feat(discovery): per-channel caption-availability memory (ADR-024)

After transcript caching (ADR-021) and the Shorts filter (ADR-023), the remaining
caption waste is the first fetch on every new video of a channel that never has
English captions — each costs one rate-limited fetch to resolve to "none", and on
a throttled IP churns the backoff machinery first.

Remember, per (user, channel), a streak of consecutive no-caption outcomes
(channel_caption_state, migration 016, RLS-scoped). Once it reaches
TAPIR_CHANNEL_CAPTIONLESS_THRESHOLD (default 5) the channel is suppressed — videos
discovered/listed but not caption-fetched — for TAPIR_CHANNEL_CAPTIONLESS_WINDOW
(default 14d), then one is re-probed (auto-recovery). A successful fetch resets
the streak; a 429 does not count; an explicit manual request bypasses suppression.
threshold=0 disables.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -86,7 +86,7 @@ Skills live in the canonical library `mathias/skills` and are wired into this re
 
 ## Current build state (start here for the first task)
 
-The repo is **green and shipping** — last tag `v0.14.0`. `task check` passes (fmt, vet, lint,
+The repo is **green and shipping** — last tag `v0.15.0`. `task check` passes (fmt, vet, lint,
 `go test -p 1 ./...`). Go is `1.26.1` (see `go.mod`).
 
 - Clean Architecture core is implemented: `internal/domain` (entities), `internal/ports`
@@ -95,7 +95,8 @@ The repo is **green and shipping** — last tag `v0.14.0`. `task check` passes (
   green against it.
 - Adapters present under `internal/adapters/`: `youtube` (captions-first `VideoSource`,
   timedtext/InnerTube acquisition per ADR-010), `summarizer` + `llm` (the copied AI router,
-  Primary→Fallback per ADR-004), `store` (Postgres, golang-migrate migrations 001–015),
+  now a resilient endpoint chain — local primary → local fallback → external worst-case,
+  parse-failure-aware, ADR-004 + ADR-022), `store` (Postgres, golang-migrate migrations 001–015),
   `secrets` (file-backed `SecretStore`). The brain HTTP sink (ADR-005) is the remaining
   optional sink.
 - Stage 1 is open (ADR-012): multi-user with **DB-enforced** isolation — Postgres RLS `FORCE`d

DECISIONS.md

@@ -803,6 +803,126 @@ never exposed in the UI.
 
 ---
 
+## ADR-022 — Summarizer is a resilient endpoint chain, not a single model
+
+**Status:** Accepted (2026-06-10). **Extends ADR-004** (the copied `llm` Primary→Fallback
+routing). Triggered by the first friendly-pilot live run, where a connected user got **zero**
+summaries after 12h.
+
+**Context.** Stage-0 ran a single summarizer model (`koala/phi4-mini`) with no fallback wired
+(`summarizer.New(primary, nil)`). The live run exposed three independent failure modes, each of
+which silently produced no summary:
+
+1. **Context overflow.** `phi4-mini` has an 8k context. Real transcripts (one was 11,602 tokens)
+   exceed it and the gateway returns HTTP 400 — and the request also sent `max_tokens=8192`, so
+   even a short transcript plus the completion budget could overflow the window.
+2. **Malformed model output.** `phi4-mini` intermittently emits `highlights` as a bare string
+   instead of an array, producing `cannot unmarshal string into []string`. The old code returned
+   the parse error **without** trying any other model — a 200-with-bad-JSON short-circuited.
+3. **No fallback existed at all** — any primary failure was terminal for that video.
+
+`phi4-mini` is kept as primary deliberately: it is fast and, on transcripts that fit, correct.
+The fix is resilience around it, not replacing it.
+
+**Decision.**
+1. **Ordered endpoint chain (`summarizer.NewChain`).** Endpoints are tried in order; the first to
+   return a *parseable* summary wins. Default chain:
+   `koala/phi4-mini` (primary, local) → `koala/phi4-14b` (fallback, local) →
+   `berget/mistral-small` (worst-case, external). All three are reached through the **one** LiteLLM
+   gateway by alias — the gateway already fronts both llama-swap and berget — so a fallback is a
+   different alias, not a second client config.
+2. **A parse failure advances the chain, same as a transport error.** "Reliably summarized" means
+   *parseable summary returned*, not *HTTP 200*. This is the behaviour the old Primary→Fallback
+   shape missed.
+3. **Tolerant parse.** `highlights`/`takeaways` coerce from a bare string (or a mixed scalar
+   array) to `[]string`, so the most common small-model quirk is absorbed **without** spending a
+   fallback round-trip — keeping the fast path fast.
+4. **Transcript truncation (`TAPIR_MAX_TRANSCRIPT_CHARS`, default 18000).** Input is bounded
+   up-front to fit a small-context primary, so overflow is prevented rather than recovered-from.
+5. **Bounded completion budget (`TAPIR_SUMMARY_MAX_TOKENS`, default 1500).** A summary needs few
+   hundred tokens; the old 8192 budget itself contributed to 8k-window overflow.
+
+**Local-first guarantee preserved.** The chain ordering *is* the guarantee: locals are tried
+first, so content reaches the external endpoint only after every local endpoint has failed.
+`TAPIR_CLOUD_FALLBACK_MODEL=""` removes the external endpoint entirely — the lever a
+**client/NDA deployment** pulls so content never leaves the local stack. With no external endpoint
+configured the `ai_routing.feature` "content only local" scenarios hold unchanged.
+
+**Reversibility.** Pure wiring + config. Setting `TAPIR_FALLBACK_MODEL` and
+`TAPIR_CLOUD_FALLBACK_MODEL` empty collapses the chain back to single-primary behaviour; the
+tolerant parse and truncation are strict supersets of the old behaviour (a previously-parseable
+reply still parses; a transcript within budget is unchanged).
+
+---
+
+## ADR-023 — Drop Shorts/livestreams at discovery to protect the caption budget
+
+**Status:** Accepted (2026-06-10). **Builds on ADR-014** (per-IP caption rate limit is the
+binding constraint) and the ADR-022 live-run findings.
+
+**Context.** The scarce resource is the unofficial timedtext caption fetch (per-egress-IP
+429, ~3 successful/pass). The first multi-user run showed the candidate set was mostly noise —
+Shorts, sub-minute clips, and live broadcasts — each of which still consumes a caption-fetch
+attempt (and a "none" result is a *completed* fetch, so it costs budget even when it yields
+nothing). Spending the rate-limited budget on content the user will not read is the waste to
+cut first; it is cheaper and lower-risk than raising the ceiling (multi-IP, Whisper).
+
+Duration and live status are NOT in the playlistItems discovery response, but they ARE in the
+Data API `videos.list` (contentDetails.duration + snippet.liveBroadcastContent) — the official
+**quota-based** API (1 unit/call, 50 ids/call), which is a *different* limit from the timedtext
+429. So one cheap quota call buys a filter that saves many expensive throttled fetches.
+
+**Decision.**
+1. `NewVideos` enriches its candidates with a single `videos.list` call and drops, before
+   returning: videos shorter than `TAPIR_MIN_VIDEO_SECONDS` (default 60) and any `live`/
+   `upcoming` broadcast. Dropped videos are never persisted, so they also declutter the list.
+2. The filter is **degrade-open**: `MinVideoSeconds=0` disables it (no quota call), and a
+   `videos.list` error returns the candidates unfiltered — discovery must never break because a
+   metadata call hiccuped (worst case = pre-ADR-023 behaviour).
+3. The paste-a-URL path (`VideoByID`) is **not** filtered — an explicit user request for a
+   specific video (even a Short) is honoured.
+
+**Reversibility.** Pure discovery-time filter + config. `TAPIR_MIN_VIDEO_SECONDS=0` restores
+the old behaviour. No schema change, no effect on already-stored videos.
+
+**Quota note.** Per-channel enrichment adds ~1 unit/channel/pass. At pilot scale (≤3 users)
+this is well under the 10k/day cap; at larger scale, batch `videos.list` across channels
+(50 ids/call) by collecting all discovered ids per pass before enriching.
+
+---
+
+## ADR-024 — Per-channel caption-availability memory
+
+**Status:** Accepted (2026-06-10). **Builds on ADR-014** (per-IP caption budget), **ADR-021**
+(shared transcript cache), **ADR-023** (Shorts filter).
+
+**Context.** After ADR-021 caches transcripts and ADR-023 drops Shorts, the remaining caption
+waste is the *first* fetch on every new video of a channel that never publishes English captions
+(foreign-language news, music, etc.). Each costs one rate-limited fetch to resolve to "none" —
+and on a throttled IP that fetch may 429 and churn the backoff machinery before it ever gets a
+verdict. A pilot user's feed had several such channels.
+
+**Decision.** Remember, per `(user, channel)`, a streak of consecutive no-caption outcomes
+(`channel_caption_state`, migration 016, RLS-scoped like the rest of the user-owned schema).
+Once the streak reaches `TAPIR_CHANNEL_CAPTIONLESS_THRESHOLD` (default 5) the channel is
+suppressed — its videos are discovered/listed but not caption-fetched — for
+`TAPIR_CHANNEL_CAPTIONLESS_WINDOW` (default 14d), after which one video is re-probed
+(auto-recovery for a channel that starts adding captions). A successful fetch resets the streak;
+a fresh 429 does NOT count (transient, not a caption verdict). An explicit manual request
+bypasses suppression. `threshold = 0` disables the feature.
+
+**Why per-user, not global.** Caption availability is really a channel property (public), so a
+global table would let users share the learning. But subscriptions are per-user (ADR-012) and at
+pilot scale users' channel sets barely overlap, so per-user + RLS keeps it consistent with the
+existing isolation model with no new non-RLS exception to justify. Promoting to a shared table
+(like transcripts, ADR-021) is a future optimisation if channel overlap grows.
+
+**Reversibility.** Migration 016 is a clean drop; `threshold = 0` disables at runtime. The
+memory only ever *suppresses fetches* — it never deletes content or affects already-stored
+summaries.
+
+---
+
 ## Rejected alternatives
 
 Approaches considered during the 2026-06-02 planning + grill session and **deliberately not

cmd/tapir/main.go

@@ -136,7 +136,8 @@ func cmdRun(ctx context.Context, log *slog.Logger) error {
 
 	r := runner.New(engine.Source, st, engine, cfg.UserID, log,
 		runner.WithBackoff(cfg.FetchBackoff),
-		runner.WithAutoWindow(cfg.AutoSummarizeWindow))
+		runner.WithAutoWindow(cfg.AutoSummarizeWindow),
+		runner.WithCaptionMemory(cfg.ChannelCaptionlessThreshold, cfg.ChannelCaptionlessWindow))
 
 	log.Info("starting run", "user", cfg.UserID, "model", cfg.SummarizerModel,
 		"gateway", cfg.GatewayURL, "poll_interval", cfg.PollInterval, "fetch_backoff", cfg.FetchBackoff,

cmd/tapir/processor.go

@@ -3,6 +3,7 @@ package main
 import (
 	"context"
 	"fmt"
+	"strings"
 
 	"gitea.d-ma.be/mathias/tapir/internal/adapters/llm"
 	"gitea.d-ma.be/mathias/tapir/internal/adapters/secrets"
@@ -42,6 +43,40 @@ func (f videoFetcher) FetchVideo(ctx context.Context, userID, videoID string) (d
 // queue-only fallback: the web UI keeps working (the button just queues) and
 // `tapir run` reports the gap via its own ValidateForRun. Missing engine config
 // is never an error here.
+// buildSummarizer wires the summarization endpoint chain (ADR-022) shared by the
+// web "Summarize now" path and the scheduler's per-user runners. The chain is:
+// primary (local, fast) → local fallback → cloud fallback (worst case). Each
+// endpoint reaches the same LiteLLM gateway with a different model alias — the
+// gateway fronts both llama-swap and berget — so a fallback is just a different
+// alias, not a second client config. Empty model entries are skipped, so a
+// client deployment can set the cloud fallback empty to keep content local.
+func buildSummarizer(cfg config.Config) *summarizer.Summarizer {
+	mk := func(model string) summarizer.Endpoint {
+		return summarizer.Endpoint{
+			Client:   llm.New(cfg.GatewayURL, cfg.GatewayKey, model, cfg.SummarizerTimeout, llm.WithMaxTokens(cfg.SummaryMaxTokens)),
+			Provider: providerOf(model),
+			Model:    model,
+		}
+	}
+	eps := []summarizer.Endpoint{mk(cfg.SummarizerModel)}
+	if cfg.FallbackModel != "" && cfg.FallbackModel != cfg.SummarizerModel {
+		eps = append(eps, mk(cfg.FallbackModel))
+	}
+	if cfg.CloudFallbackModel != "" && cfg.CloudFallbackModel != cfg.SummarizerModel {
+		eps = append(eps, mk(cfg.CloudFallbackModel))
+	}
+	return summarizer.NewChain(eps, cfg.MaxTranscriptChars)
+}
+
+// providerOf maps a model alias to the domain AIProvider recorded on summaries.
+// A "berget/" alias is an external provider; everything else is the local stack.
+func providerOf(model string) string {
+	if strings.HasPrefix(model, "berget/") {
+		return "berget"
+	}
+	return "local"
+}
+
 func buildProcessor(cfg config.Config, st *store.Store) (*usecase.Engine, error) {
 	if cfg.GatewayURL == "" || cfg.YTClientID == "" || cfg.YTClientSecret == "" || cfg.SecretsFile == "" {
 		return nil, nil
@@ -53,15 +88,10 @@ func buildProcessor(cfg config.Config, st *store.Store) (*usecase.Engine, error)
 		ClientSecret:       cfg.YTClientSecret,
 		TokenSecretRef:     cfg.YTTokenRef,
 		PreferredLanguages: []string{"en"},
+		MinVideoSeconds:    cfg.MinVideoSeconds,
 	}, secretStore)
 
