feat(hls): persistent fMP4 segment cache + integrity + stats (0.9.7)

Cache keyed by sha256(absPath|quality|audioIdx)[:8] with .complete marker; LRU + size-budget eviction; per-key writer-lock; pinned during play; startup orphan reap; integrity verify on HIT; subtitle-completeness gate; hit/miss counters + daily log line. New [downloads.hls_cache] block in config.toml (enabled/size_gb/dir, default 5GB). Smoke test: 2nd play of same source+quality is 23-31× faster (HIT path skips ffmpeg entirely).
2026-05-26 23:39:02 +02:00 · 2026-05-26 23:39:02 +02:00 · 7e96976257
commit 7e96976257
parent 834c58c25a
10 changed files with 1295 additions and 9 deletions
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@ -5,6 +5,39 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

+## [0.9.7] - 2026-05-26
+
+### Added
+
+- **hls cache**: persistent fMP4 segment cache keyed by
+  `(source, quality, audio_index)`. After a successful encode the segments
+  + `init.mp4` are kept under `~/.cache/unarr/hls-cache/{key}/` with a
+  `.complete` marker. A second play of the same file at the same quality
+  skips ffmpeg entirely (smoke-tested 23–31× faster than re-encode). LRU
+  + size-budget eviction; pinned during active play; per-key writer-lock
+  prevents two concurrent encodes from corrupting each other. Startup
+  reaps orphan dirs without `.complete` older than 10 min so a daemon
+  crash doesn't leak disk indefinitely. New `[downloads.hls_cache]` block
+  in `config.toml`: `enabled` (default true), `size_gb` (default 5,
+  min 1), `dir` (default `~/.cache/unarr/hls-cache`).
+- **hls cache integrity check**: on HIT, the daemon stats `init.mp4` +
+  last segment before reporting cache reuse — if a file was externally
+  deleted, the entry is invalidated and re-encoded transparently.
+- **hls cache stats**: hit/miss counters surface via `cache.Stats()`
+  (`Hits`, `Misses`, `EntryCount`, `TotalBytes`) and the sweeper logs a
+  daily summary line `[hls_cache] day-stats: hits=N misses=M ratio=X%
+  entries=Y size=ZMB`.
+- **subtitle integrity for cached replay**: `Close` waits up to 15 s for
+  the subtitle extractor goroutine before sealing `.complete` so a HIT
+  never serves half-written `.vtt` files. Timeout invalidates instead of
+  sealing.
+
+### Changed
+
+- `[daemon] auto_upgrade` now appears in fresh `config.toml` files as
+  `true` (it was always the implicit default; this just makes it visible
+  in default-generated configs).
+
 ## [0.9.6] - 2026-05-26

 ### Added
--- a/README.md
+++ b/README.md
@ -343,6 +343,58 @@ unarr self-update --force   # reinstall even if up to date

 `unarr doctor` checks: config file, API key, server connectivity (with latency), agent registration, download directory, disk space, and version.

+### Updating unarr
+
+unarr supports three update paths. Pick whichever fits your workflow.
+
+**1. Manual self-update (always available).**
+
+```bash
+unarr self-update                # interactive update to latest
+unarr self-update --force        # reinstall same version
+unarr self-update --allow-unsigned # accept releases without checksum signature
+```
+
+The CLI downloads the new release archive over HTTPS (from
+`torrentclaw.com/releases/download/v<ver>/`), verifies SHA-256, swaps the
+binary in place (`.backup` kept next to it), and restarts the systemd
+user unit if the daemon is running.
+
+**2. Auto-apply on server signal (default, since 0.9.6).**
+
+When you press **"Force update now"** on the web (Settings → Agent → Force
+update), the server sets a flag your daemon polls every sync (~3 s). On
+the next sync the daemon downloads the new binary, replaces itself, and
+exits — `systemd Restart=always` respawns on the new version. No SSH, no
+terminal access required. Works headless on NAS / Docker.
+
+The button shows an amber warning if your agent is below 0.9.6 (older
+daemons see the signal but only log "run unarr update" — the operator
+must run the command manually that one time).
+
+**Opt out of auto-apply.** Some users prefer reviewing CHANGELOG before
+applying. Disable in `config.toml`:
+
+```toml
+[daemon]
+auto_upgrade = false
+```
+
+With `auto_upgrade = false`, pressing the web button still flags your
+agent (so the daemon logs the new version on next sync), but the daemon
+will not download / replace anything — you run `unarr self-update` when
+you're ready.
+
+**3. Docker auto-restart with a new tag.**
+
+```bash
+docker pull torrentclaw/unarr:latest
+docker compose up -d
+```
+
+Tags published: `latest`, `0.9`, `0.9.7`, ... — pin to a minor (`0.9`)
+for opt-in patch updates without surprises.
+
 ## Clean

 Remove temporary files, logs, resume data, and other artifacts generated by unarr. Shows what will be removed and asks for confirmation before deleting.
@ -424,6 +476,7 @@ tv_shows_dir = "~/Media/TV Shows"
 [daemon]
 poll_interval = "30s"
 heartbeat_interval = "30s"
+auto_upgrade = true   # apply server-flagged upgrades in-place (since 0.9.6)

 [notifications]
 enabled = true
@ -466,6 +519,40 @@ If `transcode.enabled = true` but `ffmpeg` / `ffprobe` aren't on PATH, the
 daemon logs a warning at startup and HLS sessions are rejected at runtime
 with a clear error — install ffmpeg or set `enabled = false`.

+#### `[downloads.hls_cache]` — persistent HLS segment cache
+
+```toml
+[downloads.hls_cache]
+enabled = true   # on by default
+size_gb = 5      # disk budget; LRU eviction once exceeded
+dir     = ""     # custom path; empty = ~/.cache/unarr/hls-cache
+```
+
+| Key | Type | Default | Notes |
+|-----|------|---------|-------|
+| `enabled` | bool | `true` | Persists finished HLS encodes per `(source, quality, audio_index)`. A second play of the same file at the same quality reuses the segments — no ffmpeg, near-zero CPU, instant playback. Set to `false` to delete segments on session close (original behavior). |
+| `size_gb` | int | `5` | Cache budget in gigabytes. When exceeded the LRU sweeper evicts the least-recently-used cached encodes hourly. Minimum 1 GB (smaller values are clamped up). |
+| `dir` | string | `""` | Custom storage path. Empty defaults to `~/.cache/unarr/hls-cache` (Linux/macOS) or the user cache dir (Windows). |
+
+**What it does.** First play encodes normally (ffmpeg writes segments).
+On session close, if every segment is on disk and ffmpeg exited cleanly,
+the directory is sealed with a `.complete` marker and kept. Next time the
+same source + quality combo is requested, the daemon serves segments
+straight from disk — no transcode, no warm-up, no CPU cost.
+
+**Why per (source, quality, audio).** Renaming the file or switching
+quality invalidates the entry: the segments are tied to the exact source
+bytes and the exact ffmpeg parameters. Re-encoding generates a new key.
+
+**Eviction.** A background goroutine wakes every hour. If total cache size
+exceeds `size_gb`, it deletes the oldest entries (by mtime) until under
+budget. Active sessions are pinned — they never get evicted mid-play.
+
+**Disable.** Either edit the TOML to set `enabled = false`, or remove the
+cache directory manually (it'll be recreated as needed). Disabling does
+not delete existing cached segments — drop `dir` (or `~/.cache/unarr/hls-cache`)
+to reclaim the space.
+
 #### `[downloads.vpn]`

 | Key | Type | Default | Notes |
--- a/internal/agent/daemon.go
+++ b/internal/agent/daemon.go
@ -28,6 +28,7 @@ type DaemonConfig struct {
 	ScanPaths          []string // configured scan paths for file deletion validation
 	HWAccel            string   // detected encoder backend ("nvenc"/"qsv"/"vaapi"/"videotoolbox"/"none")
 	MaxTranscodeHeight int      // resolution cap the agent can transcode comfortably (px)
+	AutoUpgrade        bool     // honor server-flagged upgrades by downloading + restarting (default: true)
 }

