fix(transcribe): rebuild the transcript from segments so split words are whole

whisper.cpp wraps at ~60 chars and can wrap inside a word; the response's text
field joins segments with newlines, so chat saw 'nyt' / 't godkännande'. Text
is now rebuilt from the raw segments: concatenated when any segment carries a
leading space (whisper.cpp), joined by a space when none do (berget's sentence
level), whitespace collapsed; the text field is only a fallback.

Checked against the first real one-hour transcript with pianissimo's vocabulary
as an independent yardstick: 36 spans changed, 31 now consist entirely of words
pianissimo also produced, words unknown to pianissimo fell 224 -> 177 and word
agreement rose 92.4% -> 93.1%. Five rejoined spans still contain an unknown
word (names/loanwords or an unlucky glue) -- a residual risk, small.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

internal/adapters/transcribe/gateway.go

@@ -110,8 +110,9 @@ func (g *Gateway) Transcribe(ctx context.Context, path string) (domain.Transcrip
 }
 
 // verboseJSON is the subset of the verbose_json response used. Both backends
-// return it; they differ in segment granularity (iguana: whole seconds, berget:
+// return it; they differ in segment granularity (iguana: whole seconds and
-// sub-second) and in the language token, neither of which matters here.
+// sub-sentence pieces, berget: sub-second sentences), in the language token, and in
+// whether a segment's text starts with a space.
 type verboseJSON struct {
 	Text     string  `json:"text"`
 	Duration float64 `json:"duration"`
@@ -124,13 +125,21 @@ type verboseJSON struct {
 
 // parseVerbose normalises a verbose_json body. Whisper's final window can end past
 // the audio (9.0 s for 4.25 s of audio was observed), so segment ends are clamped
-// to the reported duration.
+// to the reported duration. Text is rebuilt from the segments, not taken from the
+// response's `text` field: see joinSegments.
 func parseVerbose(r io.Reader) (domain.Transcription, error) {
 	var v verboseJSON
 	if err := json.NewDecoder(r).Decode(&v); err != nil {
 		return domain.Transcription{}, fmt.Errorf("decode response: %w", err)
 	}
-	text := strings.TrimSpace(v.Text)
+	raw := make([]string, 0, len(v.Segments))
+	for _, s := range v.Segments {
+		raw = append(raw, s.Text)
+	}
+	text := joinSegments(raw)
+	if text == "" {
+		text = collapseSpace(v.Text)
+	}
 	if text == "" {
 		return domain.Transcription{}, ErrEmptyTranscript
 	}
@@ -145,3 +154,32 @@ func parseVerbose(r io.Reader) (domain.Transcription, error) {
 	}
 	return t, nil
 }
+
+// joinSegments rebuilds the running text from raw (untrimmed) segment texts.
+//
+// whisper.cpp wraps its output at ~60 characters and a wrap can land INSIDE a word
+// ("nyt" | "t godkännande"): a segment that begins with a space starts a new word,
+// one that does not continues the previous one, and the response's own `text`
+// field joins them with newlines, which breaks the word. So when any segment
+// carries a leading space the raw texts are concatenated as they are. Backends
+// that return whole sentences with no leading space (berget) are joined with a
+// space instead — concatenating those would glue sentences together.
+func joinSegments(raw []string) string {
+	leadingSpace := false
+	for _, s := range raw {
+		if strings.HasPrefix(s, " ") {
+			leadingSpace = true
+			break
+		}
+	}
+	sep := " "
+	if leadingSpace {
+		sep = ""
+	}
+	return collapseSpace(strings.Join(raw, sep))
+}
+
+// collapseSpace turns every run of whitespace (including newlines) into one space.
+func collapseSpace(s string) string {
+	return strings.Join(strings.Fields(s), " ")
+}

internal/adapters/transcribe/gateway_test.go

@@ -121,7 +121,7 @@ func TestGateway_NormalisesTheVerboseJSONResponse(t *testing.T) {
 	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
 
 	require.NoError(t, err)
-	assert.Equal(t, "God morgon allihopa.\n Vi går igenom läget.", got.Text)
+	assert.Equal(t, "God morgon allihopa. Vi går igenom läget.", got.Text)
 	assert.Equal(t, "iguana/kb-whisper", got.Endpoint, "the serving endpoint is part of the result")
 	assert.InDelta(t, 4.25, got.AudioSeconds, 1e-9)
 	assert.Equal(t, []domain.Segment{
@@ -172,3 +172,61 @@ func TestGateway_MissingAudioFileIsAnError(t *testing.T) {
 