-	// Local Primary only; no BYO fallback for the demo (fallback nil).
-	primary := summarizer.Endpoint{
-		Client:   llm.New(cfg.GatewayURL, cfg.GatewayKey, cfg.SummarizerModel, cfg.SummarizerTimeout),
-		Provider: "local",
-		Model:    cfg.SummarizerModel,
-	}
-	sum := summarizer.New(primary, nil)
+	sum := buildSummarizer(cfg)
 
 	// The store is both the summary sink and the shared transcript cache (ADR-021):
 	// the engine reads stored transcripts before any caption fetch and writes

cmd/tapir/scheduler.go

@@ -6,9 +6,7 @@ import (
 	"log/slog"
 	"time"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/llm"
 	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/summarizer"
 	"gitea.d-ma.be/mathias/tapir/internal/adapters/youtube"
 	"gitea.d-ma.be/mathias/tapir/internal/config"
 	"gitea.d-ma.be/mathias/tapir/internal/ports"
@@ -35,18 +33,20 @@ func buildUserRunner(cfg config.Config, st *store.Store, secretStore ports.Secre
 		ClientSecret:       cfg.YTClientSecret,
 		TokenSecretRef:     web.YouTubeTokenRef(userID),
 		PreferredLanguages: []string{"en"},
+		MinVideoSeconds:    cfg.MinVideoSeconds,
 	}, secretStore)
 
-	primary := summarizer.Endpoint{
-		Client:   llm.New(cfg.GatewayURL, cfg.GatewayKey, cfg.SummarizerModel, cfg.SummarizerTimeout),
-		Provider: "local",
-		Model:    cfg.SummarizerModel,
-	}
-	engine := usecase.NewEngine(src, summarizer.New(primary, nil), st)
+	engine := usecase.NewEngine(src, buildSummarizer(cfg), st)
+	// Share the transcript cache (ADR-021) on the scheduler path too — without
+	// this every scheduled pass re-fetches transcripts it already had, burning the
+	// scarce per-IP caption budget (ADR-014) and starving other users. The web
+	// "Summarize now" path already sets this; the scheduler omitting it was a bug.
+	engine.Transcripts = st
 
 	return runner.New(src, st, engine, userID, log,
 		runner.WithBackoff(cfg.FetchBackoff),
-		runner.WithAutoWindow(cfg.AutoSummarizeWindow)), nil
+		runner.WithAutoWindow(cfg.AutoSummarizeWindow),
+		runner.WithCaptionMemory(cfg.ChannelCaptionlessThreshold, cfg.ChannelCaptionlessWindow)), nil
 }
 
 // userLister enumerates every registered user and reports a user's video
@@ -64,6 +64,7 @@ type userLister interface {
 // isolation). Returns the stats summed across users.
 func runDiscoveryPass(
 	ctx context.Context,
+	pass int,
 	lister userLister,
 	runUser func(context.Context, string) (runner.Stats, error),
 	log *slog.Logger,
@@ -74,15 +75,17 @@ func runDiscoveryPass(
 		return runner.Stats{}
 	}
 
-	log.Info("scheduler: starting discovery pass", "users", len(users))
-	var total runner.Stats
+	// Keep only users with a video connection. A pass for a connectionless user
+	// (e.g. a stale Dex-era orphan identity) only tries to resolve a token that
+	// was never minted, logging a spurious "ref not found" every tick. Filtering
+	// here — BEFORE rotation — also keeps fairness honest: rotation is over the
+	// users that actually consume the caption budget, so a dead identity can't eat
+	// a rotation slot and skew the lead share.
+	var connected []store.UserIdentity
 	for _, u := range users {
 		if ctx.Err() != nil {
-			break // shutting down: stop enumerating
+			return runner.Stats{} // shutting down
 		}
-		// Skip users with no video connection. A discovery pass for them only
-		// attempts to resolve a token that was never minted, logging a spurious
-		// "ref not found" every tick (e.g. stale Dex-era orphan identities).
 		conns, err := lister.ConnectionsForUser(ctx, u.UserID)
 		if err != nil {
 			log.Warn("scheduler: list connections failed", "user", u.UserID, "err", err)
@@ -92,6 +95,22 @@ func runDiscoveryPass(
 			log.Debug("scheduler: skipping user with no video connections", "user", u.UserID)
 			continue
 		}
+		connected = append(connected, u)
+	}
+
+	// Rotate who goes first each pass. Caption fetches share one per-egress-IP
+	// rate budget (ADR-014); whoever runs first each pass spends the pre-throttle
+	// window, so a FIXED order permanently starves whoever is last (a new pilot
+	// user got 0 fetches for 12h while the first-listed user got all of them).
+	// Rotation over the connected set gives each real user the lead in turn.
+	connected = rotateUsers(connected, pass)
+
+	log.Info("scheduler: starting discovery pass", "users", len(connected))
+	var total runner.Stats
+	for _, u := range connected {
+		if ctx.Err() != nil {
+			break // shutting down: stop enumerating
+		}
 		stats, err := runUser(ctx, u.UserID)
 		total = sumStats(total, stats)
 		if err != nil {
@@ -103,6 +122,7 @@ func runDiscoveryPass(
 		"skipped_seen", total.SkippedSeen, "skipped_no_text", total.SkippedNoText,
 		"skipped_manual", total.SkippedManual, "skipped_too_old", total.SkippedTooOld,
 		"skipped_rate_limited", total.SkippedRateLimited,
+		"skipped_no_caption_channel", total.SkippedNoCaptionChannel,
 		"channel_unavailable", total.ChannelUnavailable, "errors", total.Errors)
 	return total
 }
@@ -123,7 +143,8 @@ func runScheduler(
 		return // disabled
 	}
 
-	runDiscoveryPass(ctx, lister, runUser, log)
+	pass := 0
+	runDiscoveryPass(ctx, pass, lister, runUser, log)
 
 	ticker := time.NewTicker(interval)
 	defer ticker.Stop()
@@ -132,23 +153,45 @@ func runScheduler(
 		case <-ctx.Done():
 			return
 		case <-ticker.C:
-			runDiscoveryPass(ctx, lister, runUser, log)
+			pass++
+			runDiscoveryPass(ctx, pass, lister, runUser, log)
 		}
 	}
 }
 
+// rotateUsers left-rotates users by pass positions so a different user leads each
+// pass. With n users, user i leads on every pass where pass ≡ i (mod n). A pass
+// offset that is negative or exceeds n is normalised. Order within the rotation
+// is otherwise preserved, so the set of users run is unchanged — only who is
+// first (and thus wins the scarce caption-fetch budget) rotates.
+func rotateUsers(users []store.UserIdentity, pass int) []store.UserIdentity {
+	n := len(users)
+	if n <= 1 {
+		return users
+	}
+	off := ((pass % n) + n) % n
+	if off == 0 {
+		return users
+	}
+	out := make([]store.UserIdentity, 0, n)
+	out = append(out, users[off:]...)
+	out = append(out, users[:off]...)
+	return out
+}
+
 // sumStats adds two passes' stats field-wise, so runDiscoveryPass can report a
 // per-tick aggregate across all users.
 func sumStats(a, b runner.Stats) runner.Stats {
 	return runner.Stats{
-		Candidates:         a.Candidates + b.Candidates,
-		Summarized:         a.Summarized + b.Summarized,
-		SkippedSeen:        a.SkippedSeen + b.SkippedSeen,
-		SkippedNoText:      a.SkippedNoText + b.SkippedNoText,
-		SkippedManual:      a.SkippedManual + b.SkippedManual,
-		SkippedTooOld:      a.SkippedTooOld + b.SkippedTooOld,
-		SkippedRateLimited: a.SkippedRateLimited + b.SkippedRateLimited,
-		ChannelUnavailable: a.ChannelUnavailable + b.ChannelUnavailable,
-		Errors:             a.Errors + b.Errors,
+		Candidates:              a.Candidates + b.Candidates,
+		Summarized:              a.Summarized + b.Summarized,
+		SkippedSeen:             a.SkippedSeen + b.SkippedSeen,
+		SkippedNoText:           a.SkippedNoText + b.SkippedNoText,
+		SkippedManual:           a.SkippedManual + b.SkippedManual,
+		SkippedTooOld:           a.SkippedTooOld + b.SkippedTooOld,
+		SkippedRateLimited:      a.SkippedRateLimited + b.SkippedRateLimited,
+		SkippedNoCaptionChannel: a.SkippedNoCaptionChannel + b.SkippedNoCaptionChannel,
+		ChannelUnavailable:      a.ChannelUnavailable + b.ChannelUnavailable,
+		Errors:                  a.Errors + b.Errors,
 	}
 }

cmd/tapir/scheduler_test.go

@@ -45,6 +45,7 @@ func (f fakeLister) ConnectionsForUser(_ context.Context, userID string) ([]stor
 type countingRunUser struct {
 	mu      sync.Mutex
 	calls   map[string]int
+	order   []string // userIDs in the order they were run, across all passes
 	failFor map[string]bool
 }
 
@@ -60,12 +61,19 @@ func (c *countingRunUser) run(_ context.Context, userID string) (runner.Stats, e
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	c.calls[userID]++
+	c.order = append(c.order, userID)
 	if c.failFor[userID] {
 		return runner.Stats{Errors: 1}, errors.New("boom")
 	}
 	return runner.Stats{Summarized: 1}, nil
 }
 
+func (c *countingRunUser) runOrder() []string {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return append([]string(nil), c.order...)
+}
+
 func (c *countingRunUser) count(userID string) int {
 	c.mu.Lock()
 	defer c.mu.Unlock()
@@ -94,7 +102,7 @@ func TestDiscoveryPassRunsEveryUserOnce(t *testing.T) {
 	lister := fakeLister{users: usersN("a", "b", "c")}
 	rc := newCountingRunUser()
 
-	stats := runDiscoveryPass(context.Background(), lister, rc.run, quietLog())
+	stats := runDiscoveryPass(context.Background(), 0, lister, rc.run, quietLog())
 
 	require.Equal(t, 1, rc.count("a"))
 	require.Equal(t, 1, rc.count("b"))
@@ -102,13 +110,46 @@ func TestDiscoveryPassRunsEveryUserOnce(t *testing.T) {
 	require.Equal(t, 3, stats.Summarized, "stats are summed across users")
 }
 
+// Caption fetches share one per-IP budget; a fixed user order starves whoever is
+// last. Each pass must rotate which user leads so the lead slot is shared.
+func TestDiscoveryPassRotatesLeadUser(t *testing.T) {
+	lister := fakeLister{users: usersN("a", "b", "c")}
+	rc := newCountingRunUser()
+
+	runDiscoveryPass(context.Background(), 0, lister, rc.run, quietLog())
+	runDiscoveryPass(context.Background(), 1, lister, rc.run, quietLog())
+	runDiscoveryPass(context.Background(), 2, lister, rc.run, quietLog())
+
+	require.Equal(t, []string{"a", "b", "c", "b", "c", "a", "c", "a", "b"}, rc.runOrder(),
+		"each pass left-rotates the user order so every user leads in turn")
+	// Fairness: over a full rotation cycle every user ran the same number of times.
+	require.Equal(t, 3, rc.count("a"))
+	require.Equal(t, 3, rc.count("b"))
+	require.Equal(t, 3, rc.count("c"))
+}
+
+// A connectionless orphan must not consume a rotation slot: rotation is over the
+// connected users only, so two real users alternate the lead 50/50 even with a
+// dead identity listed between them.
+func TestDiscoveryPassRotationIgnoresConnectionlessUsers(t *testing.T) {
+	lister := fakeLister{users: usersN("a", "orphan", "c"), noConn: map[string]bool{"orphan": true}}
+	rc := newCountingRunUser()
+
+	runDiscoveryPass(context.Background(), 0, lister, rc.run, quietLog())
+	runDiscoveryPass(context.Background(), 1, lister, rc.run, quietLog())
+
+	require.Equal(t, []string{"a", "c", "c", "a"}, rc.runOrder(),
+		"only connected users rotate; the orphan never runs and never holds a slot")
+	require.Equal(t, 0, rc.count("orphan"))
+}
+
 func TestDiscoveryPassSkipsUsersWithoutConnections(t *testing.T) {
 	// b never connected a video source (e.g. a stale Dex-era orphan identity).
 	// It must be skipped silently — not run and logged as a token error every pass.
 	lister := fakeLister{users: usersN("a", "b", "c"), noConn: map[string]bool{"b": true}}
 	rc := newCountingRunUser()
 