 // Daemon manages agent registration and the sync loop.
@ -237,6 +238,10 @@ func (d *Daemon) Run(ctx context.Context) error {
 			return
 		}
 		d.lastNotifiedVersion = version
+		if !d.cfg.AutoUpgrade {
+			log.Printf("[upgrade] new version available: %s — auto_upgrade=false, run `unarr update` to apply", version)
+			return
+		}
 		log.Printf("[upgrade] new version available: %s — applying auto-upgrade", version)
 		go d.applyAutoUpgrade(version)
 	}
--- a/internal/cmd/daemon.go
+++ b/internal/cmd/daemon.go
@ -162,6 +162,7 @@ func runDaemonStart() error {
 		ScanPaths:          library.ResolveScanPaths(cfg.Download.Dir, cfg.Organize.MoviesDir, cfg.Organize.TVShowsDir, cfg.Library.ScanPath),
 		HWAccel:            string(hwAccelPick),
 		MaxTranscodeHeight: maxTranscodeHeight,
+		AutoUpgrade:        cfg.Daemon.AutoUpgradeEnabled(),
 	}

 	// Create HTTP client with mirror failover so a `.com` block-out rolls
@ -299,6 +300,32 @@ func runDaemonStart() error {
 	if err := engine.CleanupHLSOrphanDirs(); err != nil {
 		log.Printf("[hls] orphan tmpdir cleanup: %v", err)
 	}
+
+	// Persistent HLS segment cache — survives across sessions so re-plays
+	// of the same file at the same quality skip ffmpeg entirely. Off when
+	// hls_cache.enabled = false; size cap from hls_cache.size_gb; path from
+	// hls_cache.dir (defaults to ~/.cache/unarr/hls-cache).
+	var hlsCache *engine.HLSCache
+	if cfg.Download.HLSCache.Enabled {
+		cacheDir := cfg.Download.HLSCache.Dir
+		if cacheDir == "" {
+			if base, err := os.UserCacheDir(); err == nil {
+				cacheDir = filepath.Join(base, "unarr", "hls-cache")
+			} else {
+				cacheDir = filepath.Join(os.TempDir(), "unarr-hls-cache")
+			}
+		}
+		c, err := engine.NewHLSCache(cacheDir, cfg.Download.HLSCache.SizeGB)
+		if err != nil {
+			log.Printf("[hls_cache] init failed (%v) — falling back to per-session tmpdirs", err)
+		} else {
+			hlsCache = c
+			hlsCache.StartSweeper(ctx, time.Hour)
+			log.Printf("[hls_cache] enabled: dir=%s budget=%dGB", cacheDir, cfg.Download.HLSCache.SizeGB)
+		}
+	} else {
+		log.Printf("[hls_cache] disabled by config — every play re-encodes from scratch")
+	}
 	if err := streamSrv.Listen(ctx); err != nil {
 		return fmt.Errorf("start stream server: %w", err)
 	}
@ -543,6 +570,7 @@ func runDaemonStart() error {
 			Quality:    sess.Quality,
 			AudioIndex: sess.AudioIndex,
 			Transcode:  tcRuntime,
+			Cache:      hlsCache,
 		}
 		hsess, err := engine.StartHLSSession(hlsCtx, hlsCfg)
 		if err != nil {
--- a/internal/cmd/version.go
+++ b/internal/cmd/version.go
@ -1,4 +1,4 @@
 package cmd

 // Version is the CLI version. Overridden by goreleaser ldflags at release time.
-var Version = "0.9.6"
+var Version = "0.9.7"
--- a/internal/config/config.go
+++ b/internal/config/config.go
@ -52,10 +52,22 @@ type DownloadConfig struct {
 	EnableUPnP       bool            `toml:"enable_upnp"`        // map StreamPort to the WAN via UPnP/NAT-PMP (default: false; opt-in because it exposes the unauthenticated /stream + /hls endpoints to the public internet)
 	CORSExtraOrigins []string        `toml:"cors_extra_origins"` // extra browser origins added on top of the baked-in allowlist (torrentclaw.com, app.torrentclaw.com, localhost:3030)
 	Transcode        TranscodeConfig `toml:"transcode"`
+	HLSCache         HLSCacheConfig  `toml:"hls_cache"`
 	VPN              VPNConfig       `toml:"vpn"`
 	Funnel           FunnelConfig    `toml:"funnel"`
 }

+// HLSCacheConfig controls the persistent HLS segment cache. A completed encode
+// is kept on disk so a second play of the same file at the same quality skips
+// ffmpeg entirely. Old entries are evicted (LRU) once the cache exceeds the
+// size budget. Enabled by default — disable to save disk space at the cost of
+// re-encoding every play.
+type HLSCacheConfig struct {
+	Enabled bool   `toml:"enabled"`  // default: true
+	SizeGB  int    `toml:"size_gb"`  // size budget in gigabytes; default: 5; minimum: 1
+	Dir     string `toml:"dir"`      // override storage path; default: ~/.cache/unarr/hls-cache
+}
+
 // FunnelConfig gates the optional CloudFlare Quick Tunnel that exposes the
 // daemon's HLS server over a public HTTPS hostname (https://<random>.try
 // cloudflare.com). Enabling it lets the web player on torrentclaw.com play
@ -101,8 +113,27 @@ type OrganizeConfig struct {

 type DaemonConfig struct {
 	StatusInterval string `toml:"status_interval"`
+	// AutoUpgrade gates the daemon's response to a server-flagged upgrade
+	// (set via the "Force update" button on the web). When true the daemon
+	// downloads + replaces the binary in-place and exits so the service
+	// supervisor respawns on the new version. When false the daemon only
+	// logs "new version available" and the operator must run `unarr update`
+	// manually. Default: true. Available since unarr 0.9.6.
+	AutoUpgrade *bool `toml:"auto_upgrade"`
 }

+// AutoUpgradeEnabled returns the resolved AutoUpgrade flag — defaults to true
+// when the user has not set it explicitly. Pointer-vs-bool because Go's
+// zero-value bool would collapse "unset" and "false" together.
+func (d DaemonConfig) AutoUpgradeEnabled() bool {
+	if d.AutoUpgrade == nil {
+		return true
+	}
+	return *d.AutoUpgrade
+}
+
+func boolPtr(v bool) *bool { return &v }
+
 type NotificationsConfig struct {
 	Enabled bool `toml:"enabled"`
 }
@ -156,11 +187,24 @@ func Default() Config {
 				// `unarr funnel off` (sets enabled=false in the TOML).
 				Enabled: true,
 			},
+			HLSCache: HLSCacheConfig{
+				// On by default — second play of a recently watched file at the
+				// same quality skips ffmpeg (instant start, near-zero CPU).
+				// Users can opt out (hls_cache.enabled=false) or shrink the
+				// budget (hls_cache.size_gb) when disk is tight.
+				Enabled: true,
+				SizeGB:  5,
+			},
+		},
+		Daemon: DaemonConfig{
+			// Pointer-to-true so Default() round-trips through TOML marshal
+			// as `auto_upgrade = true` instead of an omitted key — keeps the
+			// freshly-written config aligned with what README documents.
+			AutoUpgrade: boolPtr(true),
 		},
 		Organize: OrganizeConfig{
 			Enabled: true,
 		},
-		Daemon: DaemonConfig{},
 		Notifications: NotificationsConfig{
 			Enabled: true,
 		},
--- a/internal/engine/hls.go
+++ b/internal/engine/hls.go
@ -100,6 +100,11 @@ type HLSSessionConfig struct {
 	Quality    string // "2160p"|"1080p"|"720p"|"480p"|"original"|""
 	AudioIndex int    // 0-based ffmpeg audio stream selection (-map 0:a:N). -1 = default.
 	Transcode  TranscodeRuntime
+	// Cache is an optional persistent segment cache keyed by (source, quality,
+	// audio). When set, completed encodes are kept across sessions so re-plays
+	// of the same file at the same quality skip ffmpeg entirely. nil disables
+	// caching (per-session tmpdir, deleted on Close — original behavior).
+	Cache *HLSCache
 }

