 
 	err := s.httpServer.ListenAndServe()

backend/script/build

@@ -1,6 +1,6 @@
 #!/bin/sh
 # script/build: compile the static netwatch-server binary into the
-# backend project root, with its version and architecture stamped in.
+# backend project root, with its version stamped in.
 set -eu
 
 ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"
@@ -14,7 +14,7 @@ main() {
     version="${VERSION:-$(git describe --always --dirty 2>/dev/null || echo dev)}"
 
     CGO_ENABLED=0 go build -trimpath \
-        -ldflags "-s -w -X main.Version=$version -X main.Buildarch=$(uname -m)" \
+        -ldflags "-s -w -X main.Version=$version" \
         -o netwatch-server ./cmd/netwatch-server/
 }
 

backend/script/test

@@ -1,12 +1,20 @@
 #!/bin/sh
-# script/test: run the backend test suite.
+# script/test: run the backend test suite with the race detector and
+# coverage. Go's own -timeout bounds the tests and not their compile,
+# so a cold build cache cannot fail it. The race detector needs cgo,
+# and so a C compiler. If the tests fail, they run again with -v for
+# the details, and the script fails even if that run passes.
 set -eu
 
 ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"
 
 main() {
     cd "$ROOT"
-    timeout 30 go test ./...
+    go test -timeout 30s -race -cover ./... || {
+        echo "--- Rerunning with -v for details ---"
+        go test -timeout 30s -race -v ./...
+        exit 1
+    }
 }
 
 main "$@"

index.html

@@ -9,6 +9,9 @@
             type="image/svg+xml"
             href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><text y='.9em' font-size='90'>📡</text></svg>"
         />
+        <!-- Linked here, not imported by src/main.js, so the unit tests can
+             import that module in Node, which cannot import CSS. -->
+        <link rel="stylesheet" href="/src/styles.css" />
     </head>
     <body class="bg-gray-900 text-white min-h-screen">
         <div id="app"></div>

script/bootstrap

@@ -64,7 +64,8 @@ detect_pkgmgr() {
     fi
 }
 
-# pkg_install <nix-attr> <apt-pkg> <brew-formula> <apk-pkg>
+# pkg_install <nix-attr> <apt-pkgs> <brew-formula> <apk-pkgs>: the apt
+# and apk arguments may each list several packages, separated by spaces.
 pkg_install() {
     detect_pkgmgr
     case "$PKGMGR" in
@@ -74,10 +75,10 @@ pkg_install() {
                 $SUDO env DEBIAN_FRONTEND=noninteractive apt-get update
                 APT_UPDATED=1
             fi
-            $SUDO env DEBIAN_FRONTEND=noninteractive apt-get install -y "$2"
+            $SUDO env DEBIAN_FRONTEND=noninteractive apt-get install -y $2
             ;;
         brew) brew install "$3" ;;
-        apk) apk add --no-cache "$4" ;;
+        apk) apk add --no-cache $4 ;;
     esac
 }
 
@@ -253,6 +254,12 @@ main() {
 
     if missing make; then pkg_install gnumake make make make; fi
     if missing git; then pkg_install git git git git; fi
+    # The race detector in make test needs cgo, which Go turns on only
+    # when it finds its C compiler, gcc on Linux. apt and apk ship the C
+    # library headers apart from gcc.
+    if missing gcc; then
+        pkg_install gcc "gcc libc6-dev" gcc "gcc musl-dev"
+    fi
 
     ensure_node
     ensure_yarn

script/cibuild

@@ -16,9 +16,8 @@ main() {
     "$SCRIPT_DIR/check"
     # Own line: a failing command substitution inside an argument does
     # not trip `set -e`, so the inline form degrades silently to an
-    # empty constant. VERSION is computed here because .dockerignore
+    # empty constant. The VERSION build argument takes precedence over
-    # excludes .git, so `git describe` in a build stage yields an empty
+    # the version a build stage derives from the .git in the context.
-    # version without failing.
     version="$(git describe --tags --always --dirty 2>/dev/null || true)"
     [ -n "$version" ] || version="unknown"
     docker build --no-cache \

