docs: describe SQLite logging (logs.db), drop stale log-file docs

Logging moved to a logs.db SQLite store, but MANUAL.md, MANUAL.html and
SPEC.md still documented info_*.log / error_*.log / structured.json and a
midnight-rollover limitation that no longer exist. Replace those sections with
the logs.db reality: levels (info/error to db+stdout/stderr, debug db-only),
progress-to-stdout-only, retention via log_retention_days, and that recent
events are viewable in the dashboard / GET /status.

Closes #6

MANUAL.html

@@ -505,19 +505,19 @@ old.broadcast.tvrip.tt0066026.mkv</code></pre></div>
     <section id="logs">
       <p class="eyebrow">08 — Observability</p>
       <h2>Logs</h2>
-      <p>Three log files are written to the <strong>current working directory</strong> (not the config paths):</p>
+      <p>All logs are written to a <strong>SQLite database, <code>logs.db</code></strong>, in the current working directory (not the config paths). Run the program from where you want it to land — in Docker that's the mounted <code>/data</code>. The <code>logs</code> table has columns <code>id</code>, <code>ts</code>, <code>level</code>, <code>message</code>, <code>file</code>, <code>extra</code>.</p>
       <div class="tablewrap">
         <table>
-          <thead><tr><th>File</th><th>Contents</th></tr></thead>
+          <thead><tr><th>Level</th><th>Where</th><th>Contents</th></tr></thead>
           <tbody>
-            <tr><td><code>info_YYYY-MM-DD.log</code></td><td>INFO messages, dated</td></tr>
-            <tr><td><code>error_YYYY-MM-DD.log</code></td><td>ERROR messages, dated</td></tr>
-            <tr><td><code>structured.json</code></td><td>One JSON object per line — every entry, with <code>timestamp</code>, <code>level</code>, <code>message</code>, optional <code>error</code> &amp; <code>file</code></td></tr>
+            <tr><td><code>info</code></td><td>db + stdout</td><td>Processing / metadata hits / completions</td></tr>
+            <tr><td><code>error</code></td><td>db + stderr</td><td>Failures, with the source <code>file</code> recorded</td></tr>
+            <tr><td><code>debug</code></td><td>db only</td><td>ffprobe output, zscale width, full ffmpeg/opusenc command, OMDb/TVmaze responses (API key redacted) in <code>extra</code></td></tr>
           </tbody>
         </table>
       </div>
-      <p>Run the program from the directory where you want the logs to land. A number of <code>DEBUG</code> lines also print to stdout/stderr (ffprobe output, calculated zscale width, the full ffmpeg command) — intentional, but not written to the log files.</p>
-      <div class="note warn"><span class="tag">Known limitation</span><p>The date in <code>info_*.log</code> / <code>error_*.log</code> filenames is computed at daemon start and does not roll over at midnight. Left running across days, all writes continue into the start-day's file — restart to rotate. <code>structured.json</code> does not rotate at all.</p></div>
+      <p>Live encode progress (<code>encoding … · 47% · 3.2fps · …</code>) prints to <strong>stdout only</strong> and is deliberately not stored, so it can't flood the database.</p>
+      <div class="note"><span class="tag">Retention</span><p>Rows older than <code>log_retention_days</code> (default <code>7</code>) are purged on startup and shutdown. Editable live from the settings page (<code>/settings</code>); applies at the next purge. Recent non-debug events are viewable in the dashboard and via <code>GET /status</code> when <code>http_addr</code> is set.</p></div>
     </section>
 
     <section id="failures">
@@ -527,7 +527,7 @@ old.broadcast.tvrip.tt0066026.mkv</code></pre></div>
       <ul>
         <li>The source <code>.mkv</code> is moved to <code>paths.failed</code> (<code>os.Stat</code>-guarded — if the source is already gone the move is skipped and logged; a move error is logged too). This guarantees the source leaves <code>paths.input</code> on every failure path, so the watcher won't retry it next tick.</li>
         <li>The per-job work directory (partial wav/opus/output.mkv) is deleted unconditionally.</li>
-        <li>The error is logged to <code>error_*.log</code> and <code>structured.json</code> with the source path.</li>
+        <li>The error is logged to <code>logs.db</code> (level <code>error</code>) with the source path, and printed to stderr.</li>
       </ul>
       <p>The watcher continues with the next file; one bad rip won't stop the daemon.</p>
       <div class="note"><span class="tag">Collisions</span><p>If a finished encode would land on a name that already exists in <code>paths.output</code>, the move is refused (no silent overwrite) and the source is routed to <code>paths.failed</code>. This is the path you hit when two sources sanitize to the same name — two re-rips of the same release, or two episodes that both resolve to <code>SxxExx</code>.</p></div>