 // HLSSession owns a tmpdir + ffmpeg subprocess producing HLS fragments.
@ -139,6 +144,19 @@ type HLSSession struct {
 	exitErr  error
 	exited   bool
 	readyCh  chan struct{} // closed + replaced each time readyMax advances
+
+	// Persistent cache state. cache==nil means caching disabled for this session.
+	// fromCache=true means the session is replaying a completed encode and no
+	// ffmpeg subprocess was spawned. writerLockHeld=true means this session
+	// owns the per-key TryAcquireWriter claim — Close must ReleaseWriter.
+	// subsDone closes when the subtitle extractor goroutine returns (or is
+	// nil when the source had no subtitle tracks); MarkComplete waits on it
+	// so a HIT replay never serves partial .vtt files.
+	cache          *HLSCache
+	cacheKey       string
+	fromCache      bool
+	writerLockHeld bool
+	subsDone       chan struct{}
 }

 // hlsSeekAhead is how many segments past the writer's current position the
@ -263,11 +281,77 @@ func StartHLSSession(ctx context.Context, cfg HLSSessionConfig) (*HLSSession, er
 		return nil, errors.New("hls: source has no duration")
 	}

-	tmpDir := filepath.Join(hlsTmpDirRoot(), cfg.SessionID)
+	// Resolve tmpDir + cache placement. Three states:
+	//   1. cache disabled              → per-session tmpdir, deleted on Close.
+	//   2. cache HIT (.complete found) → read from cache dir, no ffmpeg, Pin.
+	//   3. cache MISS, writer-lock OK  → ffmpeg writes to cache dir, Pin + writer-lock.
+	//   4. cache MISS, writer-lock NO  → another session already writing this
+	//                                    key; fall back to private per-session tmpdir
+	//                                    (no caching for this session — second-writer
+	//                                    would corrupt the first one's segments).
+	var (
+		tmpDir         string
+		cacheKey       string
+		fromCache      bool
+		writerLockHeld bool
+	)
+	if cfg.Cache != nil {
+		cacheKey = cfg.Cache.KeyFor(cfg.SourcePath, cfg.Quality, cfg.AudioIndex)
+		// Integrity gate: HasComplete just stats the marker. If init.mp4 or
+		// the last segment vanished (external rm, partial-disk failure), we
+		// can't actually serve a HIT — drop the dir and re-encode.
+		segCountForVerify := int((probe.DurationSec + float64(hlsSegmentDuration) - 1) / float64(hlsSegmentDuration))
+		if segCountForVerify < 1 {
+			segCountForVerify = 1
+		}
+		if cfg.Cache.HasComplete(cacheKey) && !cfg.Cache.VerifyComplete(cacheKey, segCountForVerify) {
+			log.Printf("[hls %s] cache %s sealed but failed integrity check — re-encoding",
+				shortHLSID(cfg.SessionID), cacheKey)
+			_ = cfg.Cache.Invalidate(cacheKey)
+		}
+		if cfg.Cache.HasComplete(cacheKey) {
+			// HIT: read-only replay — many concurrent HITs are fine.
+			tmpDir = cfg.Cache.DirFor(cacheKey)
+			cfg.Cache.Pin(cacheKey)
+			fromCache = true
+			cfg.Cache.RecordHit()
+			_ = cfg.Cache.Touch(cacheKey)
+		} else if cfg.Cache.TryAcquireWriter(cacheKey) {
+			tmpDir = cfg.Cache.DirFor(cacheKey)
+			cfg.Cache.Pin(cacheKey)
+			writerLockHeld = true
+			cfg.Cache.RecordMiss()
+		} else {
+			// Another session is writing this key — fall back to private
+			// dir so we don't trample its segments.
+			log.Printf("[hls %s] cache key %s busy, falling back to per-session tmpdir",
+				shortHLSID(cfg.SessionID), cacheKey)
+			tmpDir = filepath.Join(hlsTmpDirRoot(), cfg.SessionID)
+			cacheKey = "" // disable caching for this session
+			cfg.Cache.RecordMiss()
+		}
+	} else {
+		tmpDir = filepath.Join(hlsTmpDirRoot(), cfg.SessionID)
+	}
+
+	cleanupOnError := func() {
+		if cfg.Cache != nil && cacheKey != "" {
+			cfg.Cache.Unpin(cacheKey)
+			if writerLockHeld {
+				cfg.Cache.ReleaseWriter(cacheKey)
+				_ = cfg.Cache.Invalidate(cacheKey)
+			}
+		} else {
+			_ = os.RemoveAll(tmpDir)
+		}
+	}
+
 	if err := os.MkdirAll(filepath.Join(tmpDir, "video"), 0o755); err != nil {
+		cleanupOnError()
 		return nil, fmt.Errorf("hls: mkdir video: %w", err)
 	}
 	if err := os.MkdirAll(filepath.Join(tmpDir, "subs"), 0o755); err != nil {
+		cleanupOnError()
 		return nil, fmt.Errorf("hls: mkdir subs: %w", err)
 	}

@ -285,10 +369,30 @@ func StartHLSSession(ctx context.Context, cfg HLSSessionConfig) (*HLSSession, er
 		startedAt:    time.Now(),
 		lastTouch:    time.Now(),
 		readyCh:      make(chan struct{}),
+		cache:          cfg.Cache,
+		cacheKey:       cacheKey,
+		fromCache:      fromCache,
+		writerLockHeld: writerLockHeld,
 	}
 	s.manifestVideo = renderVideoPlaylist(probe.DurationSec, segCount)
 	s.manifestRoot = renderMasterPlaylist(probe, cfg.Quality)

+	// Cache HIT: every segment + init.mp4 is already on disk. Skip ffmpeg
+	// entirely and mark readyMax so handlers don't wait. Background subtitle
+	// extraction is also unnecessary — subs were extracted on the original run.
+	if fromCache {
+		s.readyMu.Lock()
+		s.readyMax = segCount - 1
+		s.exited = true
+		close(s.readyCh)
+		s.readyCh = nil
+		s.readyMu.Unlock()
+		log.Printf("[hls %s] cache HIT %s: %s, %.1fs, %d segs (quality=%s)",
+			shortHLSID(cfg.SessionID), cacheKey, filepath.Base(cfg.SourcePath),
+			probe.DurationSec, segCount, coalesce(cfg.Quality, "auto"))
+		return s, nil
+	}
+
 	// Spawn ffmpeg under a dedicated context so Close() can kill it without
 	// touching the parent ctx.
 	ffCtx, cancel := context.WithCancel(context.Background())
@ -298,7 +402,7 @@ func StartHLSSession(ctx context.Context, cfg HLSSessionConfig) (*HLSSession, er
 	cmd.Stderr = &hlsStderrCapture{owner: s}
 	if err := cmd.Start(); err != nil {
 		cancel()
-		_ = os.RemoveAll(tmpDir)
+		cleanupOnError()
 		return nil, fmt.Errorf("hls: start ffmpeg: %w", err)
 	}
 	s.cmd = cmd
@ -307,12 +411,20 @@ func StartHLSSession(ctx context.Context, cfg HLSSessionConfig) (*HLSSession, er
 	go s.pollSegments(ffCtx)

 	if len(probe.SubtitleTracks) > 0 {
-		go s.extractSubtitles(ffCtx)
+		s.subsDone = make(chan struct{})
+		go func() {
+			defer close(s.subsDone)
+			s.extractSubtitles(ffCtx)
+		}()
 	}