script/docker

@@ -12,9 +12,8 @@ main() {
     cd "$ROOT"
     # Own line: a failing command substitution inside an argument does
     # not trip `set -e`, so the inline form degrades silently to an
-    # empty constant. VERSION is computed here because .dockerignore
+    # empty constant. The VERSION build argument takes precedence over
-    # excludes .git, so `git describe` in a build stage yields an empty
+    # the version a build stage derives from the .git in the context.
-    # version without failing.
     version="$(git describe --tags --always --dirty 2>/dev/null || true)"
     [ -n "$version" ] || version="unknown"
     docker build --no-cache \

script/frontend-test

@@ -1,13 +1,14 @@
 #!/bin/sh
-# script/frontend-test: run the frontend test suite. The frontend has no
+# script/frontend-test: run the frontend test suite: the unit tests in
-# unit tests; the production build serves as the test (fails on broken
+# test/unit/ with Node's built-in test runner, then the production
-# code).
+# build, which fails on broken code.
 set -eu
 
 ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"
 
 main() {
     cd "$ROOT"
+    timeout 30 node --test test/unit/*.test.js
     timeout 30 yarn build
 }
 

script/test

@@ -1,14 +1,17 @@
 #!/bin/sh
 # script/test: run the test suite for the whole repo: the frontend at
-# the repo root, then the Go backend in backend/. Both halves together
+# the repo root, then the Go backend in backend/. Each half has its own
-# get 30 seconds; each also keeps its own limit for the Dockerfiles.
+# 30-second limit, and there is none around both: from a cold Go build
+# cache, compiling the backend's tests with the race detector can take
+# 30 seconds on its own, and Go's -timeout leaves the compile out.
 set -eu
 
 ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"
 
 main() {
     cd "$ROOT"
-    timeout 30 sh -c 'script/frontend-test && backend/script/test'
+    script/frontend-test
+    backend/script/test
 }
 
 main "$@"

src/main.js

@@ -1,14 +1,14 @@
-import "./styles.css";
-
 // --- Configuration -----------------------------------------------------------
 
-// Timing, axis labels, and display constants. Latency above maxLatency is
+// Timing, axis labels, and display constants. A target check times out
-// clamped to "unreachable". The sparkline Y-axis is capped at
+// after requestTimeout, 80% of updateInterval, so a round's checks have
+// all finished before the next round is due; latency above maxLatency is
+// recorded as a timeout. The sparkline Y-axis is capped at
 // graphMaxLatency — values above it pin to the top of the chart but still
 // display their real value in the latency figure. The history buffer holds
 // maxHistoryPoints samples (historyDuration / updateInterval).
 // reportInterval is how often collected samples are POSTed to the backend.
-const CONFIG = {
+export const CONFIG = {
     updateInterval: 3000,
     maxHistoryPoints: 100,
     reportInterval: 60000,
@@ -16,7 +16,7 @@ const CONFIG = {
         return (this.maxHistoryPoints * this.updateInterval) / 1000;
     },
     get requestTimeout() {
-        return Math.min(this.updateInterval - 100, 3000);
+        return this.updateInterval * 0.8;
     },
     get maxLatency() {
         return this.requestTimeout;
@@ -503,12 +503,16 @@ class Reporter {
 
 // --- Latency Measurement -----------------------------------------------------
 
-async function measureLatency(url) {
+// Checks one target. The check times out after CONFIG.requestTimeout; the
+// caller can give it up sooner through the optional signal, which also ends
+// it as a timeout.
+export async function measureLatency(url, signal) {
     const controller = new AbortController();
     const timeoutId = setTimeout(
         () => controller.abort(),
         CONFIG.requestTimeout,
     );
+    signal?.addEventListener("abort", () => controller.abort());
 
     const targetUrl = new URL(url);
     targetUrl.searchParams.set("_cb", Date.now().toString());
@@ -1101,7 +1105,7 @@ function sortAndRebuildWAN(state) {
 
