fix(test): store tests take an OS-assigned free port, not a PID-derived one

PID%1000 collided with a concurrent job on the shared CI host (run on
e946709: process already listening on port 54000). freePort asks the OS;
Start is retried up to 3 times if the port is taken in between.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SZkhi49hfeSqBp5aZJNXsM

.gitea/workflows/ci.yml

@@ -53,10 +53,12 @@ jobs:
 
       - name: Derive image tags
         id: meta
+        env:
+          # signals#47: a ref name can carry $(...); pass it via env, not ${{ }}.
+          REF: ${{ github.ref }}
         run: |
           SHA=$(git rev-parse --short HEAD)
           echo "sha-tag=${SHA}" >> "$GITHUB_OUTPUT"
-          REF="${{ github.ref }}"
           if [[ "$REF" == refs/tags/v* ]]; then
             echo "version-tag=${REF#refs/tags/}" >> "$GITHUB_OUTPUT"
           fi

CLAUDE.md

@@ -27,6 +27,23 @@ docs it indexes.
 - **Language is Go (ADR-001).** Do not introduce Python or a second language. If you think you
   need one, that's a new ADR with a real justification, not a default.
 
+## Deploying: a push does not deploy
+
+Tapir runs on the koala k3s cluster under Flux GitOps. Pushing here only builds the image;
+`kubectl set image` is reverted by Flux drift correction within seconds. The manifest bump is
+manual: Flux image-automation in `mathias/infra` covers only `supervisor`, not tapir (infra#111
+is closed, but `k3s/flux/flux-system/image-automation.yaml` has no tapir policy).
+
+1. CI (`.gitea/workflows/ci.yml`) builds and pushes `localhost:5000/tapir:<short-sha>` on the
+   main push.
+2. In `mathias/infra`, bump `image:` in `k3s/apps/tapir/deployment.yaml` to that sha; commit to
+   `main` (infra's own rules and full `task check` apply).
+3. On koala: `flux reconcile kustomization apps --with-source`, then
+   `kubectl -n tapir rollout status deploy/tapir`. Both need per-action approval in a session.
+4. Confirm the running image is the new sha and the startup log shows the new code.
+
+Runbook and why: brain `homelab/failures/flux-reverts-kubectl-set-image-deploy-via-git`.
+
 ## Things that look reusable but are NOT — read before "reusing"
 
 These caused real mistakes that were caught and corrected; the corrections are load-bearing.

DECISIONS.md

@@ -1466,10 +1466,19 @@ an hour of Swedish *was* an unknown here; it is now measured, see decision 8.)
    as busy beyond that. Together with the 20-job queue and the running job that bounds the spool
    at about (4 + 20 + 1) x 200 MiB = 5 GiB of the 6 Gi `emptyDir` (infra): raising any of those or
    `TAPIR_UPLOAD_MAX_BYTES` must revisit that sum (pinned by `TestNewIntake_Defaults...`, not
-   enforced against infra). **A slow-drip body holds its slot without a time bound**: the server
-   sets `ReadHeaderTimeout` only, because a `ReadTimeout` would cut legitimate 200 MiB uploads on
-   slow links; two slow bodies take a user's two slots and two such users make everyone's upload
-   busy. It needs a valid session; recorded as a limitation with a follow-up issue.
+   enforced against infra). **A body is bounded by its silence, not its length (#56, amended
+   2026-10-04)**: a `ReadTimeout` would cut legitimate 200 MiB uploads on slow links, so instead
+   every read of the upload body re-arms the connection's read deadline 60 s ahead
+   (`App.UploadIdleTimeout`). A body that stops delivering bytes fails its read, Accept returns
+   and the slot is released; one that keeps moving is never cut. The handler never clears the
+   deadline: an armed deadline also bounds net/http's drain of an unread remainder before it
+   replies, and net/http resets it itself before its background read. A deadline still armed
+   while the file is probed (the multipart part can end before the body's last bytes arrive)
+   costs at most keep-alive reuse of that connection, not the upload. The 60 s default is the
+   deployed value (cmd/tapir does not set it) and is pinned by a test. It needs `http.ResponseController` to reach the connection, so the
+   metrics middleware's writer implements `Unwrap`; where a deadline cannot be set the handler
+   logs `upload idle bound unavailable` rather than silently running unbounded. Still not
+   bounded: a body that drips one byte per 59 s holds its slot for as long as it keeps that up.
 