-	stats := runDiscoveryPass(context.Background(), lister, rc.run, quietLog())
+	stats := runDiscoveryPass(context.Background(), 0, lister, rc.run, quietLog())
 
 	require.Equal(t, 1, rc.count("a"))
 	require.Equal(t, 0, rc.count("b"), "a user with no connection must be skipped, not run")
@@ -121,7 +162,7 @@ func TestDiscoveryPassOneUserFailureDoesNotStopOthers(t *testing.T) {
 	lister := fakeLister{users: usersN("a", "b", "c")}
 	rc := newCountingRunUser("b") // user b's pass errors
 
-	stats := runDiscoveryPass(context.Background(), lister, rc.run, quietLog())
+	stats := runDiscoveryPass(context.Background(), 0, lister, rc.run, quietLog())
 
 	require.Equal(t, 1, rc.count("a"))
 	require.Equal(t, 1, rc.count("b"))
@@ -134,7 +175,7 @@ func TestDiscoveryPassListerErrorIsContained(t *testing.T) {
 	lister := fakeLister{err: errors.New("db down")}
 	rc := newCountingRunUser()
 
-	stats := runDiscoveryPass(context.Background(), lister, rc.run, quietLog())
+	stats := runDiscoveryPass(context.Background(), 0, lister, rc.run, quietLog())
 
 	require.Equal(t, 0, rc.total(), "no users enumerated → no passes")
 	require.Equal(t, runner.Stats{}, stats)

docs/homelab-integration.md

@@ -27,6 +27,28 @@ it** — endpoints and aliases drift, and this file is a snapshot (2026-06-06),
   `iguana/deepseek-r1-14b`) is preferred for summary quality if its latency/output is acceptable.
   The `max_tokens` fix below means thinking models no longer return empty content, so they are now
   viable choices, not blocked ones. Do not assume a coder alias is right for prose.
+- **Summarizer fallback chain (ADR-022).** The primary alias is the *first* of an ordered chain;
+  on failure or unparseable output the summarizer advances to the next model. All reached through
+  the same gateway by alias.
+  - `TAPIR_FALLBACK_MODEL` — local fallback. **Default `koala/phi4-14b`.** Empty disables it.
+  - `TAPIR_CLOUD_FALLBACK_MODEL` — worst-case EXTERNAL fallback. **Default `berget/mistral-small`.**
+    **Set this empty (`""`) for any client/NDA deployment** so content never leaves the local
+    stack — the chain then contains only local endpoints.
+  - `TAPIR_SUMMARY_MAX_TOKENS` — per-summary completion budget. **Default `1500`.** Small on
+    purpose: with the old 8192 budget, prompt + completion overflowed `phi4-mini`'s 8k window.
+  - `TAPIR_MAX_TRANSCRIPT_CHARS` — transcript truncation budget sent to the model. **Default
+    `18000`** (~fits an 8k-context model). `0` disables truncation. Prevents the context-overflow
+    HTTP 400 a long transcript caused on `phi4-mini`.
+- **Discovery low-value filter (ADR-023).** `TAPIR_MIN_VIDEO_SECONDS` — **default `60`**. At
+  discovery, `NewVideos` enriches candidates with one cheap `videos.list` call (quota API, NOT
+  the timedtext 429 path) and drops videos shorter than this plus any live/upcoming broadcast,
+  so the scarce caption-fetch budget isn't spent on Shorts. `0` disables the filter. The
+  paste-a-URL path is never filtered.
+- **Per-channel caption memory (ADR-024).** `TAPIR_CHANNEL_CAPTIONLESS_THRESHOLD` — **default
+  `5`** consecutive no-caption results before a channel is suppressed (its videos listed but not
+  caption-fetched). `TAPIR_CHANNEL_CAPTIONLESS_WINDOW` — **default `336h`** (14d) suppression
+  before one video is re-probed. `THRESHOLD=0` disables. A successful fetch resets the channel;
+  a 429 does not count; an explicit manual request bypasses suppression.
 - **Thinking models need an explicit `max_tokens`.** qwen3 / deepseek-r1 spend the budget on
   reasoning and return **empty content** if `max_tokens` is too low (or unset). The summarizer's
   parser treats an empty summary as an error for exactly this reason. **Done (2026-06-02, Worker F):**

docs/use-cases/ai_routing.feature

@@ -31,5 +31,17 @@ Feature: Local-first AI with optional BYO fallback
     When any transcript is summarized
     Then my content is only ever sent to the local AI stack
 
-  # "Reliably" is operationalized as: Primary returned without error within timeout.
+  Scenario: A model returns unparseable output and the next endpoint succeeds
+    Given the local AI stack is available
+    But the primary model returns output that cannot be parsed into a summary
+    And a fallback model is configured
+    When a transcript is summarized
+    Then Tapir falls back to the next model in the chain
+    And the summary records fallback_used as true
+
+  # "Reliably" is operationalized as: an endpoint returned a PARSEABLE summary
+  # within timeout. A 200 with malformed JSON (or highlights emitted as a bare
+  # string) counts as a failure and advances the chain (ADR-022). Endpoints are
+  # tried in order, locals first, so the external worst-case model only ever sees
+  # content after every local endpoint has failed.
   # Quality scoring may be added later without changing these scenarios.

internal/adapters/llm/client.go

@@ -34,15 +34,35 @@ type Client struct {
 	httpClient *http.Client
 }
 
+// Option configures a Client at construction. Variadic so the existing 4-arg
+// call sites stay valid as new knobs are added.
+type Option func(*Client)
+
+// WithMaxTokens overrides the per-request completion budget. The summarizer uses
+// this to cap completion for small-context models (e.g. koala/phi4-mini, 8k):
+// with the default 8192 budget, prompt + max_tokens overflows an 8k context and
+// the gateway returns HTTP 400. A non-positive n is ignored (keeps the default).
+func WithMaxTokens(n int) Option {
+	return func(c *Client) {
+		if n > 0 {
+			c.maxTokens = n
+		}
+	}
+}
+
 // New constructs a Client.
-func New(baseURL, apiKey, model string, timeout time.Duration) *Client {
-	return &Client{
+func New(baseURL, apiKey, model string, timeout time.Duration, opts ...Option) *Client {
+	c := &Client{
 		baseURL:    strings.TrimRight(baseURL, "/"),
 		apiKey:     apiKey,
 		model:      model,
 		maxTokens:  defaultMaxTokens,
 		httpClient: &http.Client{Timeout: timeout},
 	}
+	for _, opt := range opts {
+		opt(c)
+	}
+	return c
 }
 
 type chatRequest struct {

internal/adapters/llm/client_test.go

@@ -64,6 +64,27 @@ func TestClient_SendsMaxTokens(t *testing.T) {
 	}
 }
 
+// TestClient_WithMaxTokens overrides the completion budget — the summarizer caps
+// it small so prompt + max_tokens fits a small-context model's window (8k).
+func TestClient_WithMaxTokens(t *testing.T) {
+	var body chatRequest
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"choices": []map[string]any{{"message": map[string]any{"content": "ok"}}},
+		})
+	}))
+	defer srv.Close()
+
+	c := New(srv.URL, "", "test-model", 10*time.Second, WithMaxTokens(1500))
+	if _, err := c.Complete(context.Background(), "sys", "user"); err != nil {
+		t.Fatalf("Complete: %v", err)
+	}
+	if body.MaxTokens != 1500 {
+		t.Errorf("max_tokens = %d, want 1500", body.MaxTokens)
+	}
+}
+
 func TestClient_ReturnsErrorOnNon200(t *testing.T) {
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		http.Error(w, "overloaded", http.StatusServiceUnavailable)

internal/adapters/store/channel_caption.go

@@ -0,0 +1,83 @@
+package store
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"github.com/jackc/pgx/v5"
+)
+
+// CaptionlessChannels returns the set of channel ids currently suppressed for the
+// user — channels whose recent videos all yielded no captions, within their
+// suppression window (ADR-024). The runner skips caption fetches for these
+// channels' videos. A channel whose window has expired is not returned, so its
+// next video is re-probed (auto-recovery).
+func (s *Store) CaptionlessChannels(ctx context.Context, userID string) (map[string]bool, error) {
+	out := map[string]bool{}
+	err := s.withUser(ctx, userID, func(tx pgx.Tx) error {
+		rows, err := tx.Query(ctx, `
+			SELECT channel_id FROM channel_caption_state
+			WHERE user_id = $1 AND captionless_until IS NOT NULL AND captionless_until > now()`,
+			userID)
+		if err != nil {
+			return fmt.Errorf("store: caption-less channels: %w", err)
+		}
+		defer rows.Close()
+		for rows.Next() {
+			var ch string
+			if err := rows.Scan(&ch); err != nil {
+				return fmt.Errorf("store: scan caption-less channel: %w", err)
+			}
+			out[ch] = true
+		}
+		return rows.Err()
+	})
+	return out, err
+}
+
+// RecordChannelCaptionOutcome updates a channel's caption-availability memory
+// after a fetch attempt (ADR-024). hadCaptions resets the channel (consecutive
+// count to 0, suppression cleared). Otherwise the consecutive no-caption count is
+// incremented; once it reaches threshold the channel is suppressed for window.
+// threshold <= 0 is a no-op (feature disabled). An empty channelID is ignored
+// (some sources may not carry one).
+func (s *Store) RecordChannelCaptionOutcome(ctx context.Context, userID, channelID string, hadCaptions bool, threshold int, window time.Duration) error {
+	if channelID == "" || threshold <= 0 {
+		return nil
+	}
+	return s.withUser(ctx, userID, func(tx pgx.Tx) error {
+		if hadCaptions {
+			_, err := tx.Exec(ctx, `
+				INSERT INTO channel_caption_state (user_id, channel_id, consecutive_none, captionless_until, updated_at)
+				VALUES ($1, $2, 0, NULL, now())
+				ON CONFLICT (user_id, channel_id)
+				DO UPDATE SET consecutive_none = 0, captionless_until = NULL, updated_at = now()`,
+				userID, channelID)
+			if err != nil {
+				return fmt.Errorf("store: reset channel caption state: %w", err)
+			}
+			return nil
+		}
+		// No captions: increment the streak; suppress once it reaches threshold.
+		// captionless_until is set from the NEW count inside the same statement so
+		// the decision is atomic with the increment.
+		until := time.Now().Add(window)
+		_, err := tx.Exec(ctx, `
+			INSERT INTO channel_caption_state (user_id, channel_id, consecutive_none, captionless_until, updated_at)
+			VALUES ($1, $2, 1, CASE WHEN 1 >= $3 THEN $4::timestamptz ELSE NULL END, now())
+			ON CONFLICT (user_id, channel_id)
+			DO UPDATE SET
+				consecutive_none = channel_caption_state.consecutive_none + 1,
+				captionless_until = CASE
+					WHEN channel_caption_state.consecutive_none + 1 >= $3 THEN $4::timestamptz
+					ELSE channel_caption_state.captionless_until
+				END,
+				updated_at = now()`,
+			userID, channelID, threshold, until)
+		if err != nil {
+			return fmt.Errorf("store: record channel no-caption: %w", err)
+		}
+		return nil
+	})
+}