 // --- Main Loop ---------------------------------------------------------------
 
-async function tick(state, onOffline) {
+async function tick(state, signal, onOffline) {
     const ts = Date.now();
 
     if (state.paused) {
@@ -1123,11 +1127,12 @@ async function tick(state, onOffline) {
     log.debug(`Tick #${state.tickCount + 1} started`);
 
     const results = await Promise.all(
-        state.allHosts.map((h) => measureLatency(h.url)),
+        state.allHosts.map((h) => measureLatency(h.url, signal)),
     );
 
-    // User may have paused while awaiting results — discard them
+    // User may have paused, or the next round may have given up this
-    if (state.paused) return;
+    // one's checks, while awaiting results — discard them
+    if (state.paused || signal.aborted) return;
 
     state.tickCount++;
 
@@ -1168,9 +1173,10 @@ async function tick(state, onOffline) {
 
 // --- Recovery Probe ----------------------------------------------------------
 
-// When offline, rapidly poll 4 random WAN hosts every 500ms. As soon as any
+// When offline, check 4 random WAN hosts every 500ms, giving up the checks
-// responds, stop probing and fire a normal tick to refresh all hosts.
+// started 500ms before, so at most 4 are ever waiting. As soon as one
-function startRecoveryProbe(state, triggerTick) {
+// answers, stop probing and start a new round at once.
+function startRecoveryProbe(state, startRounds) {
     if (state._recoveryProbeId) return; // already running
     const candidates = [...state.wan];
     for (let i = candidates.length - 1; i > 0; i--) {
@@ -1181,15 +1187,18 @@ function startRecoveryProbe(state, triggerTick) {
     log.notice(
         `Recovery probe started (${canaries.map((h) => h.name).join(", ")})`,
     );
-    state._recoveryProbeId = setInterval(async () => {
+    state._recoveryProbeId = setInterval(() => {
         if (state.paused) return;
-        const results = await Promise.all(
+        state._recoveryProbeChecks?.abort();
-            canaries.map((h) => measureLatency(h.url)),
+        const checks = new AbortController();
-        );
+        state._recoveryProbeChecks = checks;
-        if (results.some((r) => r.error === null)) {
+        for (const host of canaries) {
+            measureLatency(host.url, checks.signal).then((r) => {
+                if (r.error !== null || checks.signal.aborted) return;
                 log.notice("Recovery probe: connectivity detected");
                 stopRecoveryProbe(state);
-            triggerTick();
+                startRounds();
+            });
         }
     }, 500);
 }
@@ -1198,6 +1207,7 @@ function stopRecoveryProbe(state) {
     if (state._recoveryProbeId) {
         clearInterval(state._recoveryProbeId);
         state._recoveryProbeId = null;
+        state._recoveryProbeChecks?.abort();
     }
 }
 
@@ -1374,18 +1384,34 @@ async function init() {
     updateClocks();
     setInterval(updateClocks, 1000);
 
+    // Rounds never overlap: a round first gives up the last round's checks
+    // if they are still waiting, and the last round then records nothing.
+    // At a steady interval they never are, as they time out at 80% of it;
+    // they can be when a round starts early, after an interval change or
+    // when the recovery probe finds a target answering.
+    let roundChecks = new AbortController();
     function doTick() {
-        tick(state, () => startRecoveryProbe(state, doTick));
+        roundChecks.abort();
+        roundChecks = new AbortController();
+        tick(state, roundChecks.signal, () =>
+            startRecoveryProbe(state, startRounds),
+        );
     }
 