 8. **Chat reuses `chat.Service`; it is not forked.** The uploaded-transcript chat handlers
    look the row up through the RLS-scoped store, so a non-owner gets 404 by the same mechanism

internal/adapters/store/port_test.go

@@ -0,0 +1,19 @@
+package store_test
+
+import (
+	"fmt"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+// TestMain's Postgres already holds a port. A second instance started in the
+// same process must still come up, i.e. the port is not derived from something
+// two starters share (the PID). Observes: startTestPostgres succeeds while
+// another instance from this process is listening.
+func TestStartTestPostgres_SecondInstanceStartsWhileFirstHoldsItsPort(t *testing.T) {
+	pg, _, rt, err := startTestPostgres(fmt.Sprintf("tapir-epg-store-second-%d", os.Getpid()))
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = pg.Stop(); _ = os.RemoveAll(rt) })
+}

internal/adapters/store/store_test.go

@@ -3,6 +3,7 @@ package store_test
 import (
 	"context"
 	"fmt"
+	"net"
 	"os"
 	"path/filepath"
 	"testing"
@@ -25,27 +26,15 @@ var _ ports.Sink = (*store.Store)(nil)
 var dsn string
 
 func TestMain(m *testing.M) {
-	// Port + runtime/data dirs are per-process (PID-derived) so two concurrent
-	// `go test` invocations — e.g. a push-run and a tag-run firing together in CI —
-	// don't collide on a fixed port or a shared data dir (which silently failed
-	// both runs). CachePath is shared so the PG archive is downloaded once, not
-	// per process. Base 54000 keeps this package's range distinct from web's.
-	port := uint32(54000 + os.Getpid()%1000)
-	dsn = fmt.Sprintf("postgres://postgres:postgres@localhost:%d/postgres?sslmode=disable", port)
-
-	rt := filepath.Join(os.TempDir(), fmt.Sprintf("tapir-epg-store-%d", os.Getpid()))
-	pg := embeddedpostgres.NewDatabase(
-		embeddedpostgres.DefaultConfig().
-			Port(port).
-			RuntimePath(rt).
-			DataPath(filepath.Join(rt, "data")).
-			BinariesPath(filepath.Join(rt, "bin")).
-			CachePath(filepath.Join(os.TempDir(), "tapir-epg-cache")),
-	)
-	if err := pg.Start(); err != nil {
+	// Runtime/data dirs are per-process (PID-named) and the port is OS-assigned,
+	// so concurrent `go test` runs on a shared CI host don't collide on either.
+	// CachePath is shared so the PG archive is downloaded once, not per process.
+	pg, port, rt, err := startTestPostgres(fmt.Sprintf("tapir-epg-store-%d", os.Getpid()))
+	if err != nil {
 		fmt.Fprintf(os.Stderr, "embedded-postgres start: %v\n", err)
 		os.Exit(1)
 	}
+	dsn = fmt.Sprintf("postgres://postgres:postgres@localhost:%d/postgres?sslmode=disable", port)
 
 	code := m.Run()
 
@@ -56,6 +45,44 @@ func TestMain(m *testing.M) {
 	os.Exit(code)
 }
 
+// freePort asks the OS for an unused TCP port. The port is released before
+// return, so another process can grab it first; startTestPostgres retries.
+func freePort() (uint32, error) {
+	l, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		return 0, fmt.Errorf("listen for free port: %w", err)
+	}
+	defer func() { _ = l.Close() }()
+	return uint32(l.Addr().(*net.TCPAddr).Port), nil
+}
+
+// startTestPostgres starts an embedded Postgres on an OS-assigned free port and
+// returns it with that port and its runtime dir (caller removes it). A port
+// taken between freePort and Start (shared CI host) is retried with a new one.
+func startTestPostgres(name string) (*embeddedpostgres.EmbeddedPostgres, uint32, string, error) {
+	rt := filepath.Join(os.TempDir(), name)
+	var lastErr error
+	for range 3 {
+		port, err := freePort()
+		if err != nil {
+			return nil, 0, rt, err
+		}
+		pg := embeddedpostgres.NewDatabase(
+			embeddedpostgres.DefaultConfig().
+				Port(port).
+				RuntimePath(rt).
+				DataPath(filepath.Join(rt, "data")).
+				BinariesPath(filepath.Join(rt, "bin")).
+				CachePath(filepath.Join(os.TempDir(), "tapir-epg-cache")),
+		)
+		if lastErr = pg.Start(); lastErr == nil {
+			return pg, port, rt, nil
+		}
+		lastErr = fmt.Errorf("start on port %d: %w", port, lastErr)
+	}
+	return nil, 0, rt, lastErr
+}
+
 // uuids — fixed so tests are deterministic. user_id/video_id are UUID columns.
 const (
 	userA  = "11111111-1111-1111-1111-111111111111"