internal/adapters/store/channel_caption_test.go

@@ -0,0 +1,67 @@
+package store_test
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestChannelCaptionMemory_SuppressesAfterThreshold(t *testing.T) {
+	ctx := context.Background()
+	s := newStore(t)
+	super := rawPool(t)
+	resetDB(t, super)
+	seedUser(t, super, userA)
+
+	const threshold = 3
+	window := time.Hour
+
+	// Below threshold: not yet suppressed.
+	require.NoError(t, s.RecordChannelCaptionOutcome(ctx, userA, "chanX", false, threshold, window))
+	require.NoError(t, s.RecordChannelCaptionOutcome(ctx, userA, "chanX", false, threshold, window))
+	got, err := s.CaptionlessChannels(ctx, userA)
+	require.NoError(t, err)
+	require.NotContains(t, got, "chanX", "2 < threshold 3: not suppressed yet")
+
+	// Crossing the threshold suppresses the channel.
+	require.NoError(t, s.RecordChannelCaptionOutcome(ctx, userA, "chanX", false, threshold, window))
+	got, err = s.CaptionlessChannels(ctx, userA)
+	require.NoError(t, err)
+	require.Contains(t, got, "chanX", "3 consecutive no-caption results suppress the channel")
+
+	// A successful caption fetch resets it.
+	require.NoError(t, s.RecordChannelCaptionOutcome(ctx, userA, "chanX", true, threshold, window))
+	got, err = s.CaptionlessChannels(ctx, userA)
+	require.NoError(t, err)
+	require.NotContains(t, got, "chanX", "a captioned video clears suppression")
+}
+
+func TestChannelCaptionMemory_WindowExpiryReProbes(t *testing.T) {
+	ctx := context.Background()
+	s := newStore(t)
+	super := rawPool(t)
+	resetDB(t, super)
+	seedUser(t, super, userA)
+
+	// A negative window means captionless_until lands in the past — modelling an
+	// elapsed suppression window, which must make the channel eligible again.
+	require.NoError(t, s.RecordChannelCaptionOutcome(ctx, userA, "chanY", false, 1, -time.Hour))
+	got, err := s.CaptionlessChannels(ctx, userA)
+	require.NoError(t, err)
+	require.NotContains(t, got, "chanY", "an expired window re-enables the channel for a re-probe")
+}
+
+func TestChannelCaptionMemory_DisabledThresholdIsNoOp(t *testing.T) {
+	ctx := context.Background()
+	s := newStore(t)
+	super := rawPool(t)
+	resetDB(t, super)
+	seedUser(t, super, userA)
+
+	require.NoError(t, s.RecordChannelCaptionOutcome(ctx, userA, "chanZ", false, 0, time.Hour))
+	got, err := s.CaptionlessChannels(ctx, userA)
+	require.NoError(t, err)
+	require.Empty(t, got, "threshold 0 disables the memory — nothing recorded")
+}

internal/adapters/store/migrate_test.go

@@ -53,7 +53,9 @@ func TestMigration010LoginEventsUpDown(t *testing.T) {
 	require.True(t, loginEventsExists(t), "login_events must exist at latest migration")
 
 	m := fileMigrator(t)
-	// 011..015 sit above 010; step them down first so 010 is exercised in isolation.
+	// 011..016 sit above 010; step them down first so 010 is exercised in isolation.
+	require.NoError(t, m.Steps(-1), "down 016 drops channel_caption_state, login_events intact")
+	require.True(t, loginEventsExists(t), "016 down leaves login_events intact")
 	require.NoError(t, m.Steps(-1), "down 015 reshapes transcripts, login_events intact")
 	require.True(t, loginEventsExists(t), "015 down leaves login_events intact")
 	require.NoError(t, m.Steps(-1), "down 014 drops channel_title, login_events intact")
@@ -68,7 +70,7 @@ func TestMigration010LoginEventsUpDown(t *testing.T) {
 	require.NoError(t, m.Steps(-1), "down 010 must drop login_events")
 	require.False(t, loginEventsExists(t), "login_events must be gone after the down migration")
 
-	require.NoError(t, m.Steps(6), "up must recreate 010 then re-apply 011..015")
+	require.NoError(t, m.Steps(7), "up must recreate 010 then re-apply 011..016")
 	require.True(t, loginEventsExists(t), "login_events must be restored after the up migration")
 }
 
@@ -91,6 +93,7 @@ func TestMigration011AutoSummarizeDefaultUpDown(t *testing.T) {
 	require.Equal(t, "true", autoSummarizeDefault(t), "011 sets the default to TRUE")
 
 	m := fileMigrator(t)
+	require.NoError(t, m.Steps(-1), "down 016 drops channel_caption_state")
 	require.NoError(t, m.Steps(-1), "down 015 reshapes transcripts")
 	require.NoError(t, m.Steps(-1), "down 014 drops channel_title")
 	require.NoError(t, m.Steps(-1), "down 013 drops channel_errors")
@@ -104,6 +107,7 @@ func TestMigration011AutoSummarizeDefaultUpDown(t *testing.T) {
 	require.NoError(t, m.Steps(1), "up 013 creates channel_errors")
 	require.NoError(t, m.Steps(1), "up 014 recreates channel_title")
 	require.NoError(t, m.Steps(1), "up 015 reshapes transcripts to shared")
+	require.NoError(t, m.Steps(1), "up 016 recreates channel_caption_state (HEAD)")
 }
 
 // channelTitleExists reports whether videos.channel_title is present.
@@ -123,6 +127,8 @@ func TestMigration014VideoChannelTitleUpDown(t *testing.T) {
 	require.True(t, channelTitleExists(t), "channel_title exists at latest migration")
 
 	m := fileMigrator(t)
+	require.NoError(t, m.Steps(-1), "down 016 drops channel_caption_state, channel_title intact")
+	require.True(t, channelTitleExists(t), "016 down leaves channel_title intact")
 	require.NoError(t, m.Steps(-1), "down 015 reshapes transcripts, channel_title intact")
 	require.True(t, channelTitleExists(t), "015 down leaves channel_title intact")
 	require.NoError(t, m.Steps(-1), "down 014 must drop channel_title")
@@ -130,7 +136,8 @@ func TestMigration014VideoChannelTitleUpDown(t *testing.T) {
 
 	require.NoError(t, m.Steps(1), "up 014 must recreate channel_title")
 	require.True(t, channelTitleExists(t), "channel_title must be restored after the up migration")
-	require.NoError(t, m.Steps(1), "up 015 restores the shared transcripts shape (HEAD)")
+	require.NoError(t, m.Steps(1), "up 015 restores the shared transcripts shape")
+	require.NoError(t, m.Steps(1), "up 016 recreates channel_caption_state (HEAD)")
 }
 
 // TestMigration012FixAutoSummarizeRLS proves 012 runs cleanly and flips any

internal/adapters/store/migrations/016_channel_caption_state.down.sql

@@ -0,0 +1 @@
+DROP TABLE channel_caption_state;

internal/adapters/store/migrations/016_channel_caption_state.up.sql

@@ -0,0 +1,30 @@
+-- Migration 016: per-(user, channel) caption-availability memory (ADR-024).
+--
+-- Some channels never publish English captions (foreign-language news, music,
+-- etc.). Each of their new videos still costs ONE rate-limited caption fetch
+-- (ADR-014) before resolving to "none" — and on a throttled egress IP that fetch
+-- may 429 and churn through the backoff machinery first. This table remembers
+-- channels that repeatedly yield no captions so discovery can stop attempting
+-- their videos, freeing the scarce fetch budget for channels that do have them.
+--
+-- consecutive_none counts no-caption outcomes in a row; a successful fetch resets
+-- it to 0. Once it crosses the threshold the channel is suppressed until
+-- captionless_until, after which one video is re-probed (auto-recovery for a
+-- channel that starts adding captions). Per-user + RLS-scoped, consistent with
+-- the rest of the user-owned schema (subscriptions are per-user; ADR-012).
+CREATE TABLE channel_caption_state (
+    user_id           UUID        NOT NULL REFERENCES users(id) ON DELETE CASCADE,
+    channel_id        TEXT        NOT NULL,
+    consecutive_none  INT         NOT NULL DEFAULT 0,
+    captionless_until TIMESTAMPTZ,
+    updated_at        TIMESTAMPTZ NOT NULL DEFAULT now(),
+    PRIMARY KEY (user_id, channel_id)
+);
+
+CREATE INDEX idx_channel_caption_state_user_id ON channel_caption_state(user_id);
+
+ALTER TABLE channel_caption_state ENABLE ROW LEVEL SECURITY;
+ALTER TABLE channel_caption_state FORCE ROW LEVEL SECURITY;
+CREATE POLICY channel_caption_state_isolation ON channel_caption_state
+    FOR ALL
+    USING (user_id = current_setting('tapir.current_user_id', true)::uuid);

internal/adapters/summarizer/summarizer.go

@@ -1,17 +1,21 @@
 // Package summarizer implements ports.Summarizer backed by the copied llm
-// package's local-Primary -> BYO-Fallback routing (ADR-004). It is the only
-// place content ever leaves the engine toward an AI model, so it is also the
-// enforcement point for the local-first guarantee in
-// docs/use-cases/ai_routing.feature: a user with no BYO provider configured has
-// their content sent to the local stack and nowhere else.
+// package's routing (ADR-004, extended by ADR-022). It is the only place content
+// ever leaves the engine toward an AI model, so it is also the enforcement point
+// for the local-first guarantee in docs/use-cases/ai_routing.feature: endpoints
+// are tried in order, locals first, so content only reaches an external model
+// after every local endpoint has failed — and never at all when no external
+// endpoint is configured.
 package summarizer
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"strings"
 	"time"
+	"unicode/utf8"
 
 	"gitea.d-ma.be/mathias/tapir/internal/domain"
 )
@@ -29,19 +33,41 @@ type Endpoint struct {
 	Model    string // resolved alias, e.g. "iguana/deepseek-r1-14b"
 }
 
-// Summarizer routes a transcript through the local endpoint first, then the
-// optional BYO endpoint. It owns its routing (rather than delegating to
-// llm.Router) so it can record which provider answered and whether the fallback
-// was used — information llm.Router collapses away.
+// Summarizer routes a transcript through an ordered chain of endpoints, trying
+// each in turn until one returns a parseable summary. It owns its routing
+// (rather than delegating to llm.Router) so it can record which provider answered
+// and whether a fallback was used — information llm.Router collapses away. The
+// chain ordering is the local-first guarantee: callers place local endpoints
+// first and any external endpoint last, so content only reaches an external model
+// after every local endpoint has failed.
 type Summarizer struct {
-	primary  Endpoint
-	fallback *Endpoint // nil => no BYO; primary errors are returned, never sent externally
-	now      func() time.Time
+	endpoints     []Endpoint
+	maxInputChars int // transcript truncation budget; 0 = no limit
+	now           func() time.Time
 }
 
-// New constructs a Summarizer. fallback may be nil (no BYO provider configured).
+// New constructs a Summarizer from a primary endpoint and an optional fallback
+// (the historical local-Primary -> BYO-Fallback shape, ADR-004). A nil fallback
+// means a single-endpoint chain: errors are returned, content never leaves it.
 func New(primary Endpoint, fallback *Endpoint) *Summarizer {
-	return &Summarizer{primary: primary, fallback: fallback, now: time.Now}
+	eps := []Endpoint{primary}
+	if fallback != nil {
+		eps = append(eps, *fallback)
+	}
+	return &Summarizer{endpoints: eps, now: time.Now}
+}
+
+// NewChain constructs a Summarizer over an ordered endpoint chain (ADR-022).
+// endpoints are tried in order; the first to return a parseable summary wins, and
+// FallbackUsed is recorded true for any endpoint past the first. maxInputChars
+// bounds the transcript text sent to every endpoint (0 = unbounded), so a long
+// transcript does not overflow a small-context primary model's window. It panics
+// on an empty chain — a wiring bug, not a runtime condition.
+func NewChain(endpoints []Endpoint, maxInputChars int) *Summarizer {
+	if len(endpoints) == 0 {
+		panic("summarizer: NewChain requires at least one endpoint")
+	}
+	return &Summarizer{endpoints: endpoints, maxInputChars: maxInputChars, now: time.Now}
 }
 
 const systemPrompt = `You are Tapir, a video-summarization assistant.
@@ -53,31 +79,35 @@ Respond with ONLY a JSON object, no prose and no code fences:
 - "takeaways": the actionable conclusions a viewer should leave with.
 Output the JSON object and nothing else.`
 
-// Summarize implements ports.Summarizer.
+// Summarize implements ports.Summarizer. It walks the endpoint chain in order:
+// the first endpoint whose reply parses into a non-empty summary wins. An
+// endpoint is considered failed — and the next one tried — when the model call
+// errors OR when its reply cannot be parsed (a 200 with malformed JSON or a
+// highlights field the model emitted as a bare string). Truncation is applied
+// once, up front, so every endpoint sees the same bounded prompt. When the whole
+// chain fails, the joined error is returned so the engine queues the work for
+// retry and delivers no summary.
 func (s *Summarizer) Summarize(ctx context.Context, v domain.Video, t domain.Transcript) (domain.Summary, error) {
 	if !t.HasText() {
 		return domain.Summary{}, fmt.Errorf("summarize: transcript for video %s has no text", v.ID)
 	}
-	user := buildUserPrompt(v, t)
+	user := buildUserPrompt(v, t, s.maxInputChars)
 