-	log.Printf("[hls %s] started: %s, %.1fs, %d segs (quality=%s)",
+	cachedNote := ""
+	if cfg.Cache != nil {
+		cachedNote = fmt.Sprintf(" (cache-miss %s)", cacheKey)
+	}
+	log.Printf("[hls %s] started: %s, %.1fs, %d segs (quality=%s)%s",
 		shortHLSID(cfg.SessionID), filepath.Base(cfg.SourcePath),
-		probe.DurationSec, segCount, coalesce(cfg.Quality, "auto"))
+		probe.DurationSec, segCount, coalesce(cfg.Quality, "auto"), cachedNote)
 	return s, nil
 }

@ -385,8 +497,15 @@ func (s *HLSSession) Touch() {
 	s.mu.Unlock()
 }

-// Close stops ffmpeg, deletes the tmpdir, and prevents further requests from
-// blocking on segment readiness. Idempotent.
+// Close stops ffmpeg and prevents further requests from blocking on segment
+// readiness. Idempotent.
+//
+// Disk lifecycle:
+//   - cache disabled → delete tmpDir (original behavior).
+//   - cache enabled + this session was a HIT → keep dir, just unpin.
+//   - cache enabled + this was a write session → if ffmpeg exited cleanly and
+//     every segment is on disk, persist with .complete and keep dir. Otherwise
+//     drop the dir so a half-written cache doesn't survive into the next play.
 func (s *HLSSession) Close() error {
 	s.mu.Lock()
 	if s.closed {
@ -407,7 +526,47 @@ func (s *HLSSession) Close() error {
 		s.readyCh = nil
 	}
 	s.exited = true
+	exitErr := s.exitErr
 	s.readyMu.Unlock()
+
+	if s.cache != nil && s.cacheKey != "" {
+		defer s.cache.Unpin(s.cacheKey)
+		if s.writerLockHeld {
+			defer s.cache.ReleaseWriter(s.cacheKey)
+		}
+		if s.fromCache {
+			log.Printf("[hls %s] closed (cache reuse)", shortHLSID(s.cfg.SessionID))
+			return nil
+		}
+		// Wait briefly for the subtitle extractor to finish so a cached
+		// replay never serves half-written .vtt files. Bounded so a stuck
+		// extractor can't block Close indefinitely; on timeout we treat
+		// the cache as incomplete and drop it.
+		subsOK := true
+		if s.subsDone != nil {
+			select {
+			case <-s.subsDone:
+			case <-time.After(15 * time.Second):
+				log.Printf("[hls %s] subtitle extractor timeout — not caching", shortHLSID(s.cfg.SessionID))
+				subsOK = false
+			}
+		}
+		if subsOK && exitErr == nil && s.allSegmentsPresent() {
+			if err := s.cache.MarkComplete(s.cacheKey); err == nil {
+				log.Printf("[hls %s] cache persisted %s", shortHLSID(s.cfg.SessionID), s.cacheKey)
+				return nil
+			} else {
+				log.Printf("[hls %s] cache persist failed: %v", shortHLSID(s.cfg.SessionID), err)
+			}
+		}
+		// Partial / failed → drop so we re-encode next time.
+		if err := s.cache.Invalidate(s.cacheKey); err != nil {
+			log.Printf("[hls %s] cache invalidate failed: %v", shortHLSID(s.cfg.SessionID), err)
+		}
+		log.Printf("[hls %s] closed (cache discarded)", shortHLSID(s.cfg.SessionID))
+		return nil
+	}
+
 	if tmpDir != "" {
 		_ = os.RemoveAll(tmpDir)
 	}
@ -415,6 +574,31 @@ func (s *HLSSession) Close() error {
 	return nil
 }