internal/metrics/metrics.go

@@ -125,6 +125,11 @@ type statusWriter struct {
 	wroteHeader bool
 }
 
+// Unwrap exposes the underlying writer to http.ResponseController, so a handler behind this
+// middleware can still set a connection read deadline (#56: without it the upload idle bound
+// was silently unavailable in production).
+func (s *statusWriter) Unwrap() http.ResponseWriter { return s.ResponseWriter }
+
 func (s *statusWriter) WriteHeader(code int) {
 	if !s.wroteHeader {
 		s.code = code

internal/web/export_test.go

@@ -0,0 +1,11 @@
+package web
+
+import "time"
+
+// DefaultUploadIdleTimeout exposes the production idle bound (#56) to the black-box tests:
+// cmd/tapir never sets App.UploadIdleTimeout, so this constant IS the deployed value.
+const DefaultUploadIdleTimeout = defaultUploadIdleTimeout
+
+// UploadIdleTimeoutForTest exposes the bound an App resolves to (#56), so a test can pin the
+// value the zero-value App -- production's -- actually uses rather than only the constant.
+func (a *App) UploadIdleTimeoutForTest() time.Duration { return a.uploadIdleTimeout() }

internal/web/handlers.go

@@ -116,6 +116,11 @@ type App struct {
 	// front rather than discovered by failing.
 	Uploads      Uploader
 	UploadLimits uploads.Limits
+	// UploadIdleTimeout is how long an upload body may go without delivering a byte
+	// before the connection read is cut off and its receive slot freed (#56). Zero
+	// means defaultUploadIdleTimeout. It bounds silence, not the whole body, so a
+	// large upload on a slow but moving link still completes.
+	UploadIdleTimeout time.Duration
 	// UploadChatMaxChars is the transcript budget for chat over an uploaded
 	// recording. The caption default (18,000 chars, ~22 minutes of speech) would
 	// keep only the first third of an hour-long meeting.

internal/web/uploads.go

@@ -84,7 +84,7 @@ func (a *App) handleUploadPost(w http.ResponseWriter, r *http.Request) {
 	if !ok {
 		return
 	}
-	r.Body = http.MaxBytesReader(w, r.Body, a.UploadLimits.MaxBytes+uploadOverhead)
+	r.Body = http.MaxBytesReader(w, a.idleBoundedBody(w, r.Body), a.UploadLimits.MaxBytes+uploadOverhead)
 	mr, err := r.MultipartReader()
 	if err != nil {
 		a.renderUploadsPage(w, r, userID, http.StatusBadRequest, "Choose an audio file to upload.")
@@ -158,6 +158,53 @@ func (a *App) handleUploadPost(w http.ResponseWriter, r *http.Request) {
 	a.renderUploadsPage(w, r, userID, http.StatusBadRequest, "Choose an audio file to upload.")
 }
 