-	// Primary = local stack. Only on its failure is anything sent externally,
-	// and only when a BYO fallback is configured.
-	out, err := s.primary.Client.Complete(ctx, systemPrompt, user)
-	if err == nil {
-		return s.build(v, s.primary, false, out)
+	var errs []error
+	for i, ep := range s.endpoints {
+		out, err := ep.Client.Complete(ctx, systemPrompt, user)
+		if err != nil {
+			errs = append(errs, fmt.Errorf("%s/%s call: %w", ep.Provider, ep.Model, err))
+			continue
+		}
+		sum, perr := s.build(v, ep, i > 0, out)
+		if perr != nil {
+			errs = append(errs, fmt.Errorf("%s/%s output: %w", ep.Provider, ep.Model, perr))
+			continue
+		}
+		return sum, nil
 	}
-
-	if s.fallback == nil {
-		// No BYO: content was sent to the local stack only. Surface the error so
-		// the engine can queue the work for retry; deliver no summary.
-		return domain.Summary{}, fmt.Errorf("summarize: local AI failed and no BYO provider configured: %w", err)
-	}
-
-	out, ferr := s.fallback.Client.Complete(ctx, systemPrompt, user)
-	if ferr != nil {
-		return domain.Summary{}, fmt.Errorf("summarize: local AI failed: %w; BYO %s failed: %v", err, s.fallback.Provider, ferr)
-	}
-	return s.build(v, *s.fallback, true, out)
+	return domain.Summary{}, fmt.Errorf("summarize: all %d endpoint(s) failed: %w", len(s.endpoints), errors.Join(errs...))
 }
 
 func (s *Summarizer) build(v domain.Video, ep Endpoint, fallbackUsed bool, raw string) (domain.Summary, error) {
@@ -89,8 +119,8 @@ func (s *Summarizer) build(v domain.Video, ep Endpoint, fallbackUsed bool, raw s
 		UserID:       v.UserID,
 		VideoID:      v.ID,
 		Summary:      parsed.Summary,
-		Highlights:   parsed.Highlights,
-		Takeaways:    parsed.Takeaways,
+		Highlights:   []string(parsed.Highlights),
+		Takeaways:    []string(parsed.Takeaways),
 		AIProvider:   ep.Provider,
 		AIModel:      ep.Model,
 		FallbackUsed: fallbackUsed,
@@ -98,20 +128,94 @@ func (s *Summarizer) build(v domain.Video, ep Endpoint, fallbackUsed bool, raw s
 	}, nil
 }
 
-func buildUserPrompt(v domain.Video, t domain.Transcript) string {
+func buildUserPrompt(v domain.Video, t domain.Transcript, maxInputChars int) string {
 	var b strings.Builder
 	fmt.Fprintf(&b, "Title: %s\n", v.Title)
 	if v.URL != "" {
 		fmt.Fprintf(&b, "URL: %s\n", v.URL)
 	}
-	fmt.Fprintf(&b, "\nTranscript:\n%s", t.Content)
+	fmt.Fprintf(&b, "\nTranscript:\n%s", truncate(t.Content, maxInputChars))
 	return b.String()
 }
 
+// truncate caps content to max bytes on a UTF-8 rune boundary, appending a
+// marker so the model knows the transcript was cut. A non-positive max (or a
+// content already within budget) returns content unchanged. Bounding the input
+// keeps a long transcript from overflowing a small-context model's window — the
+// production failure mode where koala/phi4-mini's 8k context returned HTTP 400 on
+// a 11.6k-token transcript.
+func truncate(content string, max int) string {
+	if max <= 0 || len(content) <= max {
+		return content
+	}
+	cut := max
+	for cut > 0 && !utf8.RuneStart(content[cut]) {
+		cut--
+	}
+	return content[:cut] + "\n…[transcript truncated to fit the model context]"
+}
+
+// flexStrings is a []string that also unmarshals from a single JSON string or a
+// JSON array of scalars. Small local models (koala/phi4-mini) sometimes emit
+// "highlights": "one point" instead of an array, or mix in a number; rather than
+// fail the whole summary on that quirk, coerce to []string. Empty/whitespace
+// elements are dropped.
+type flexStrings []string
+
+func (f *flexStrings) UnmarshalJSON(b []byte) error {
+	b = bytes.TrimSpace(b)
+	if len(b) == 0 || string(b) == "null" {
+		*f = nil
+		return nil
+	}
+	if b[0] == '[' {
+		var raw []json.RawMessage
+		if err := json.Unmarshal(b, &raw); err != nil {
+			return err
+		}
+		out := make([]string, 0, len(raw))
+		for _, r := range raw {
+			s, err := rawToString(r)
+			if err != nil {
+				return err
+			}
+			if strings.TrimSpace(s) != "" {
+				out = append(out, s)
+			}
+		}
+		*f = out
+		return nil
+	}
+	s, err := rawToString(b)
+	if err != nil {
+		return err
+	}
+	if strings.TrimSpace(s) == "" {
+		*f = nil
+	} else {
+		*f = flexStrings{s}
+	}
+	return nil
+}
+
+// rawToString renders a JSON scalar as text: a quoted string is unquoted; any
+// other scalar (number, bool) is kept as its literal source so no content is lost.
+func rawToString(r json.RawMessage) (string, error) {
+	r = bytes.TrimSpace(r)
+	if len(r) > 0 && r[0] == '"' {
+		var s string
+		if err := json.Unmarshal(r, &s); err != nil {
+			return "", err
+		}
+		return s, nil
+	}
+	return string(r), nil
+}
+
 type parsedSummary struct {
-	Summary    string   `json:"summary"`
-	Highlights []string `json:"highlights"`
-	Takeaways  []string `json:"takeaways"`
+	Summary    string      `json:"summary"`
+	Highlights flexStrings `json:"highlights"`
+	Takeaways  flexStrings `json:"takeaways"`
 }
 
 // parse extracts the JSON object from a model reply. Thinking models (qwen3,

internal/adapters/summarizer/summarizer_test.go

@@ -129,8 +129,8 @@ func TestSummarize_NoBYO_ContentOnlyLocal(t *testing.T) {
 	local := &fakeClient{reply: goodReply}
 	s := New(Endpoint{Client: local, Provider: "local", Model: "iguana/deepseek-r1-14b"}, nil)
 
-	if s.fallback != nil {
-		t.Fatal("no BYO configured but fallback endpoint is non-nil")
+	if len(s.endpoints) != 1 {
+		t.Fatalf("no BYO configured but chain has %d endpoints, want 1", len(s.endpoints))
 	}
 	for i := 0; i < 3; i++ {
 		sum, err := s.Summarize(context.Background(), testVideo(), testTranscript())
@@ -176,3 +176,97 @@ func TestParse_EmptySummaryRejected(t *testing.T) {
 		t.Fatal("want error for empty summary (thinking model returned no content)")
 	}
 }
+
+// parse tolerates a small model emitting "highlights" as a bare string instead
+// of an array — the production koala/phi4-mini quirk that errored with
+// "cannot unmarshal string into Go struct field ... highlights of type []string".
+func TestParse_ToleratesStringHighlights(t *testing.T) {
+	p, err := parse(`{"summary":"s","highlights":"one big point","takeaways":["a","b"]}`)
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(p.Highlights) != 1 || p.Highlights[0] != "one big point" {
+		t.Errorf("highlights = %v, want [\"one big point\"]", p.Highlights)
+	}
+	if len(p.Takeaways) != 2 {
+		t.Errorf("takeaways = %v, want 2", p.Takeaways)
+	}
+}
+
+// Chain: an endpoint that returns a 200 with unparseable output is treated as a
+// failure, and the next endpoint in the chain is tried. This is the case the old
+// primary->fallback shape missed — a parse error short-circuited instead of
+// falling back.
+func TestSummarize_FallsBackOnMalformedOutput(t *testing.T) {
+	bad := &fakeClient{reply: `{"summary": not json`}
+	good := &fakeClient{reply: goodReply}
+	s := NewChain([]Endpoint{
+		{Client: bad, Provider: "local", Model: "koala/phi4-mini"},
+		{Client: good, Provider: "local", Model: "koala/phi4-14b"},
+	}, 0)
+
+	sum, err := s.Summarize(context.Background(), testVideo(), testTranscript())
+	if err != nil {
+		t.Fatalf("Summarize: %v", err)
+	}
+	if sum.AIModel != "koala/phi4-14b" {
+		t.Errorf("AIModel = %q, want koala/phi4-14b (fell back past malformed primary)", sum.AIModel)
+	}
+	if !sum.FallbackUsed {
+		t.Error("FallbackUsed = false, want true")
+	}
+	if bad.calls != 1 || good.calls != 1 {
+		t.Errorf("calls: bad=%d good=%d, want 1 and 1", bad.calls, good.calls)
+	}
+}
+
+// Chain: when every endpoint fails, no summary is produced and the joined error
+// names each failure so the engine queues the work for retry.
+func TestSummarize_ChainAllEndpointsFail(t *testing.T) {
+	a := &fakeClient{err: errors.New("context overflow")}
+	b := &fakeClient{reply: "not even json"}
+	s := NewChain([]Endpoint{
+		{Client: a, Provider: "local", Model: "m1"},
+		{Client: b, Provider: "berget", Model: "m2"},
+	}, 0)
+
+	if _, err := s.Summarize(context.Background(), testVideo(), testTranscript()); err == nil {
+		t.Fatal("want error when all endpoints fail")
+	}
+	if a.calls != 1 || b.calls != 1 {
+		t.Errorf("calls: a=%d b=%d, want 1 and 1", a.calls, b.calls)
+	}
+}
+
+// A transcript longer than the chain's input budget is truncated before it
+// reaches any model, so a small-context primary does not overflow its window.
+func TestSummarize_TruncatesLongTranscript(t *testing.T) {
+	local := &fakeClient{reply: goodReply}
+	const budget = 100
+	s := NewChain([]Endpoint{{Client: local, Provider: "local", Model: "m"}}, budget)
+
+	long := domain.Transcript{
+		VideoID: "vid-1", UserID: "user-1", Source: domain.SourceCaptions,
+		Content: strings.Repeat("word ", 1000), // 5000 bytes, well over budget
+	}
+	if _, err := s.Summarize(context.Background(), testVideo(), long); err != nil {
+		t.Fatalf("Summarize: %v", err)
+	}
+	// The prompt carries title/URL framing plus the truncation marker, so allow
+	// headroom over the raw transcript budget — but it must be far below 5000.
+	if len(local.lastUser) > budget+300 {
+		t.Errorf("prompt length = %d, want <= %d (transcript not truncated)", len(local.lastUser), budget+300)
+	}
+	if !strings.Contains(local.lastUser, "truncated") {
+		t.Error("truncation marker missing from prompt")
+	}
+}
+
+func TestNewChain_PanicsOnEmptyChain(t *testing.T) {
+	defer func() {
+		if recover() == nil {
+			t.Fatal("want panic on empty endpoint chain")
+		}
+	}()
+	NewChain(nil, 0)
+}

internal/adapters/youtube/youtube.go

@@ -66,6 +66,13 @@ type Config struct {
 	// poll. Zero means defaultMaxVideos.
 	MaxVideosPerSubscription int
 
+	// MinVideoSeconds drops videos shorter than this from discovery (Shorts/clips,
+	// ADR-023). NewVideos enriches candidates with a single cheap videos.list call
+	// (contentDetails.duration + snippet.liveBroadcastContent) and filters before
+	// returning, so the scarce caption-fetch budget is never spent on them. Live
+	// and upcoming broadcasts are dropped too. Zero disables the filter.
+	MinVideoSeconds int
+
 	// BaseURL overrides the Data API root. Empty means defaultBaseURL.
 	BaseURL string
 