+// allSegmentsPresent reports whether every expected segment (and init.mp4) is
+// on disk AND validated by the segment poller. Used to decide whether a
+// finished session is cacheable. We trust readyMax (advanced by pollSegments
+// only after the next segment exists, proving the predecessor is fully closed)
+// over a naive Size>0 stat that could accept truncated mid-write files.
+func (s *HLSSession) allSegmentsPresent() bool {
+	if fi, err := os.Stat(filepath.Join(s.tmpDir, "video", "init.mp4")); err != nil || fi.Size() == 0 {
+		return false
+	}
+	s.readyMu.Lock()
+	readyMax := s.readyMax
+	s.readyMu.Unlock()
+	if readyMax < s.segmentCount-1 {
+		return false
+	}
+	for i := 0; i < s.segmentCount; i++ {
+		path := filepath.Join(s.tmpDir, "video", fmt.Sprintf("seg-%d.m4s", i))
+		fi, err := os.Stat(path)
+		if err != nil || fi.Size() == 0 {
+			return false
+		}
+	}
+	return true
+}
+
 // waitFFmpeg reaps the ffmpeg process and records its exit error for handlers.
 //
 // Auto-restart supervisor: if ffmpeg crashes (non-graceful exit) and the
--- a/internal/engine/hls_cache.go
+++ b/internal/engine/hls_cache.go
@ -0,0 +1,410 @@
+package engine
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"log"
+	"os"
+	"path/filepath"
+	"sort"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// HLSCache persists transcoded HLS segments per (source, quality, audio) so a
+// second play of the same file at the same quality skips ffmpeg entirely.
+//
+// Layout on disk:
+//
+//	{root}/{key}/init.mp4
+//	{root}/{key}/seg-0.m4s
+//	{root}/{key}/seg-N.m4s
+//	{root}/{key}/.complete
+//
+// Atomicity: the .complete marker is written only when ffmpeg exits 0 AND all
+// segments are on disk. A dir without .complete is treated as a partial run —
+// next session can reuse the segments already present, ffmpeg fills the gaps.
+//
+// Concurrency: Pin/Unpin increments a ref counter per key so the LRU sweeper
+// never evicts a directory that an active session is reading from.
+type HLSCache struct {
+	root     string
+	maxBytes int64
+
+	mu      sync.Mutex
+	refs    map[string]int
+	writers map[string]bool // exclusive ffmpeg writer per key; nil entries are absent
+
+	// Counters surfaced via Stats() — useful for /api/internal/agent/cache-stats
+	// and for the sweeper's daily log line. atomic so RecordHit/RecordMiss are
+	// safe to call from any goroutine without taking the cache mutex.
+	hits   atomic.Uint64
+	misses atomic.Uint64
+}
+
+const (
+	hlsCacheCompleteMarker = ".complete"
+	// hlsCacheMinBudgetGB clamps absurd / zero / negative SizeGB values to
+	// a sane floor. NOT a guarantee that any single encode fits — a long
+	// 4K HEVC re-encode can exceed it. Operators should set size_gb based
+	// on their actual workload.
+	hlsCacheMinBudgetGB = 1
+	// hlsCacheStartupOrphanAge: directories without .complete older than
+	// this are removed on cache startup. Long enough that a daemon crash
+	// during an in-progress encode (which legitimately leaves a partial
+	// dir) doesn't get nuked too aggressively if the daemon restarts fast.
+	hlsCacheStartupOrphanAge = 10 * time.Minute
+)
+
+// NewHLSCache creates the cache rooted at the given dir with a size budget in
+// gigabytes. A budget < hlsCacheMinBudgetGB is clamped up so a single play
+// doesn't get instantly evicted mid-stream.
+func NewHLSCache(root string, sizeGB int) (*HLSCache, error) {
+	if root == "" {
+		return nil, errors.New("hls_cache: empty root")
+	}
+	if sizeGB < hlsCacheMinBudgetGB {
+		sizeGB = hlsCacheMinBudgetGB
+	}
+	if err := os.MkdirAll(root, 0o755); err != nil {
+		return nil, fmt.Errorf("hls_cache: mkdir root: %w", err)
+	}
+	c := &HLSCache{
+		root:     root,
+		maxBytes: int64(sizeGB) * 1024 * 1024 * 1024,
+		refs:     make(map[string]int),
+		writers:  make(map[string]bool),
+	}
+	// Reap dirs left over from a crashed encode. A dir without .complete that
+	// hasn't been touched recently was almost certainly orphaned by an
+	// ungraceful daemon exit — keeping it just feeds the unbounded growth
+	// pattern the hourly LRU is too slow to contain.
+	if removed, err := c.cleanStartupOrphans(); err != nil {
+		log.Printf("[hls_cache] startup orphan cleanup: %v", err)
+	} else if removed > 0 {
+		log.Printf("[hls_cache] startup: removed %d orphan dir(s) without .complete", removed)
+	}
+	return c, nil
+}
+
+// cleanStartupOrphans removes cache subdirectories that lack a .complete
+// marker AND haven't been modified within hlsCacheStartupOrphanAge. Called
+// once at construction. Safe at startup because no sessions are active yet,
+// so Pin can't race with us.
+func (c *HLSCache) cleanStartupOrphans() (int, error) {
+	entries, err := os.ReadDir(c.root)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return 0, nil
+		}
+		return 0, err
+	}
+	cutoff := time.Now().Add(-hlsCacheStartupOrphanAge)
+	removed := 0
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		dir := filepath.Join(c.root, e.Name())
+		if _, err := os.Stat(filepath.Join(dir, hlsCacheCompleteMarker)); err == nil {
+			continue // sealed, keep
+		}
+		info, err := e.Info()
+		if err != nil {
+			continue
+		}
+		if info.ModTime().After(cutoff) {
+			continue // too recent — might be a daemon that just restarted mid-encode
+		}
+		if err := os.RemoveAll(dir); err == nil {
+			removed++
+		}
+	}
+	return removed, nil
+}
+
+// TryAcquireWriter attempts to claim exclusive ffmpeg-write access to a key.
+// Returns true on success — the caller is then responsible for ReleaseWriter
+// when ffmpeg exits / fails. Returns false if another session is already
+// writing this key, in which case the caller must fall back to a private
+// per-session tmpdir (no caching for that session).
+func (c *HLSCache) TryAcquireWriter(key string) bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if c.writers[key] {
+		return false
+	}
+	c.writers[key] = true
+	return true
+}
+
+// ReleaseWriter releases the writer claim acquired via TryAcquireWriter.
+// Idempotent on unknown keys.
+func (c *HLSCache) ReleaseWriter(key string) {
+	c.mu.Lock()
+	delete(c.writers, key)
+	c.mu.Unlock()
+}
+
+// KeyFor derives a stable cache key for (source, quality, audioIndex). Using
+// the absolute source path means renaming a file invalidates the cache, which
+// is correct — segment content is tied to the encoded source.
+func (c *HLSCache) KeyFor(sourcePath, quality string, audioIndex int) string {
+	abs, err := filepath.Abs(sourcePath)
+	if err != nil {
+		abs = sourcePath
+	}
+	h := sha256.Sum256([]byte(fmt.Sprintf("%s|%s|%d", abs, quality, audioIndex)))
+	return hex.EncodeToString(h[:8]) // 16 hex chars — collision-safe enough for per-host cache
+}
+
+// DirFor returns the on-disk directory for a cache key. Caller is responsible
+// for creating it.
+func (c *HLSCache) DirFor(key string) string {
+	return filepath.Join(c.root, key)
+}
+
+// HasComplete returns true when the .complete marker is present, meaning the