+// defaultUploadIdleTimeout is how long an upload body may deliver nothing before its
+// connection read is cut off (#56). A stall this long is not a slow link but a dead or
+// hostile one; a legitimate upload, however slow, keeps delivering bytes.
+const defaultUploadIdleTimeout = 60 * time.Second
+
+// uploadIdleTimeout is the idle bound this App applies: UploadIdleTimeout when set, else
+// defaultUploadIdleTimeout. cmd/tapir leaves it unset, so production runs the default.
+func (a *App) uploadIdleTimeout() time.Duration {
+	if a.UploadIdleTimeout <= 0 {
+		return defaultUploadIdleTimeout
+	}
+	return a.UploadIdleTimeout
+}
+
+// idleBoundedBody bounds the SILENCE in an upload body, not its total duration (#56): every
+// read first pushes the connection's read deadline idle into the future, so a body that
+// keeps moving is never cut, and one that stops fails its read with a timeout. That read
+// error reaches Accept as a client read error, so Accept returns and its receive slot is
+// released. A whole-request ReadTimeout would instead cut a large upload on a slow link.
+//
+// If the writer cannot set a read deadline (a wrapper without Unwrap, or a test recorder)
+// the body is returned unbounded and that is logged: a control that silently does not
+// apply is worse than one that says so.
+func (a *App) idleBoundedBody(w http.ResponseWriter, body io.ReadCloser) io.ReadCloser {
+	idle := a.uploadIdleTimeout()
+	rc := http.NewResponseController(w)
+	if err := rc.SetReadDeadline(time.Now().Add(idle)); err != nil {
+		a.logger().Warn("upload idle bound unavailable", "err", err)
+		return body
+	}
+	return &idleDeadlineBody{ReadCloser: body, rc: rc, idle: idle}
+}
+
+type idleDeadlineBody struct {
+	io.ReadCloser
+	rc   *http.ResponseController
+	idle time.Duration
+}
+
+func (b *idleDeadlineBody) Read(p []byte) (int, error) {
+	_ = b.rc.SetReadDeadline(time.Now().Add(b.idle))
+	// Never cleared here: net/http clears the read deadline itself before its background
+	// read once the handler has the body, and an armed deadline after an error also bounds
+	// net/http's drain of an unread remainder before it replies.
+	return b.ReadCloser.Read(p)
+}
+
 // abandoned reports, logs at INFO and answers 499, because the client has already gone: its request
 // context is cancelled, so there is nobody to answer and rendering a page would run its
 // database read on a dead context and log a false ERROR. A hang-up mid-upload is routine.