+    // Starts a round now and then one every CONFIG.updateInterval.
+    let tickIntervalId;
+    function startRounds() {
+        clearInterval(tickIntervalId);
         doTick();
-    let tickIntervalId = setInterval(doTick, CONFIG.updateInterval);
+        tickIntervalId = setInterval(doTick, CONFIG.updateInterval);
+    }
+
+    startRounds();
 
     document
         .getElementById("interval-select")
         .addEventListener("change", (e) => {
             const newInterval = parseInt(e.target.value, 10);
-            clearInterval(tickIntervalId);
             CONFIG.updateInterval = newInterval;
             log.notice(
                 `Interval changed to ${humanDuration(newInterval / 1000)}, history reset`,
@@ -1434,8 +1460,7 @@ async function init() {
 
             // Start immediately with new interval
             stopRecoveryProbe(state);
-            doTick();
+            startRounds();
-            tickIntervalId = setInterval(doTick, CONFIG.updateInterval);
         });
 
     window.addEventListener("resize", () => handleResize(state));
@@ -1444,7 +1469,7 @@ async function init() {
 
 // Bootstrap only when loaded as the page: a real DOM containing the #app
 // mount point this module renders into. Importing the module in a unit test
-// (which has no #app) runs nothing, so buildReport can be tested in isolation.
+// (which has no #app) runs nothing, so its exports can be tested in isolation.
 if (typeof document !== "undefined" && document.getElementById("app")) {
     if (document.readyState === "loading") {
         document.addEventListener("DOMContentLoaded", init);

test/unit/main.test.js

@@ -0,0 +1,66 @@
+// Unit tests for src/main.js, run by script/frontend-test with Node's
+// built-in test runner. Importing the module does not start the page.
+
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { CONFIG, measureLatency } from "../../src/main.js";
+
+// measureLatency writes timeouts to the debug log, which looks for its
+// panel in the page. There is no page here.
+globalThis.document = { getElementById: () => null };
+
+// Mocks the clock for test t, so that a check lasting seconds takes no real
+// time, and replaces fetch with a target that answers after answerAfter
+// milliseconds of that clock, or never when answerAfter is Infinity. Both
+// are restored when the test ends.
+function mockTarget(t, answerAfter) {
+    t.mock.timers.enable({ apis: ["setTimeout", "Date"] });
+    t.mock.method(performance, "now", () => Date.now());
+    t.mock.method(
+        globalThis,
+        "fetch",
+        (url, { signal }) =>
+            new Promise((resolve, reject) => {
+                if (answerAfter !== Infinity) setTimeout(resolve, answerAfter);
+                signal.addEventListener("abort", () => reject(signal.reason));
+            }),
+    );
+}
+
+// The result of check if it has ended, otherwise "still waiting".
+function settled(check) {
+    return Promise.race([
+        check,
+        new Promise((resolve) => setImmediate(resolve, "still waiting")),
+    ]);
+}
+
+for (const interval of [10000, 30000]) {
+    const timeout = interval * 0.8;
+    // Over 3 seconds, which the timeout used to be capped at.
+    const slowAnswer = timeout - 1000;
+
+    test(`at a ${interval}ms interval, an answer after ${slowAnswer}ms is recorded with its real time`, async (t) => {
+        CONFIG.updateInterval = interval;
+        mockTarget(t, slowAnswer);
+        const check = measureLatency("https://target.test");
+        t.mock.timers.tick(slowAnswer);
+        assert.deepEqual(await settled(check), {
+            latency: slowAnswer,
+            error: null,
+        });
+    });
+
+    test(`at a ${interval}ms interval, a target that never answers is recorded as a timeout after ${timeout}ms`, async (t) => {
+        CONFIG.updateInterval = interval;
+        mockTarget(t, Infinity);
+        const check = measureLatency("https://target.test");
+        t.mock.timers.tick(timeout - 1);
+        assert.equal(await settled(check), "still waiting");
+        t.mock.timers.tick(1);
+        assert.deepEqual(await settled(check), {
+            latency: null,
+            error: "timeout",
+        });
+    });
+}