+// directory holds a full set of segments from a successful encode.
+func (c *HLSCache) HasComplete(key string) bool {
+	if _, err := os.Stat(filepath.Join(c.DirFor(key), hlsCacheCompleteMarker)); err == nil {
+		return true
+	}
+	return false
+}
+
+// MarkComplete writes the .complete marker. Call only after verifying ffmpeg
+// exited cleanly AND every expected segment is on disk. The dir must already
+// exist — StartHLSSession created it on the writer path.
+func (c *HLSCache) MarkComplete(key string) error {
+	return os.WriteFile(filepath.Join(c.DirFor(key), hlsCacheCompleteMarker), nil, 0o644)
+}
+
+// RecordHit increments the hit counter; called by StartHLSSession on a
+// cache-HIT path.
+func (c *HLSCache) RecordHit() { c.hits.Add(1) }
+
+// RecordMiss increments the miss counter; called when a session has to
+// encode from scratch (or fails an integrity check on a stale HIT).
+func (c *HLSCache) RecordMiss() { c.misses.Add(1) }
+
+// CacheStats is a snapshot of the cache's runtime counters + on-disk size.
+// The size fields are best-effort (computed via dirSize) so callers paying
+// for them should cache the result, not poll in a hot loop.
+type CacheStats struct {
+	Hits       uint64
+	Misses     uint64
+	EntryCount int
+	TotalBytes int64
+}
+
+// Stats returns a snapshot of the cache counters and size. Walks the root
+// to total disk usage — O(N segments). Call at most every few minutes.
+func (c *HLSCache) Stats() CacheStats {
+	s := CacheStats{
+		Hits:   c.hits.Load(),
+		Misses: c.misses.Load(),
+	}
+	entries, err := os.ReadDir(c.root)
+	if err != nil {
+		return s
+	}
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		size, err := dirSize(filepath.Join(c.root, e.Name()))
+		if err != nil {
+			continue
+		}
+		s.EntryCount++
+		s.TotalBytes += size
+	}
+	return s
+}
+
+// hitRatePercent returns the current hit/(hit+miss) percentage rounded to
+// the nearest int; 0 when no calls have been recorded.
+func (c *HLSCache) hitRatePercent() int {
+	h := c.hits.Load()
+	m := c.misses.Load()
+	total := h + m
+	if total == 0 {
+		return 0
+	}
+	return int((h*100 + total/2) / total)
+}
+
+// VerifyComplete checks that the .complete marker is present AND the
+// essential files (init.mp4 + last segment) exist with non-zero size. A
+// dir that passes HasComplete but fails VerifyComplete is treated as
+// corrupted — typically external `rm` or a partial-disk-failure scenario.
+// When it returns false, callers should Invalidate and re-encode.
+func (c *HLSCache) VerifyComplete(key string, segmentCount int) bool {
+	if !c.HasComplete(key) {
+		return false
+	}
+	dir := c.DirFor(key)
+	if fi, err := os.Stat(filepath.Join(dir, "video", "init.mp4")); err != nil || fi.Size() == 0 {
+		return false
+	}
+	if segmentCount > 0 {
+		lastSeg := filepath.Join(dir, "video", fmt.Sprintf("seg-%d.m4s", segmentCount-1))
+		if fi, err := os.Stat(lastSeg); err != nil || fi.Size() == 0 {
+			return false
+		}
+	}
+	return true
+}
+
+// Pin increments the ref counter for a key. The sweeper checks this before
+// evicting, so a pinned dir is safe even if its mtime is old.
+func (c *HLSCache) Pin(key string) {
+	c.mu.Lock()
+	c.refs[key]++
+	c.mu.Unlock()
+}
+
+// Unpin decrements; safe to call on unknown keys (no-op).
+func (c *HLSCache) Unpin(key string) {
+	c.mu.Lock()
+	if c.refs[key] > 0 {
+		c.refs[key]--
+		if c.refs[key] == 0 {
+			delete(c.refs, key)
+		}
+	}
+	c.mu.Unlock()
+}
+
+func (c *HLSCache) isPinned(key string) bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.refs[key] > 0
+}
+
+// Touch updates the directory mtime so LRU picks fresher entries as recently
+// used. Called when a session starts reading from a cached dir.
+func (c *HLSCache) Touch(key string) error {
+	dir := c.DirFor(key)
+	now := time.Now()
+	return os.Chtimes(dir, now, now)
+}
+
+// Sweep enforces the size budget by deleting the least-recently-used cache
+// dirs (ignoring pinned ones) until the total size is at or below maxBytes.
+// Returns the number of bytes freed.
+func (c *HLSCache) Sweep() (int64, error) {
+	entries, err := os.ReadDir(c.root)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return 0, nil
+		}
+		return 0, fmt.Errorf("hls_cache: read root: %w", err)
+	}
+
+	type item struct {
+		key   string
+		path  string
+		size  int64
+		mtime time.Time
+	}
+	items := make([]item, 0, len(entries))
+	var total, pinned int64
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		info, err := e.Info()
+		if err != nil {
+			continue
+		}
+		key := e.Name()
+		path := filepath.Join(c.root, key)
+		size, err := dirSize(path)
+		if err != nil {
+			continue
+		}
+		items = append(items, item{key: key, path: path, size: size, mtime: info.ModTime()})
+		total += size
+		if c.isPinned(key) {
+			pinned += size
+		}
+	}
+
+	if total <= c.maxBytes {
+		return 0, nil
+	}
+	if pinned >= c.maxBytes {
+		// Every pinned byte already exceeds the budget — even evicting
+		// every unpinned dir won't bring us under. Warn loudly so the
+		// operator knows to bump size_gb (or kill the long-running session).
+		log.Printf("[hls_cache] warn: pinned bytes (%.1f MB) exceed budget (%.1f MB) — cannot enforce limit until sessions release",
+			float64(pinned)/(1024*1024), float64(c.maxBytes)/(1024*1024))
+		return 0, nil
+	}
+
+	// Oldest first.
+	sort.Slice(items, func(i, j int) bool {
+		return items[i].mtime.Before(items[j].mtime)
+	})
+
+	var freed int64
+	for _, it := range items {
+		if total-freed <= c.maxBytes {
+			break
+		}
+		if c.isPinned(it.key) {
+			continue
+		}
+		if err := os.RemoveAll(it.path); err != nil {
+			log.Printf("[hls_cache] evict %s failed: %v", it.key, err)
+			continue
+		}
+		log.Printf("[hls_cache] evicted %s (%.1f MB, age %s)",
+			it.key, float64(it.size)/(1024*1024), time.Since(it.mtime).Round(time.Second))
+		freed += it.size
+	}
+	return freed, nil
+}
+
+// StartSweeper kicks off the LRU sweeper goroutine. Cancels on ctx done.
+// In addition to enforcing the size budget, logs a daily summary of hit-rate
+// + disk usage so operators can see the cache's value at a glance.
+func (c *HLSCache) StartSweeper(ctx context.Context, interval time.Duration) {
+	if interval <= 0 {
+		interval = time.Hour
+	}
+	go func() {
+		t := time.NewTicker(interval)
+		defer t.Stop()
+		statsTick := time.NewTicker(24 * time.Hour)
+		defer statsTick.Stop()
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case <-t.C:
+				if _, err := c.Sweep(); err != nil {
+					log.Printf("[hls_cache] sweep error: %v", err)
+				}
+			case <-statsTick.C:
+				s := c.Stats()
+				log.Printf("[hls_cache] day-stats: hits=%d misses=%d ratio=%d%% entries=%d size=%.1fMB",
+					s.Hits, s.Misses, c.hitRatePercent(), s.EntryCount,
+					float64(s.TotalBytes)/(1024*1024))
+			}
+		}
+	}()
+}
+
+// Invalidate removes a cache entry — used when ffmpeg fails to encode the
+// source so we don't reuse a half-written dir next time.
+func (c *HLSCache) Invalidate(key string) error {