internal/web/uploads_idle_handler_test.go

@@ -0,0 +1,161 @@
+package web_test
+
+// #56: a request body that stops arriving must not hold a receive slot forever. The bound is a
+// read deadline on the connection, re-armed on every read, so it only exercises over a real
+// socket and through the production handler stack (metrics.HTTPMiddleware wraps the writer).
+
+import (
+	"bufio"
+	"bytes"
+	"errors"
+	"fmt"
+	"log/slog"
+	"mime/multipart"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/metrics"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+	"git.d-ma.be/mathias/tapir/internal/web"
+)
+
+func multipartBody(t *testing.T, content []byte) (body []byte, contentType string) {
+	t.Helper()
+	var buf bytes.Buffer
+	mw := multipart.NewWriter(&buf)
+	fw, err := mw.CreateFormFile("file", "a.m4a")
+	require.NoError(t, err)
+	_, err = fw.Write(content)
+	require.NoError(t, err)
+	require.NoError(t, mw.Close())
+	return buf.Bytes(), mw.FormDataContentType()
+}
+
+// rawUpload opens a connection and sends the request headers; the caller feeds the body.
+func rawUpload(t *testing.T, addr, contentType string, length int) net.Conn {
+	t.Helper()
+	c, err := net.Dial("tcp", addr)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = c.Close() })
+	_, err = fmt.Fprintf(c, "POST /uploads HTTP/1.1\r\nHost: x\r\nContent-Type: %s\r\nContent-Length: %d\r\n\r\n", contentType, length)
+	require.NoError(t, err)
+	return c
+}
+
+// serverEnded reports whether the server ended the request (answered it or closed the
+// connection) before wait elapsed, rather than leaving the client waiting.
+func serverEnded(t *testing.T, c net.Conn, wait time.Duration) bool {
+	t.Helper()
+	require.NoError(t, c.SetReadDeadline(time.Now().Add(wait)))
+	_, err := http.ReadResponse(bufio.NewReader(c), nil)
+	var ne net.Error
+	return !errors.As(err, &ne) || !ne.Timeout()
+}
+
+// readStatus returns the response status code, or 0 if none arrived before wait elapsed.
+func readStatus(t *testing.T, c net.Conn, wait time.Duration) int {
+	t.Helper()
+	require.NoError(t, c.SetReadDeadline(time.Now().Add(wait)))
+	resp, err := http.ReadResponse(bufio.NewReader(c), nil)
+	if err != nil {
+		return 0
+	}
+	_ = resp.Body.Close()
+	return resp.StatusCode
+}
+
+func newIdleBoundServer(t *testing.T, idle time.Duration) (addr string) {
+	t.Helper()
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+	app.UploadIdleTimeout = idle
+	app.Uploads.(*uploads.Intake).MaxReceivesPerUser = 1
+	srv := httptest.NewServer(metrics.HTTPMiddleware(app.Router()))
+	t.Cleanup(srv.Close)
+	return strings.TrimPrefix(srv.URL, "http://")
+}
+
+func TestUploadPost_AConnectionReadDeadlineEndsABodyThatStallsAndReleasesItsReceiveSlot(t *testing.T) {
+	addr := newIdleBoundServer(t, 200*time.Millisecond)
+	body, ct := multipartBody(t, bytes.Repeat([]byte("a"), 8000))
+
+	stalled := rawUpload(t, addr, ct, len(body))
+	_, err := stalled.Write(body[:4000]) // half the body, then nothing more
+	require.NoError(t, err)
+
+	require.True(t, serverEnded(t, stalled, 3*time.Second),
+		"the server never ended the stalled request: it still holds its slot")
+
+	next := rawUpload(t, addr, ct, len(body))
+	_, err = next.Write(body)
+	require.NoError(t, err)
+	require.Equal(t, http.StatusSeeOther, readStatus(t, next, 3*time.Second),
+		"the same user's next upload gets the slot the stalled one held")
+}
+
+func TestUploadPost_TheReadDeadlineIsReArmedOnEveryReadSoASlowButMovingBodyIsAccepted(t *testing.T) {
+	addr := newIdleBoundServer(t, 300*time.Millisecond)
+	body, ct := multipartBody(t, bytes.Repeat([]byte("a"), 2000))
+
+	c := rawUpload(t, addr, ct, len(body))
+	// ~1.2 s in total, four times the idle bound, but never 300 ms without a byte.
+	const chunks = 20
+	step := len(body)/chunks + 1
+	for i := 0; i < len(body); i += step {
+		_, err := c.Write(body[i:min(i+step, len(body))])
+		require.NoError(t, err)
+		time.Sleep(60 * time.Millisecond)
+	}
+
+	require.Equal(t, http.StatusSeeOther, readStatus(t, c, 3*time.Second))
+}
+
+// cmd/tapir never sets UploadIdleTimeout, so the default is the value production runs.
+// Shorter and legitimate uploads on a slow link are cut; much longer and the bound is moot.
+func TestDefaultUploadIdleTimeoutIsSixtySeconds(t *testing.T) {
+	require.Equal(t, 60*time.Second, web.DefaultUploadIdleTimeout)
+}
+
+// The bound an App actually resolves to, not the constant: production's App leaves
+// UploadIdleTimeout unset, so the zero value must resolve to exactly 60 s, and a set value
+// must be used as given.
+func TestAppUploadIdleTimeout_ResolvesUnsetToSixtySecondsAndHonoursASetValue(t *testing.T) {
+	require.Equal(t, 60*time.Second, (&web.App{}).UploadIdleTimeoutForTest())
+	require.Equal(t, 60*time.Second, (&web.App{UploadIdleTimeout: -time.Second}).UploadIdleTimeoutForTest())
+	require.Equal(t, 300*time.Millisecond, (&web.App{UploadIdleTimeout: 300 * time.Millisecond}).UploadIdleTimeoutForTest())
+}
+
+// With UploadIdleTimeout unset, a body with 100 ms gaps between reads is not cut (a resolved
+// bound in milliseconds or nanoseconds would cut it). The exact value is pinned above.
+func TestUploadPost_AnUnsetIdleTimeoutDoesNotCutABodyWithSubSecondGaps(t *testing.T) {
+	addr := newIdleBoundServer(t, 0)
+	body, ct := multipartBody(t, bytes.Repeat([]byte("a"), 1000))
+
+	c := rawUpload(t, addr, ct, len(body))
+	const chunks = 5
+	step := len(body)/chunks + 1
+	for i := 0; i < len(body); i += step {
+		time.Sleep(100 * time.Millisecond)
+		_, err := c.Write(body[i:min(i+step, len(body))])
+		require.NoError(t, err)
+	}
+
+	require.Equal(t, http.StatusSeeOther, readStatus(t, c, 3*time.Second))
+}
+
+// A writer that cannot set a read deadline leaves the body unbounded; that must be said in
+// the log, not happen silently. The in-process recorder is such a writer.
+func TestUploadPost_AWriterWithoutReadDeadlineSupportLogsThatTheIdleBoundIsUnavailable(t *testing.T) {
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+	buf := &lockedBuf{}
+	app.Log = slog.New(slog.NewTextHandler(buf, nil))
+
+	do(t, app, uploadReq(t, "t", "a.m4a", []byte("x")))
+
+	require.Contains(t, buf.String(), "upload idle bound unavailable")
+}