 	assert.Error(t, err)
 }
+
+// whisper.cpp wraps its output at ~60 characters, and a wrap can land INSIDE a word:
+// the continuation segment has no leading space, the `text` field joins segments
+// with newlines, and a segment can be a lone ".". Measured on real gateway output
+// 2026-10-01: 7% of a one-hour recording's segments were such fragments. Joining by
+// newline hands the chat model broken words ("nyt\nt"), so Text is rebuilt from the
+// segments: a segment that begins with a space starts a new word, one that does not
+// continues the previous one.
+func TestGateway_RejoinsWordsWhisperCppSplitAcrossSegments(t *testing.T) {
+	reply := `{
+	  "text": " Vi ligger just nu 12 procent över den ursprungliga kalkylen\n.\n Det kräver ett nyt\nt godkännande från beställaren.\n Framförallt på arbetspl\natsen.\n",
+	  "language": "swedish", "duration": 20,
+	  "segments": [
+	    {"start": 0, "end": 4, "text": " Vi ligger just nu 12 procent över den ursprungliga kalkylen"},
+	    {"start": 4, "end": 5, "text": "."},
+	    {"start": 5, "end": 8, "text": " Det kräver ett nyt"},
+	    {"start": 8, "end": 11, "text": "t godkännande från beställaren."},
+	    {"start": 11, "end": 14, "text": " Framförallt på arbetspl"},
+	    {"start": 14, "end": 16, "text": "atsen."}
+	  ]}`
+	srv := serve(t, 200, reply, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "iguana/kb-whisper", srv.Client())
+
+	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	require.NoError(t, err)
+	assert.Equal(t, "Vi ligger just nu 12 procent över den ursprungliga kalkylen. "+
+		"Det kräver ett nytt godkännande från beställaren. Framförallt på arbetsplatsen.", got.Text)
+	assert.Len(t, got.Segments, 6, "the timed segments are kept as the backend gave them (trimmed)")
+	assert.Equal(t, ".", got.Segments[1].Text)
+}
+
+func TestGateway_SentenceLevelSegmentsWithNoLeadingSpaceAreJoinedWithASpace(t *testing.T) {
+	// berget/kb-whisper returns whole sentences with no leading space; rebuilding by
+	// concatenation would glue them ("allihopa!Vi").
+	reply := `{"text": "God morgon allihopa!\nVi går igenom läget.", "language": "sv", "duration": 5,
+	  "segments": [
+	    {"start": 0.05, "end": 1.2, "text": "God morgon allihopa!"},
+	    {"start": 1.6, "end": 4.2, "text": "Vi går igenom läget."}
+	  ]}`
+	srv := serve(t, 200, reply, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "berget/kb-whisper", srv.Client())
+
+	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	require.NoError(t, err)
+	assert.Equal(t, "God morgon allihopa! Vi går igenom läget.", got.Text)
+}
+
+func TestGateway_WithNoSegmentsTheTextFieldIsUsedWithWhitespaceCollapsed(t *testing.T) {
+	srv := serve(t, 200, `{"text": "  Hej  allihopa.\n Vi börjar.  ", "duration": 3, "segments": []}`, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "m", srv.Client())
+
+	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	require.NoError(t, err)
+	assert.Equal(t, "Hej allihopa. Vi börjar.", got.Text)
+}

internal/web/uploads.go

@@ -111,12 +111,15 @@ func (a *App) uploadFailure(w http.ResponseWriter, r *http.Request, userID strin
 	var tooBig *http.MaxBytesError
 	switch {
 	case errors.As(err, &rej):
+		// Logged by reason only: the filename and title are private content.
+		a.logger().Info("upload rejected", "reason", string(rej.Reason))
 		status := http.StatusUnprocessableEntity
 		if rej.Reason == uploads.ReasonTooLarge {
 			status = http.StatusRequestEntityTooLarge
 		}
 		a.renderUploadsPage(w, r, userID, status, rej.Message)
 	case errors.As(err, &tooBig):
+		a.logger().Info("upload rejected", "reason", string(uploads.ReasonTooLarge))
 		a.renderUploadsPage(w, r, userID, http.StatusRequestEntityTooLarge,
 			"This file is larger than the upload limit.")
 	default:

internal/web/uploads_handler_test.go

@@ -3,6 +3,7 @@ package web_test
 import (
 	"bytes"
 	"context"
+	"log/slog"
 	"mime/multipart"
 	"net/http"
 	"net/http/httptest"
@@ -336,3 +337,21 @@ func TestUploadPage_AlsoCarriesTheVideoNotice(t *testing.T) {
 
 	assert.Contains(t, body(t, do(t, app, get("/uploads/"+id))), "Only its audio was used")
 }
+
+func TestUploadPost_ARejectionIsLoggedWithItsReasonAndNeverTheFilenameOrTitle(t *testing.T) {
+	// Rejections are synchronous and leave no row, so without a log line an operator
+	// cannot tell "the user never retried" from "the page refused it". The reason is
+	// operational; the filename and title are the user's private content.
+	var buf bytes.Buffer
+	app, _ := newUploadApp(t, uploads.Probe{Duration: time.Minute, HasVideo: true}, 1<<20)
+	app.Log = slog.New(slog.NewTextHandler(&buf, nil))
+
+	rec := do(t, app, uploadReq(t, "Secret project notes", "client-acme-roadmap.mp4", []byte("x")))
+
+	require.Equal(t, http.StatusUnprocessableEntity, rec.Code)
+	logged := buf.String()
+	assert.Contains(t, logged, "upload rejected")
+	assert.Contains(t, logged, "reason=no_audio")
+	assert.NotContains(t, logged, "Secret project notes", "the title is private content")
+	assert.NotContains(t, logged, "client-acme-roadmap", "the filename is private content")
+}