+	return os.RemoveAll(c.DirFor(key))
+}
+
--- a/internal/engine/hls_cache_smoke_test.go
+++ b/internal/engine/hls_cache_smoke_test.go
@ -0,0 +1,134 @@
+//go:build smoke
+
+package engine
+
+import (
+	"context"
+	"os/exec"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+// TestHLSCacheSmoke exercises the end-to-end cache flow against real ffmpeg:
+//   - First session encodes a 5s test pattern; expect MISS, ffmpeg runs,
+//     .complete written, MarkComplete logs.
+//   - Second session for identical (source, quality, audio); expect HIT,
+//     no ffmpeg, instant Start.
+//
+// Build tag `smoke` keeps it out of the default `go test ./...` run because
+// it depends on a working ffmpeg/ffprobe and takes ~5–10 s.
+//
+//	go test -tags=smoke -run TestHLSCacheSmoke -v ./internal/engine/
+func TestHLSCacheSmoke(t *testing.T) {
+	ffmpeg, err := exec.LookPath("ffmpeg")
+	if err != nil {
+		t.Skipf("ffmpeg not on PATH: %v", err)
+	}
+	ffprobe, err := exec.LookPath("ffprobe")
+	if err != nil {
+		t.Skipf("ffprobe not on PATH: %v", err)
+	}
+
+	tmp := t.TempDir()
+	source := filepath.Join(tmp, "source.mp4")
+	t.Logf("generating 5 s test pattern → %s", source)
+	if out, err := exec.Command(ffmpeg,
+		"-y", "-loglevel", "error",
+		"-f", "lavfi", "-i", "testsrc=duration=5:size=640x480:rate=30",
+		"-f", "lavfi", "-i", "sine=frequency=1000:duration=5",
+		"-c:v", "libx264", "-preset", "ultrafast", "-pix_fmt", "yuv420p",
+		"-c:a", "aac",
+		source,
+	).CombinedOutput(); err != nil {
+		t.Fatalf("ffmpeg generate: %v\n%s", err, out)
+	}
+
+	cacheRoot := filepath.Join(tmp, "cache")
+	cache, err := NewHLSCache(cacheRoot, 1)
+	if err != nil {
+		t.Fatalf("NewHLSCache: %v", err)
+	}
+
+	cfg := HLSSessionConfig{
+		SessionID:  "smoke1",
+		SourcePath: source,
+		FileName:   "source.mp4",
+		Quality:    "720p",
+		AudioIndex: 0,
+		Transcode: TranscodeRuntime{
+			FFmpegPath:  ffmpeg,
+			FFprobePath: ffprobe,
+			Preset:      "ultrafast",
+		},
+		Cache: cache,
+	}
+
+	// First run — expect MISS, ffmpeg runs.
+	t.Log("session 1: expect MISS")
+	t0 := time.Now()
+	s1, err := StartHLSSession(context.Background(), cfg)
+	if err != nil {
+		t.Fatalf("StartHLSSession #1: %v", err)
+	}
+	if s1.fromCache {
+		t.Fatal("session 1 reported cache HIT on a fresh cache")
+	}
+
+	// Wait for all segments to land. 5 s source @ 4 s segments → 2 segments.
+	deadline := time.Now().Add(60 * time.Second)
+	for {
+		s1.readyMu.Lock()
+		ready := s1.readyMax
+		exited := s1.exited
+		s1.readyMu.Unlock()
+		if ready >= s1.segmentCount-1 && exited {
+			break
+		}
+		if time.Now().After(deadline) {
+			_ = s1.Close()
+			t.Fatalf("session 1 didn't finish in 60 s (readyMax=%d/%d, exited=%v)",
+				ready, s1.segmentCount-1, exited)
+		}
+		time.Sleep(100 * time.Millisecond)
+	}
+	if err := s1.Close(); err != nil {
+		t.Fatalf("Close #1: %v", err)
+	}
+	encodeDur := time.Since(t0)
+	t.Logf("session 1: MISS completed in %s", encodeDur.Round(time.Millisecond))
+
+	key := cache.KeyFor(source, "720p", 0)
+	if !cache.HasComplete(key) {
+		t.Fatalf("cache.HasComplete(%s) is false after successful encode", key)
+	}
+
+	// Second run — expect HIT, no ffmpeg.
+	t.Log("session 2: expect HIT")
+	cfg.SessionID = "smoke2"
+	t1 := time.Now()
+	s2, err := StartHLSSession(context.Background(), cfg)
+	if err != nil {
+		t.Fatalf("StartHLSSession #2: %v", err)
+	}
+	if !s2.fromCache {
+		t.Fatal("session 2 should have reported cache HIT")
+	}
+	if s2.cmd != nil {
+		t.Fatal("session 2 should not have spawned ffmpeg (s.cmd != nil)")
+	}
+	hitDur := time.Since(t1)
+	t.Logf("session 2: HIT in %s (%.1f×  faster than MISS)",
+		hitDur.Round(time.Millisecond), float64(encodeDur)/float64(hitDur))
+	if hitDur > 500*time.Millisecond {
+		t.Errorf("HIT path too slow: %s — expected <500 ms", hitDur)
+	}
+	if err := s2.Close(); err != nil {
+		t.Fatalf("Close #2: %v", err)
+	}
+
+	// After the HIT session closes, the cache dir + .complete must still exist.
+	if !cache.HasComplete(key) {
+		t.Fatal(".complete disappeared after HIT session closed")
+	}
+}
--- a/internal/engine/hls_cache_test.go
+++ b/internal/engine/hls_cache_test.go
@ -0,0 +1,361 @@
+package engine
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"sync"
+	"testing"
+	"time"
+)
+
+func newTestCache(t *testing.T, sizeGB int) *HLSCache {
+	t.Helper()
+	root := t.TempDir()
+	c, err := NewHLSCache(root, sizeGB)
+	if err != nil {
+		t.Fatalf("NewHLSCache: %v", err)
+	}
+	return c
+}
+
+func TestKeyForStable(t *testing.T) {
+	c := newTestCache(t, 1)
+	k1 := c.KeyFor("/a/b/movie.mkv", "1080p", 0)
+	k2 := c.KeyFor("/a/b/movie.mkv", "1080p", 0)
+	if k1 != k2 {
+		t.Fatalf("expected stable keys, got %q vs %q", k1, k2)
+	}
+	if c.KeyFor("/a/b/movie.mkv", "720p", 0) == k1 {
+		t.Fatal("quality should change key")
+	}
+	if c.KeyFor("/a/b/movie.mkv", "1080p", 1) == k1 {
+		t.Fatal("audio index should change key")
+	}
+	if c.KeyFor("/x/y/other.mkv", "1080p", 0) == k1 {
+		t.Fatal("path should change key")
+	}
+}
+
+func TestMarkCompleteAndHas(t *testing.T) {
+	c := newTestCache(t, 1)
+	key := "abc123"
+	if c.HasComplete(key) {
+		t.Fatal("fresh cache should not report complete")
+	}
+	// Production callers create the dir during StartHLSSession; MarkComplete
+	// trusts that invariant and fails if the dir was wiped meanwhile.
+	if err := os.MkdirAll(c.DirFor(key), 0o755); err != nil {
+		t.Fatalf("mkdir: %v", err)
+	}
+	if err := c.MarkComplete(key); err != nil {
+		t.Fatalf("MarkComplete: %v", err)
+	}
+	if !c.HasComplete(key) {
+		t.Fatal("after MarkComplete, HasComplete must be true")
+	}
+}
+
+func TestMarkCompleteFailsWithoutDir(t *testing.T) {
+	c := newTestCache(t, 1)
+	if err := c.MarkComplete("never-created"); err == nil {
+		t.Fatal("MarkComplete should error when dir doesn't exist")
+	}
+}
+
+func TestPinPreventsEviction(t *testing.T) {
+	c := newTestCache(t, 1) // 1 GB budget, but min clamp keeps it usable
+	c.maxBytes = 1024       // squeeze budget for the test
+
+	// Write two entries past the budget.
+	for i, key := range []string{"old", "new"} {
+		dir := c.DirFor(key)
+		if err := os.MkdirAll(dir, 0o755); err != nil {
+			t.Fatalf("mkdir %s: %v", dir, err)
+		}
+		path := filepath.Join(dir, "seg.bin")
+		if err := os.WriteFile(path, make([]byte, 800), 0o644); err != nil {
+			t.Fatalf("write %s: %v", path, err)
+		}
+		now := time.Now().Add(time.Duration(i) * time.Hour) // "old" mtime < "new"
+		_ = os.Chtimes(dir, now, now)
+	}
+
+	c.Pin("old") // protect the older one
+	freed, err := c.Sweep()
+	if err != nil {
+		t.Fatalf("Sweep: %v", err)
+	}
+	if freed == 0 {
+		t.Fatal("expected some eviction")
+	}
+	if _, err := os.Stat(c.DirFor("old")); err != nil {
+		t.Fatal("pinned 'old' was evicted")
+	}
+	if _, err := os.Stat(c.DirFor("new")); err == nil {
+		t.Fatal("'new' should have been evicted to make room")
+	}
+}
+
+func TestSweepNoOpUnderBudget(t *testing.T) {
+	c := newTestCache(t, 1)
+	dir := c.DirFor("small")
+	_ = os.MkdirAll(dir, 0o755)