@@ -242,7 +249,97 @@ func (a *Adapter) NewVideos(ctx context.Context, sub domain.Subscription) ([]dom
 			break
 		}
 	}
-	return videos, nil
+
+	// Drop Shorts/sub-minute clips and live/upcoming broadcasts before they ever
+	// reach the rate-limited caption path (ADR-023). One cheap videos.list call
+	// (quota API, not the timedtext throttle) supplies duration + live status.
+	return a.filterLowValue(ctx, client, videos), nil
+}
+
+// filterLowValue removes videos shorter than cfg.MinVideoSeconds and any live or
+// upcoming broadcast, using a single videos.list lookup for duration +
+// liveBroadcastContent. The filter is best-effort: if MinVideoSeconds is 0 (off)
+// or the lookup fails, the input is returned unfiltered — discovery must not break
+// because a metadata call hiccuped; the worst case is the pre-ADR-023 behaviour.
+func (a *Adapter) filterLowValue(ctx context.Context, client *http.Client, videos []domain.Video) []domain.Video {
+	if a.cfg.MinVideoSeconds <= 0 || len(videos) == 0 {
+		return videos
+	}
+
+	ids := make([]string, 0, len(videos))
+	for _, v := range videos {
+		ids = append(ids, v.ProviderVideoID)
+	}
+	q := url.Values{
+		"part": {"contentDetails,snippet"},
+		"id":   {strings.Join(ids, ",")},
+	}
+	var resp videoListResponse
+	if err := a.getJSON(ctx, client, "/videos", q, &resp); err != nil {
+		// Degrade open: keep the candidates rather than lose discovery.
+		return videos
+	}
+
+	type meta struct {
+		seconds int
+		live    string
+	}
+	byID := make(map[string]meta, len(resp.Items))
+	for _, it := range resp.Items {
+		byID[it.ID] = meta{seconds: parseISO8601Seconds(it.ContentDetails.Duration), live: it.Snippet.LiveBroadcastContent}
+	}
+
+	kept := videos[:0]
+	for _, v := range videos {
+		m, ok := byID[v.ProviderVideoID]
+		if !ok {
+			kept = append(kept, v) // unknown metadata: keep, let the fetch decide
+			continue
+		}
+		if m.live != "" && m.live != "none" {
+			continue // live or upcoming broadcast
+		}
+		if m.seconds > 0 && m.seconds < a.cfg.MinVideoSeconds {
+			continue // Short / sub-threshold clip
+		}
+		kept = append(kept, v)
+	}
+	return kept
+}
+
+// parseISO8601Seconds parses an ISO 8601 duration as returned by the YouTube Data
+// API (e.g. "PT1H2M3S", "PT45S", "PT3M") into seconds. Only the hour/minute/second
+// components YouTube emits are handled; an unparseable or zero value returns 0,
+// which the caller treats as "unknown" (not filtered on duration).
+func parseISO8601Seconds(d string) int {
+	if !strings.HasPrefix(d, "PT") {
+		return 0
+	}
+	d = d[2:]
+	total, num := 0, 0
+	seen := false
+	for _, r := range d {
+		switch {
+		case r >= '0' && r <= '9':
+			num = num*10 + int(r-'0')
+			seen = true
+		case r == 'H':
+			total += num * 3600
+			num, seen = 0, false
+		case r == 'M':
+			total += num * 60
+			num, seen = 0, false
+		case r == 'S':
+			total += num
+			num, seen = 0, false
+		default:
+			return 0 // unexpected component (days/weeks) — treat as unknown
+		}
+	}
+	if seen {
+		return 0 // trailing digits without a unit: malformed
+	}
+	return total
 }
 
 // VideoByID fetches a single video's metadata (videos.list, snippet) for an
@@ -377,11 +474,16 @@ type playlistItemListResponse struct {
 
 type videoListResponse struct {
 	Items []struct {
+		ID      string `json:"id"`
 		Snippet struct {
-			Title        string    `json:"title"`
-			ChannelTitle string    `json:"channelTitle"`
-			PublishedAt  time.Time `json:"publishedAt"`
+			Title                string    `json:"title"`
+			ChannelTitle         string    `json:"channelTitle"`
+			PublishedAt          time.Time `json:"publishedAt"`
+			LiveBroadcastContent string    `json:"liveBroadcastContent"`
 		} `json:"snippet"`
+		ContentDetails struct {
+			Duration string `json:"duration"` // ISO 8601, e.g. "PT1M30S"
+		} `json:"contentDetails"`
 	} `json:"items"`
 }
 

internal/adapters/youtube/youtube_test.go

@@ -186,6 +186,91 @@ func TestNewVideosCapsAtMax(t *testing.T) {
 	}
 }
 
+// TestNewVideosFiltersShortsAndLive: with MinVideoSeconds set, discovery enriches
+// candidates via videos.list and drops sub-threshold clips (Shorts) and
+// live/upcoming broadcasts before they reach the rate-limited caption path.
+func TestNewVideosFiltersShortsAndLive(t *testing.T) {
+	a, _ := newTestAdapter(t, func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/playlistItems":
+			_, _ = w.Write([]byte(`{
+				"items": [
+					{"snippet": {"title": "Real Talk", "publishedAt": "2026-06-03T10:00:00Z", "resourceId": {"videoId": "long1"}}},
+					{"snippet": {"title": "A Short", "publishedAt": "2026-06-03T09:00:00Z", "resourceId": {"videoId": "short1"}}},
+					{"snippet": {"title": "Live Now", "publishedAt": "2026-06-03T08:00:00Z", "resourceId": {"videoId": "live1"}}}
+				]
+			}`))
+		case "/videos":
+			if got := r.URL.Query().Get("part"); got != "contentDetails,snippet" {
+				t.Errorf("videos.list part=%q, want contentDetails,snippet", got)
+			}
+			_, _ = w.Write([]byte(`{
+				"items": [
+					{"id": "long1", "contentDetails": {"duration": "PT12M30S"}, "snippet": {"liveBroadcastContent": "none"}},
+					{"id": "short1", "contentDetails": {"duration": "PT45S"}, "snippet": {"liveBroadcastContent": "none"}},
+					{"id": "live1", "contentDetails": {"duration": "PT0S"}, "snippet": {"liveBroadcastContent": "live"}}
+				]
+			}`))
+		default:
+			t.Errorf("unexpected path %q", r.URL.Path)
+		}
+	})
+	a.cfg.MinVideoSeconds = 60
+
+	vids, err := a.NewVideos(context.Background(), domain.Subscription{ID: "s1", UserID: "u1", ChannelID: "UC_acme"})
+	if err != nil {
+		t.Fatalf("NewVideos: %v", err)
+	}
+	if len(vids) != 1 || vids[0].ProviderVideoID != "long1" {
+		t.Fatalf("expected only long1 to survive the filter, got %+v", vids)
+	}
+}
+
+// TestNewVideosNoFilterWhenDisabled: MinVideoSeconds=0 keeps the pre-ADR-023
+// behaviour — no videos.list call, no filtering.
+func TestNewVideosNoFilterWhenDisabled(t *testing.T) {
+	a, _ := newTestAdapter(t, func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/videos" {
+			t.Errorf("videos.list must not be called when MinVideoSeconds is 0")
+		}
+		_, _ = w.Write([]byte(`{"items": [
+			{"snippet": {"title": "A Short", "publishedAt": "2026-06-03T09:00:00Z", "resourceId": {"videoId": "short1"}}}
+		]}`))
+	})
+	a.cfg.MinVideoSeconds = 0
+
+	vids, err := a.NewVideos(context.Background(), domain.Subscription{ID: "s1", UserID: "u1", ChannelID: "UC_acme"})
+	if err != nil {
+		t.Fatalf("NewVideos: %v", err)
+	}
+	if len(vids) != 1 {
+		t.Fatalf("filter disabled must keep all videos, got %d", len(vids))
+	}
+}
+
+func TestParseISO8601Seconds(t *testing.T) {
+	cases := []struct {
+		in   string
+		want int
+	}{
+		{"PT45S", 45},
+		{"PT1M30S", 90},
+		{"PT3M", 180},
+		{"PT1H2M3S", 3723},
+		{"PT2H", 7200},
+		{"PT0S", 0},
+		{"", 0},
+		{"garbage", 0},
+		{"P1D", 0},  // days component not handled → unknown
+		{"PT10", 0}, // trailing digits without a unit → malformed
+	}
+	for _, c := range cases {
+		if got := parseISO8601Seconds(c.in); got != c.want {
+			t.Errorf("parseISO8601Seconds(%q) = %d, want %d", c.in, got, c.want)
+		}
+	}
+}
+
 // TestUploadsPlaylistID covers the zero-cost UC->UU derivation, including
 // non-standard ids that must fall through unchanged (handled via fallback).
 func TestUploadsPlaylistID(t *testing.T) {

internal/config/config.go

@@ -28,8 +28,38 @@ type Config struct {
 	GatewayURL string
 	// GatewayKey authorizes the gateway. Read from env, never committed.
 	GatewayKey string
-	// SummarizerModel is the alias in host/name form, e.g. "koala/phi4-mini".
+	// SummarizerModel is the primary summarizer alias in host/name form, tried
+	// first on every video, e.g. "koala/phi4-mini".
 	SummarizerModel string
+	// FallbackModel is the LOCAL fallback alias tried when the primary fails or
+	// returns unparseable output (ADR-022). Kept local so content stays on the
+	// homelab stack. Empty disables it. Default a bigger-context local model.
+	FallbackModel string
+	// CloudFallbackModel is the worst-case EXTERNAL fallback alias, tried only
+	// after every local endpoint has failed (ADR-022). For client deployments set
+	// this empty so content never leaves the local stack. Default a berget alias.
+	CloudFallbackModel string
+	// SummaryMaxTokens caps the completion budget per summary call. Small-context
+	// models (koala/phi4-mini, 8k) overflow when prompt + max_tokens exceeds the
+	// window; a summary needs only a few hundred tokens, so the default is small.
+	SummaryMaxTokens int
+	// MaxTranscriptChars bounds the transcript text sent to the model so a long
+	// transcript does not overflow a small-context primary. 0 disables truncation.
+	MaxTranscriptChars int
+
+	// MinVideoSeconds drops videos shorter than this from discovery (Shorts and
+	// other sub-minute clips that are noise and waste the scarce caption-fetch
+	// budget, ADR-014/ADR-023). Enforced via a cheap Data API videos.list lookup at
+	// discovery, never the rate-limited caption path. 0 disables the filter.
+	MinVideoSeconds int
+
+	// ChannelCaptionlessThreshold is how many consecutive no-caption results a
+	// channel may yield before its videos are suppressed from caption fetching
+	// (ADR-024). 0 disables the per-channel caption memory entirely.
+	ChannelCaptionlessThreshold int
+	// ChannelCaptionlessWindow is how long a suppressed channel stays suppressed
+	// before one video is re-probed (auto-recovery for a channel that adds captions).
+	ChannelCaptionlessWindow time.Duration
 	// SummarizerTimeout bounds a single completion call. Thinking models are
 	// slow, so the default is generous.
 	SummarizerTimeout time.Duration
@@ -118,6 +148,13 @@ func (c Config) DexConfigured() bool { return strings.TrimSpace(c.OIDCIssuer) !=
 const (
 	defaultGatewayURL            = "http://koala:30401/v1"
 	defaultSummarizerModel       = "koala/phi4-mini"
+	defaultFallbackModel         = "koala/phi4-14b"
+	defaultCloudFallbackModel    = "berget/mistral-small"
+	defaultSummaryMaxTokens      = 1500
+	defaultMaxTranscriptChars    = 18000
+	defaultMinVideoSeconds       = 60
+	defaultCaptionlessThreshold  = 5
+	defaultCaptionlessWindow     = 14 * 24 * time.Hour
 	defaultSummarizerTimeout     = 5 * time.Minute
 	defaultYTTokenRef            = "youtube/refresh_token"
 	defaultYTConnectRedirectURL  = "https://tapir.d-ma.be/oauth/youtube/callback"
@@ -141,6 +178,8 @@ func Load() (Config, error) {
 		GatewayURL:           envOr("TAPIR_GATEWAY_URL", defaultGatewayURL),
 		GatewayKey:           os.Getenv("TAPIR_GATEWAY_KEY"),
 		SummarizerModel:      envOr("TAPIR_SUMMARIZER_MODEL", defaultSummarizerModel),
+		FallbackModel:        lookupOr("TAPIR_FALLBACK_MODEL", defaultFallbackModel),
+		CloudFallbackModel:   lookupOr("TAPIR_CLOUD_FALLBACK_MODEL", defaultCloudFallbackModel),
 		DBDSN:                os.Getenv("TAPIR_DB_DSN"),
 		YTClientID:           os.Getenv("TAPIR_YT_CLIENT_ID"),
 		YTClientSecret:       os.Getenv("TAPIR_YT_CLIENT_SECRET"),
@@ -193,6 +232,45 @@ func Load() (Config, error) {
 	}
 	c.AutoSummarizeWindow = autoWindow
 