+	_ = os.WriteFile(filepath.Join(dir, "x"), []byte("tiny"), 0o644)
+	freed, err := c.Sweep()
+	if err != nil {
+		t.Fatalf("Sweep: %v", err)
+	}
+	if freed != 0 {
+		t.Fatalf("expected 0 freed under budget, got %d", freed)
+	}
+	if _, err := os.Stat(dir); err != nil {
+		t.Fatal("under-budget entry was wrongly evicted")
+	}
+}
+
+func TestSweepEmptyRoot(t *testing.T) {
+	c := newTestCache(t, 1)
+	freed, err := c.Sweep()
+	if err != nil {
+		t.Fatalf("Sweep empty: %v", err)
+	}
+	if freed != 0 {
+		t.Fatalf("freed=%d, want 0", freed)
+	}
+}
+
+func TestInvalidateRemovesDir(t *testing.T) {
+	c := newTestCache(t, 1)
+	key := "drop"
+	dir := c.DirFor(key)
+	_ = os.MkdirAll(dir, 0o755)
+	_ = os.WriteFile(filepath.Join(dir, "x"), []byte("y"), 0o644)
+	if err := c.Invalidate(key); err != nil {
+		t.Fatalf("Invalidate: %v", err)
+	}
+	if _, err := os.Stat(dir); err == nil {
+		t.Fatal("dir still present after Invalidate")
+	}
+}
+
+func TestTouchUpdatesMtime(t *testing.T) {
+	c := newTestCache(t, 1)
+	key := "touch"
+	dir := c.DirFor(key)
+	_ = os.MkdirAll(dir, 0o755)
+	old := time.Now().Add(-2 * time.Hour)
+	_ = os.Chtimes(dir, old, old)
+
+	if err := c.Touch(key); err != nil {
+		t.Fatalf("Touch: %v", err)
+	}
+	info, err := os.Stat(dir)
+	if err != nil {
+		t.Fatalf("stat: %v", err)
+	}
+	if !info.ModTime().After(old.Add(time.Minute)) {
+		t.Fatalf("mtime not refreshed: %v", info.ModTime())
+	}
+}
+
+func TestPinUnpinSymmetry(t *testing.T) {
+	c := newTestCache(t, 1)
+	c.Pin("k")
+	c.Pin("k")
+	if !c.isPinned("k") {
+		t.Fatal("Pin twice should leave pinned")
+	}
+	c.Unpin("k")
+	if !c.isPinned("k") {
+		t.Fatal("Unpin once should keep pinned (refs=1)")
+	}
+	c.Unpin("k")
+	if c.isPinned("k") {
+		t.Fatal("Unpin twice should drop pin")
+	}
+	c.Unpin("k") // safe no-op
+}
+
+func TestConcurrentPinUnpin(t *testing.T) {
+	c := newTestCache(t, 1)
+	var wg sync.WaitGroup
+	for i := 0; i < 100; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			c.Pin("race")
+			time.Sleep(time.Microsecond)
+			c.Unpin("race")
+		}()
+	}
+	wg.Wait()
+	if c.isPinned("race") {
+		t.Fatal("refs leaked")
+	}
+}
+
+func TestSweeperLoopExits(t *testing.T) {
+	c := newTestCache(t, 1)
+	ctx, cancel := context.WithCancel(context.Background())
+	c.StartSweeper(ctx, 10*time.Millisecond)
+	time.Sleep(30 * time.Millisecond)
+	cancel()
+	// If StartSweeper doesn't exit on cancel the test would leak a goroutine;
+	// the leak detector in the test runner will surface it.
+	time.Sleep(20 * time.Millisecond)
+}
+
+func TestMinBudgetClamp(t *testing.T) {
+	root := t.TempDir()
+	c, err := NewHLSCache(root, 0) // below floor
+	if err != nil {
+		t.Fatalf("NewHLSCache: %v", err)
+	}
+	if c.maxBytes != int64(hlsCacheMinBudgetGB)*1024*1024*1024 {
+		t.Fatalf("budget not clamped to min: got %d", c.maxBytes)
+	}
+}
+
+func TestTryAcquireWriterExclusive(t *testing.T) {
+	c := newTestCache(t, 1)
+	if !c.TryAcquireWriter("k") {
+		t.Fatal("first acquire should succeed")
+	}
+	if c.TryAcquireWriter("k") {
+		t.Fatal("second acquire for same key must fail")
+	}
+	if !c.TryAcquireWriter("other") {
+		t.Fatal("different key should not conflict")
+	}
+	c.ReleaseWriter("k")
+	if !c.TryAcquireWriter("k") {
+		t.Fatal("acquire after release should succeed")
+	}
+	c.ReleaseWriter("k")
+	c.ReleaseWriter("k") // idempotent
+}
+
+func TestStartupOrphanCleanup(t *testing.T) {
+	root := t.TempDir()
+
+	// Pre-seed: one sealed dir + one orphan old enough + one orphan fresh.
+	sealed := filepath.Join(root, "sealed")
+	_ = os.MkdirAll(sealed, 0o755)
+	_ = os.WriteFile(filepath.Join(sealed, hlsCacheCompleteMarker), nil, 0o644)
+
+	staleOrphan := filepath.Join(root, "stale_orphan")
+	_ = os.MkdirAll(staleOrphan, 0o755)
+	old := time.Now().Add(-2 * hlsCacheStartupOrphanAge)
+	_ = os.Chtimes(staleOrphan, old, old)
+
+	freshOrphan := filepath.Join(root, "fresh_orphan")
+	_ = os.MkdirAll(freshOrphan, 0o755)
+
+	if _, err := NewHLSCache(root, 1); err != nil {
+		t.Fatalf("NewHLSCache: %v", err)
+	}
+
+	if _, err := os.Stat(sealed); err != nil {
+		t.Fatal("sealed dir was wrongly removed")
+	}
+	if _, err := os.Stat(staleOrphan); err == nil {
+		t.Fatal("stale orphan should have been removed at startup")
+	}
+	if _, err := os.Stat(freshOrphan); err != nil {
+		t.Fatal("fresh orphan should be kept (might be a mid-restart encode)")
+	}
+}
+
+func TestHitMissCounters(t *testing.T) {
+	c := newTestCache(t, 1)
+	if s := c.Stats(); s.Hits != 0 || s.Misses != 0 {
+		t.Fatalf("fresh cache stats not zero: %+v", s)
+	}
+	c.RecordHit()
+	c.RecordHit()
+	c.RecordMiss()
+	s := c.Stats()
+	if s.Hits != 2 || s.Misses != 1 {
+		t.Fatalf("counters wrong: %+v", s)
+	}
+	// 2/3 = 67%
+	if got := c.hitRatePercent(); got != 67 {
+		t.Fatalf("hitRatePercent=%d, want 67", got)
+	}
+}
+
+func TestStatsEntryCount(t *testing.T) {
+	c := newTestCache(t, 1)
+	for _, k := range []string{"a", "b", "c"} {
+		dir := c.DirFor(k)
+		_ = os.MkdirAll(dir, 0o755)
+		_ = os.WriteFile(filepath.Join(dir, "x"), []byte("hello"), 0o644)
+	}
+	s := c.Stats()
+	if s.EntryCount != 3 {
+		t.Fatalf("EntryCount=%d, want 3", s.EntryCount)
+	}
+	if s.TotalBytes != 15 {
+		t.Fatalf("TotalBytes=%d, want 15", s.TotalBytes)
+	}
+}
+
+func TestVerifyCompleteRejectsMissingFiles(t *testing.T) {
+	c := newTestCache(t, 1)
+	key := "v"
+	dir := c.DirFor(key)
+	_ = os.MkdirAll(filepath.Join(dir, "video"), 0o755)
+
+	// No .complete yet → reject.
+	if c.VerifyComplete(key, 2) {
+		t.Fatal("VerifyComplete should reject without .complete")
+	}
+
+	// Mark complete but no files → reject.
+	if err := c.MarkComplete(key); err != nil {
+		t.Fatalf("MarkComplete: %v", err)
+	}
+	if c.VerifyComplete(key, 2) {
+		t.Fatal("VerifyComplete should reject when init.mp4 missing")
+	}
+
+	// Write init.mp4, last seg missing → reject.
+	_ = os.WriteFile(filepath.Join(dir, "video", "init.mp4"), []byte("..."), 0o644)
+	if c.VerifyComplete(key, 2) {
+		t.Fatal("VerifyComplete should reject when last segment missing")
+	}
+
+	// Write last seg → pass.
+	_ = os.WriteFile(filepath.Join(dir, "video", "seg-1.m4s"), []byte("..."), 0o644)
+	if !c.VerifyComplete(key, 2) {
+		t.Fatal("VerifyComplete should pass with all files present")
+	}
+
+	// Zero-size last seg → reject.
+	_ = os.WriteFile(filepath.Join(dir, "video", "seg-1.m4s"), nil, 0o644)
+	if c.VerifyComplete(key, 2) {
+		t.Fatal("VerifyComplete should reject zero-size last segment")
+	}
+}
+
+func TestSweepRespectsPinnedExceedsBudget(t *testing.T) {
+	c := newTestCache(t, 1)
+	c.maxBytes = 256 // squeeze
+
+	pinned := c.DirFor("pinned")
+	_ = os.MkdirAll(pinned, 0o755)
+	_ = os.WriteFile(filepath.Join(pinned, "x"), make([]byte, 1024), 0o644)
+	c.Pin("pinned")
+
+	freed, err := c.Sweep()
+	if err != nil {
+		t.Fatalf("Sweep: %v", err)
+	}
+	if freed != 0 {
+		t.Fatalf("nothing should have been freed: got %d", freed)
+	}
+	if _, err := os.Stat(pinned); err != nil {
+		t.Fatal("pinned dir wrongly removed despite over-budget pin")
+	}
+}