+	summaryTokens, err := intOr("TAPIR_SUMMARY_MAX_TOKENS", defaultSummaryMaxTokens)
+	if err != nil {
+		return Config{}, err
+	}
+	c.SummaryMaxTokens = summaryTokens
+
+	maxChars, err := intOr("TAPIR_MAX_TRANSCRIPT_CHARS", defaultMaxTranscriptChars)
+	if err != nil {
+		return Config{}, err
+	}
+	if maxChars < 0 {
+		maxChars = 0
+	}
+	c.MaxTranscriptChars = maxChars
+
+	minVideo, err := intOr("TAPIR_MIN_VIDEO_SECONDS", defaultMinVideoSeconds)
+	if err != nil {
+		return Config{}, err
+	}
+	if minVideo < 0 {
+		minVideo = 0
+	}
+	c.MinVideoSeconds = minVideo
+
+	captionThreshold, err := intOr("TAPIR_CHANNEL_CAPTIONLESS_THRESHOLD", defaultCaptionlessThreshold)
+	if err != nil {
+		return Config{}, err
+	}
+	if captionThreshold < 0 {
+		captionThreshold = 0
+	}
+	c.ChannelCaptionlessThreshold = captionThreshold
+
+	captionWindow, err := durationOr("TAPIR_CHANNEL_CAPTIONLESS_WINDOW", defaultCaptionlessWindow)
+	if err != nil {
+		return Config{}, err
+	}
+	c.ChannelCaptionlessWindow = captionWindow
+
 	onboard, err := intOr("TAPIR_ONBOARD_SUMMARIZE_COUNT", defaultOnboardSummarizeCount)
 	if err != nil {
 		return Config{}, err
@@ -269,6 +347,17 @@ func envOr(key, fallback string) string {
 	return fallback
 }
 
+// lookupOr returns the env value when the key is PRESENT (even if empty), else
+// fallback. Unlike envOr it lets an explicit empty value override the default —
+// needed to DISABLE an optional fallback model (e.g. set the cloud fallback empty
+// for a client deployment so content never leaves the local stack).
+func lookupOr(key, fallback string) string {
+	if v, ok := os.LookupEnv(key); ok {
+		return v
+	}
+	return fallback
+}
+
 func intOr(key string, fallback int) (int, error) {
 	v := os.Getenv(key)
 	if v == "" {

internal/config/config_test.go

@@ -1,11 +1,29 @@
 package config
 
 import (
+	"os"
 	"strings"
 	"testing"
 	"time"
 )
 
+// unset removes an env key for the duration of the test, restoring it after.
+// Needed to observe a default for a key read with LookupEnv (where present-empty
+// means "explicitly disabled", not "use default").
+func unset(t *testing.T, key string) {
+	t.Helper()
+	if old, ok := os.LookupEnv(key); ok {
+		t.Cleanup(func() {
+			if err := os.Setenv(key, old); err != nil {
+				t.Fatalf("restore %s: %v", key, err)
+			}
+		})
+	}
+	if err := os.Unsetenv(key); err != nil {
+		t.Fatalf("unset %s: %v", key, err)
+	}
+}
+
 // setEnv sets env vars for the test and clears them afterward, so cases don't
 // leak into one another. t.Setenv handles restoration.
 func setEnv(t *testing.T, kv map[string]string) {
@@ -53,6 +71,46 @@ func TestLoad_AppliesDefaults(t *testing.T) {
 	}
 }
 
+func TestLoad_SummarizerChainDefaults(t *testing.T) {
+	setEnv(t, map[string]string{
+		"TAPIR_SUMMARIZER_MODEL":     "",
+		"TAPIR_SUMMARY_MAX_TOKENS":   "",
+		"TAPIR_MAX_TRANSCRIPT_CHARS": "",
+	})
+	unset(t, "TAPIR_FALLBACK_MODEL")
+	unset(t, "TAPIR_CLOUD_FALLBACK_MODEL")
+
+	c, err := Load()
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if c.FallbackModel != defaultFallbackModel {
+		t.Errorf("FallbackModel = %q, want %q", c.FallbackModel, defaultFallbackModel)
+	}
+	if c.CloudFallbackModel != defaultCloudFallbackModel {
+		t.Errorf("CloudFallbackModel = %q, want %q", c.CloudFallbackModel, defaultCloudFallbackModel)
+	}
+	if c.SummaryMaxTokens != defaultSummaryMaxTokens {
+		t.Errorf("SummaryMaxTokens = %d, want %d", c.SummaryMaxTokens, defaultSummaryMaxTokens)
+	}
+	if c.MaxTranscriptChars != defaultMaxTranscriptChars {
+		t.Errorf("MaxTranscriptChars = %d, want %d", c.MaxTranscriptChars, defaultMaxTranscriptChars)
+	}
+}
+
+// An explicitly empty cloud-fallback env disables external routing — the lever a
+// client deployment pulls so content never leaves the local stack.
+func TestLoad_EmptyCloudFallbackDisables(t *testing.T) {
+	t.Setenv("TAPIR_CLOUD_FALLBACK_MODEL", "")
+	c, err := Load()
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if c.CloudFallbackModel != "" {
+		t.Errorf("CloudFallbackModel = %q, want empty (disabled)", c.CloudFallbackModel)
+	}
+}
+
 func TestLoad_OnboardSummarizeCount(t *testing.T) {
 	cases := []struct {
 		name, env string

internal/runner/runner.go

@@ -27,8 +27,9 @@ import (
 // passCandidate is a video that passed all pre-filters (seen/manual/backoff)
 // and is queued for transcript fetch + summarization in this pass.
 type passCandidate struct {
-	v   domain.Video
-	pos int // discovery position — used as a stable tiebreak when published_at ties
+	v         domain.Video
+	channelID string // owning channel — keys the caption-availability memory (ADR-024)
+	pos       int    // discovery position — used as a stable tiebreak when published_at ties
 }
 
 // compareNewestFirst orders candidates by published_at descending, NULLS LAST,
@@ -74,6 +75,14 @@ type VideoStore interface {
 	// Called when NewVideos returns domain.ErrChannelUnavailable; best-effort, errors
 	// are logged and never abort the pass.
 	UpsertChannelError(ctx context.Context, userID, channelID, channelTitle string) error
+	// CaptionlessChannels returns channel ids currently suppressed because their
+	// recent videos all yielded no captions (ADR-024). The loop skips caption
+	// fetches for these channels' (non-requested) videos.
+	CaptionlessChannels(ctx context.Context, userID string) (map[string]bool, error)
+	// RecordChannelCaptionOutcome updates a channel's caption memory after a fetch:
+	// hadCaptions resets it, otherwise the no-caption streak grows and the channel
+	// is suppressed for window once it reaches threshold. A no-op when threshold<=0.
+	RecordChannelCaptionOutcome(ctx context.Context, userID, channelID string, hadCaptions bool, threshold int, window time.Duration) error
 }
 
 // Processor runs the core use case for a single video. *usecase.Engine
@@ -93,6 +102,9 @@ type Runner struct {
 	backoff    time.Duration    // rate-limit retry window; 0 = always retry
 	autoWindow time.Duration    // recency bound for auto-summarize; 0 = no bound
 	now        func() time.Time // injectable clock (tests); defaults to time.Now
+
+	captionThreshold int           // consecutive no-caption results before a channel is suppressed; 0 = feature off
+	captionWindow    time.Duration // how long a caption-less channel stays suppressed before re-probe
 }
 
 // Option configures a Runner at construction. Variadic so existing call sites
@@ -114,6 +126,14 @@ func WithClock(now func() time.Time) Option { return func(r *Runner) { r.now = n
 // bypasses the bound. 0 (the default) disables it (summarize every unseen video).
 func WithAutoWindow(d time.Duration) Option { return func(r *Runner) { r.autoWindow = d } }
 
+// WithCaptionMemory enables per-channel caption-availability suppression
+// (ADR-024): after threshold consecutive no-caption results a channel's videos
+// are skipped (no caption fetch) for window, then one is re-probed. threshold<=0
+// (the default) disables the feature entirely.
+func WithCaptionMemory(threshold int, window time.Duration) Option {
+	return func(r *Runner) { r.captionThreshold = threshold; r.captionWindow = window }
+}
+
 // New builds a Runner. A nil logger falls back to slog.Default.
 func New(src ports.VideoSource, store VideoStore, engine Processor, userID string, log *slog.Logger, opts ...Option) *Runner {
 	if log == nil {
@@ -131,15 +151,16 @@ func New(src ports.VideoSource, store VideoStore, engine Processor, userID strin
 
 // Stats summarizes one RunOnce pass.
 type Stats struct {
-	Candidates         int
-	Summarized         int
-	SkippedSeen        int
-	SkippedNoText      int
-	SkippedManual      int // discovered but not queued, in manual mode
-	SkippedTooOld      int // auto mode: published outside the recency window (not requested)
-	SkippedRateLimited int // 429'd previously and still inside the backoff window
-	Errors             int
-	ChannelUnavailable int // channels that returned HTTP 404 (deleted/private)
+	Candidates              int
+	Summarized              int
+	SkippedSeen             int
+	SkippedNoText           int
+	SkippedManual           int // discovered but not queued, in manual mode
+	SkippedTooOld           int // auto mode: published outside the recency window (not requested)
+	SkippedRateLimited      int // 429'd previously and still inside the backoff window
+	SkippedNoCaptionChannel int // channel suppressed as caption-less (ADR-024)
+	Errors                  int
+	ChannelUnavailable      int // channels that returned HTTP 404 (deleted/private)
 }
 
 // tooOld reports whether a video published at publishedAt falls outside the
@@ -213,6 +234,17 @@ func (r *Runner) RunOnce(ctx context.Context) (Stats, error) {
 		}
 	}
 
+	// Per-channel caption memory (ADR-024): channels whose recent videos all
+	// yielded no captions are suppressed so their new videos don't burn the scarce
+	// fetch budget. Loaded only when the feature is enabled (threshold > 0).
+	var captionless map[string]bool
+	if r.captionThreshold > 0 {
+		captionless, err = r.store.CaptionlessChannels(ctx, r.userID)
+		if err != nil {
+			return stats, fmt.Errorf("runner: load caption-less channels: %w", err)
+		}
+	}
+
 	subs, err := r.src.ListSubscriptions(ctx, r.userID)
 	if err != nil {
 		return stats, fmt.Errorf("runner: list subscriptions: %w", err)
@@ -273,6 +305,14 @@ func (r *Runner) RunOnce(ctx context.Context) (Stats, error) {
 				continue
 			}
 
+			// Caption-less channel (ADR-024): its recent videos all returned no
+			// captions, so skip the fetch entirely. The video is still listed
+			// (UpsertVideo above); an explicit manual request bypasses the skip.
+			if !requested[id] && captionless[sub.ChannelID] {
+				stats.SkippedNoCaptionChannel++
+				continue
+			}
+
 			// Still inside the rate-limit backoff window: skip without fetching.
 			if at, ok := rateLimited[id]; ok && r.now().Sub(at) < r.backoff {
 				stats.SkippedRateLimited++
@@ -280,7 +320,7 @@ func (r *Runner) RunOnce(ctx context.Context) (Stats, error) {
 				continue
 			}
 
-			candidates = append(candidates, passCandidate{v: v, pos: pos})
+			candidates = append(candidates, passCandidate{v: v, channelID: sub.ChannelID, pos: pos})
 			pos++
 		}
 	}
@@ -320,6 +360,11 @@ func (r *Runner) RunOnce(ctx context.Context) (Stats, error) {
 				errs = append(errs, fmt.Errorf("set none status %q: %w", c.v.ProviderVideoID, err))
 				stats.Errors++
 			}
+			// No captions: grow this channel's no-caption streak (ADR-024).
+			if err := r.store.RecordChannelCaptionOutcome(ctx, r.userID, c.channelID, false, r.captionThreshold, r.captionWindow); err != nil {
+				errs = append(errs, fmt.Errorf("record no-caption %q: %w", c.v.ProviderVideoID, err))
+				stats.Errors++
+			}
 			r.log.Info("skipped video (no transcript)", "video", c.v.ProviderVideoID, "title", c.v.Title)
 		case res.Summary != nil:
 			stats.Summarized++
@@ -327,6 +372,11 @@ func (r *Runner) RunOnce(ctx context.Context) (Stats, error) {
 				errs = append(errs, fmt.Errorf("set fetched status %q: %w", c.v.ProviderVideoID, err))
 				stats.Errors++
 			}
+			// Captions present: reset this channel's caption memory (ADR-024).
+			if err := r.store.RecordChannelCaptionOutcome(ctx, r.userID, c.channelID, true, r.captionThreshold, r.captionWindow); err != nil {
+				errs = append(errs, fmt.Errorf("record has-caption %q: %w", c.v.ProviderVideoID, err))
+				stats.Errors++
+			}
 			// In manual mode the video was explicitly queued; clear the flag so
 			// it is not re-summarized and the UI drops the "Queued" chip.
 			if !auto {
@@ -354,6 +404,7 @@ func (r *Runner) Loop(ctx context.Context, interval time.Duration) error {
 			"skipped_seen", stats.SkippedSeen, "skipped_no_text", stats.SkippedNoText,
 			"skipped_manual", stats.SkippedManual, "skipped_too_old", stats.SkippedTooOld,
 			"skipped_rate_limited", stats.SkippedRateLimited,
+			"skipped_no_caption_channel", stats.SkippedNoCaptionChannel,
 			"channel_unavailable", stats.ChannelUnavailable, "errors", stats.Errors)
 		if err != nil {
 			r.log.Warn("run pass had errors", "err", err)

internal/runner/runner_test.go

@@ -51,6 +51,13 @@ type fakeStore struct {
 	cleared     []string
 	rateLimited map[string]time.Time // id -> when 429'd (seeds the backoff window)
 	statuses    map[string]string    // id -> last SetTranscriptStatus value
+	captionless map[string]bool      // channel ids currently suppressed (ADR-024)
+	captionRecs []captionRec         // RecordChannelCaptionOutcome calls, in order
+}
+
+type captionRec struct {
+	channelID string
+	had       bool
 }
 
 func (f *fakeStore) UpsertVideo(_ context.Context, v domain.Video) (string, error) {
@@ -93,6 +100,22 @@ func (f *fakeStore) RateLimitedVideoIDs(_ context.Context, _ string) (map[string
 
 func (f *fakeStore) UpsertChannelError(_ context.Context, _, _, _ string) error { return nil }
 
+func (f *fakeStore) CaptionlessChannels(_ context.Context, _ string) (map[string]bool, error) {
+	cp := make(map[string]bool, len(f.captionless))
+	for k, v := range f.captionless {
+		cp[k] = v
+	}
+	return cp, nil
+}
+
+func (f *fakeStore) RecordChannelCaptionOutcome(_ context.Context, _, channelID string, hadCaptions bool, threshold int, _ time.Duration) error {
+	if threshold <= 0 {
+		return nil
+	}
+	f.captionRecs = append(f.captionRecs, captionRec{channelID: channelID, had: hadCaptions})
+	return nil
+}
+
 func (f *fakeStore) SetTranscriptStatus(_ context.Context, _, videoID, status string) error {
 	if f.statuses == nil {
 		f.statuses = map[string]string{}
@@ -284,6 +307,55 @@ func TestRunOnce_AutoMode_OldVideoRequestedBypassesWindow(t *testing.T) {
 	require.Len(t, sink.delivered, 1)
 }
 
+// TestRunOnce_CaptionlessChannelSkipped: a channel flagged caption-less (ADR-024)
+// has its videos skipped from fetching but still discovered/listed, while a
+// normal channel's video is summarized.
+func TestRunOnce_CaptionlessChannelSkipped(t *testing.T) {
+	src := &fakeSource{
+		subs: []domain.Subscription{sub("dead", "Dead Channel"), sub("live", "Live Channel")},
+		videos: map[string][]domain.Video{
+			"dead": {vid("d1", "Dead One")},
+			"live": {vid("l1", "Live One")},
+		},
+	}
+	st := &fakeStore{seen: map[string]bool{}, auto: true, captionless: map[string]bool{"dead": true}}
+	sink := &recordingSink{}
+	eng := usecase.NewEngine(src, fakeSummarizer{}, sink)
+	r := runner.New(src, st, eng, testUser, quietLogger(),
+		runner.WithCaptionMemory(5, 14*24*time.Hour))
+
+	stats, err := r.RunOnce(context.Background())
+	require.NoError(t, err)
+	require.Equal(t, 1, stats.SkippedNoCaptionChannel, "dead channel's video skipped from fetch")
+	require.Equal(t, 1, stats.Summarized, "live channel's video still summarized")
+	require.Len(t, st.upserted, 2, "both videos are still discovered and listed")
+}
+
+// TestRunOnce_RecordsCaptionOutcomes: a no-caption result grows the channel's
+// streak (had=false); a successful summary resets it (had=true).
+func TestRunOnce_RecordsCaptionOutcomes(t *testing.T) {
+	src := &fakeSource{
+		subs: []domain.Subscription{sub("c1", "Has Caps"), sub("c2", "No Caps")},
+		videos: map[string][]domain.Video{
+			"c1": {vid("good", "Good")},
+			"c2": {vid("bad", "Bad")},
+		},
+		transcripts: map[string]domain.Transcript{
+			"bad": {Source: domain.SourceNone}, // no usable text → engine skips
+		},
+	}
+	st := &fakeStore{seen: map[string]bool{}, auto: true}
+	sink := &recordingSink{}
+	eng := usecase.NewEngine(src, fakeSummarizer{}, sink)
+	r := runner.New(src, st, eng, testUser, quietLogger(),
+		runner.WithCaptionMemory(5, 14*24*time.Hour))
+
+	_, err := r.RunOnce(context.Background())
+	require.NoError(t, err)
+	require.Contains(t, st.captionRecs, captionRec{channelID: "c1", had: true}, "captioned channel reset")
+	require.Contains(t, st.captionRecs, captionRec{channelID: "c2", had: false}, "no-caption channel streak grown")
+}
+
 // TestRunOnce_AutoWindowZero_SummarizesOld: a zero window disables the bound —
 // the pre-recency behaviour (summarize every unseen video) is preserved.
 func TestRunOnce_AutoWindowZero_SummarizesOld(t *testing.T) {

internal/web/handlers.go

@@ -233,11 +233,20 @@ func (a *App) handleList(w http.ResponseWriter, r *http.Request) {
 	}
 	hasConnected := len(conns) > 0
 
-	if isHTMX(r) {
-		a.render(w, r, summaryList(buckets, hasConnected))
+	// Summarization mode drives the backlog copy: an auto user is told summaries
+	// land gradually; a manual user is told to click Summarize (the first pilot
+	// user sat in manual mode reading "land automatically" and waited forever).
+	autoSummarize, err := a.Store.GetAutoSummarize(r.Context(), userID)
+	if err != nil {
+		a.serverError(w, r, "summarize mode", err)
 		return
 	}
-	a.render(w, r, ListPage(buckets, f, stats, takeFlash(w, r), hasConnected, channels))
+
+	if isHTMX(r) {
+		a.render(w, r, summaryList(buckets, hasConnected, autoSummarize))
+		return
+	}
+	a.render(w, r, ListPage(buckets, f, stats, takeFlash(w, r), hasConnected, channels, autoSummarize))
 }
 
 // handleDetail renders one summary in full (highlights, takeaways, action group).

internal/web/handlers_test.go

@@ -477,3 +477,36 @@ func postAction(t *testing.T, app *web.App, videoID, action string, htmx bool) *
 	}
 	return do(t, app, req)
 }
+
+// TestListManualModeBannerCopy: a manual-mode user with un-summarized videos
+// sees the manual prompt (click Summarize), NOT the "summaries land
+// automatically" copy that misled the first pilot user into waiting forever.
+func TestListManualModeBannerCopy(t *testing.T) {
+	ctx := context.Background()
+	app := newApp(t)
+	p := rawPool(t)
+	resetDB(t, p)
+	seedVideo(t, p, videoX, "X Title", "https://x", time.Time{}) // pending, un-summarized
+	require.NoError(t, app.Store.SetAutoSummarize(ctx, userID, false))
+
+	html := body(t, do(t, app, httptest.NewRequest(http.MethodGet, "/", nil)))
+	require.Contains(t, html, "Manual mode")
+	require.Contains(t, html, "are not summarized automatically")
+	require.NotContains(t, html, "land gradually",
+		"manual-mode user must not be told summaries arrive automatically")
+}
+
+// TestListAutoModeBannerCopy: an auto-mode user with a backlog sees the
+// gradual-delivery copy, not the manual prompt.
+func TestListAutoModeBannerCopy(t *testing.T) {
+	ctx := context.Background()
+	app := newApp(t)
+	p := rawPool(t)
+	resetDB(t, p)
+	seedVideo(t, p, videoX, "X Title", "https://x", time.Time{})
+	require.NoError(t, app.Store.SetAutoSummarize(ctx, userID, true))
+
+	html := body(t, do(t, app, httptest.NewRequest(http.MethodGet, "/", nil)))
+	require.Contains(t, html, "land gradually")
+	require.NotContains(t, html, "are not summarized automatically")
+}

internal/web/paste_test.go

@@ -65,7 +65,7 @@ func TestParseYouTubeVideoID(t *testing.T) {
 func TestListPageShowsPasteFormOnlyWhenConnected(t *testing.T) {
 	render := func(connected bool) string {
 		var buf bytes.Buffer
-		if err := ListPage(listBuckets{}, Filter{}, PipelineStats{}, "", connected, nil).Render(context.Background(), &buf); err != nil {
+		if err := ListPage(listBuckets{}, Filter{}, PipelineStats{}, "", connected, nil, true).Render(context.Background(), &buf); err != nil {
 			t.Fatalf("render: %v", err)
 		}
 		return buf.String()

internal/web/views.templ

@@ -104,7 +104,7 @@ templ flashBanner(code string) {
 // #summary-list region; a non-HTMX request renders the whole page. flash carries
 // a one-shot notification (e.g. "connected", "registered") surfaced on arrival
 // after a POST→redirect.
-templ ListPage(b listBuckets, f Filter, stats PipelineStats, flash string, hasConnected bool, channels []string) {
+templ ListPage(b listBuckets, f Filter, stats PipelineStats, flash string, hasConnected bool, channels []string, autoSummarize bool) {
 	@Layout("Tapir — Summaries") {
 		@flashBanner(flash)
 		if hasConnected {
@@ -116,14 +116,23 @@ templ ListPage(b listBuckets, f Filter, stats PipelineStats, flash string, hasCo
 		if stats.RateLimited > 0 || stats.Pending > 0 || stats.NoText > 0 {
 			@pipelineBar(stats)
 		}
-		if stats.RateLimited+stats.Pending > 0 {
+		if (stats.RateLimited+stats.Pending) > 0 && autoSummarize {
 			<p class="pipeline-note muted">
 				Tapir fetches captions slowly on purpose, to respect YouTube's limits —
 				new summaries land gradually. Check back tomorrow.
 			</p>
 		}
+		if (stats.RateLimited+stats.Pending) > 0 && !autoSummarize {
+			<p class="pipeline-note muted">
+				You are in Manual mode: new videos appear here but are not summarized
+				automatically. Use the Summarize button on the ones you want.
+			</p>
+			<p class="pipeline-note muted">
+				<a href="/account">Switch to Automatic</a> to have new videos summarized for you.
+			</p>
+		}
 		<div id="summary-list">
-			@summaryList(b, hasConnected)
+			@summaryList(b, hasConnected, autoSummarize)
 		</div>
 	}
 }
@@ -203,12 +212,16 @@ templ filterForm(f Filter, channels []string) {
 // and a single disclosure holding the older un-summarized back-catalogue. Cards
 // reflow to a single column on mobile; an empty list shows a friendly first-run
 // state instead of a blank table.
-templ summaryList(b listBuckets, hasConnected bool) {
+templ summaryList(b listBuckets, hasConnected bool, autoSummarize bool) {
 	if b.empty() {
 		if hasConnected {
 			<div class="empty empty-connected">
 				<strong>Your account is connected</strong>
-				<span>Tapir is finding your subscriptions and fetching captions — summaries appear here gradually. Check back later.</span>
+				if autoSummarize {
+					<span>Tapir is finding your subscriptions and fetching captions — summaries appear here gradually. Check back later.</span>
+				} else {
+					<span>Tapir is finding your subscriptions. You are in Manual mode, so videos appear here with a Summarize button — pick the ones you want, or switch to Automatic in your account.</span>
+				}
 			</div>
 		} else {
 			<div class="empty">

test/acceptance/scenario_coverage_test.go

@@ -26,10 +26,11 @@ import (
 // longer matches a real non-pending scenario.
 var scenarioCoverage = map[string]string{
 	// ai_routing.feature
-	"Local AI produces the summary":                             "TestSummarize_LocalSucceeds",
-	"Local AI fails and the user has a BYO provider configured": "TestSummarize_FallsBackToBYO",
-	"Local AI fails and the user has no BYO provider":           "TestSummarize_LocalFailsNoBYO_NoExternalSend",
-	"A user without BYO never has content sent externally":      "TestSummarize_NoBYO_ContentOnlyLocal",
+	"Local AI produces the summary":                                     "TestSummarize_LocalSucceeds",
+	"Local AI fails and the user has a BYO provider configured":         "TestSummarize_FallsBackToBYO",
+	"Local AI fails and the user has no BYO provider":                   "TestSummarize_LocalFailsNoBYO_NoExternalSend",
+	"A user without BYO never has content sent externally":              "TestSummarize_NoBYO_ContentOnlyLocal",
+	"A model returns unparseable output and the next endpoint succeeds": "TestSummarize_FallsBackOnMalformedOutput",
 
 	// landing_page.feature
 	"An unauthenticated visit to the root is sent to the welcome page":    "TestUnauthenticatedRootRedirectsToWelcome",