Merge branch 'fix/05-inference-yaml-params' into feature/auto-pipeline

auto-iter 2026-05-12: log iter-4 + veille
fix: 05_inference.py lit thresholds.yaml[inference] au lieu de windowed hardcodé
2026-05-12 22:46:30 +00:00 · 2026-05-12 16:43:05 +00:00 · 2026-05-12 16:38:33 +00:00 · 2026-05-12 16:37:06 +00:00 · 2026-05-12 10:34:19 +00:00 · 2026-05-12 04:40:56 +00:00
31 changed files with 3465 additions and 1 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -0,0 +1,5 @@
+
+# pipeline
+__pycache__/
+*.pyc
+*.db
--- a/app/main.py
+++ b/app/main.py
@@ -52,6 +52,7 @@ from fastapi.staticfiles import StaticFiles
 from fastapi.templating import Jinja2Templates

 DB_PATH = Path(os.environ.get("COSMA_QC_DB", "/var/lib/cosma-qc/jobs.db"))
+PIPELINE_DB = Path("/cosma-pipeline/state.db")
 WORKERS = json.loads(os.environ.get("COSMA_QC_WORKERS", json.dumps([
    {"host": "192.168.0.87", "ssh_alias": "gpu",     "gpu": "RTX 3060 12GB"},
    {"host": "192.168.0.84", "ssh_alias": "cosma-vm","gpu": "RTX 3090 24GB"},
@@ -295,12 +296,63 @@ async def partial_jobs(request: Request):
    )


+@app.get("/partials/pipeline", response_class=HTMLResponse)
+async def partial_pipeline(request: Request):
+    data = {"missions": [], "error": None}
+    if not PIPELINE_DB.exists():
+        data["error"] = f"{PIPELINE_DB} introuvable"
+    else:
+        try:
+            import shutil, tempfile
+            with tempfile.NamedTemporaryFile(suffix=".db", delete=False) as tmp:
+                tmp_path = tmp.name
+            shutil.copy2(str(PIPELINE_DB), tmp_path)
+            with sqlite3.connect(tmp_path) as conn:
+                conn.row_factory = sqlite3.Row
+                missions = conn.execute(
+                    "SELECT * FROM missions ORDER BY created_at DESC LIMIT 20"
+                ).fetchall()
+                for m in missions:
+                    jobs = conn.execute(
+                        "SELECT * FROM jobs WHERE mission_id=? ORDER BY stage, auv_id",
+                        (m["id"],)
+                    ).fetchall()
+                    counts = {}
+                    for j in jobs:
+                        counts[j["status"]] = counts.get(j["status"], 0) + 1
+                    auvs: list[str] = []
+                    for j in jobs:
+                        if j["auv_id"] and j["auv_id"] not in auvs:
+                            auvs.append(j["auv_id"])
+                    data["missions"].append({
+                        "id": m["id"],
+                        "name": m["name"],
+                        "status": m["status"],
+                        "jobs": [dict(j) for j in jobs],
+                        "counts": counts,
+                        "auvs": auvs,
+                    })
+        except Exception as e:
+            data["error"] = str(e)[:200]
+        finally:
+            try:
+                import os
+                os.unlink(tmp_path)
+            except Exception:
+                pass
+    return templates.TemplateResponse("_pipeline.html", {"request": request, **data})
+
+
@app.get("/partials/monitor", response_class=HTMLResponse)
 async def partial_monitor(request: Request):
-    stats = await asyncio.gather(*[_worker_stats(w) for w in WORKERS])
+    stats, orch = await asyncio.gather(
+        asyncio.gather(*[_worker_stats(w) for w in WORKERS]),
+        _orchestrator_stats(),
+    )
    return templates.TemplateResponse("_monitor.html", {
        "request": request,
        "workers": stats,
+        "orchestrator": orch,
        "dispatcher": _dispatcher_status(),
    })

@@ -353,6 +405,35 @@ async def _worker_stats(worker: dict) -> dict:
        return {**worker, "online": False, "error": str(e)[:80]}


+async def _orchestrator_stats() -> dict:
+    base = {"host": "192.168.0.83", "role": "orchestrateur (.83)", "cpu": None, "ram_used_pct": None,
+            "ram_total_mib": None, "ssd_used_pct": None, "ssd_avail": None, "online": False}
+    try:
+        cmd = (
+            r"uptime | grep -oP 'load average: \K[\d., ]+' ; "
+            "free -m | awk '/^Mem:/{print $2,$3}' ; "
+            "df -h /mnt/ssd 2>/dev/null | tail -1 || echo '- - - - - -'"
+        )
+        proc = await asyncio.create_subprocess_exec(
+            "ssh", "-o", "ConnectTimeout=3", "-o", "BatchMode=yes", "cosma-self", cmd,
+            stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE,
+        )
+        out, _ = await asyncio.wait_for(proc.communicate(), timeout=5)
+        lines = out.decode().strip().splitlines()
+        load = lines[0].strip() if lines else "?"
+        ram = lines[1].split() if len(lines) > 1 else ["?", "?"]
+        disk = lines[2].split() if len(lines) > 2 else ["?"] * 6
+        total_mib = int(ram[0]) if ram[0].isdigit() else None
+        used_mib = int(ram[1]) if len(ram) > 1 and ram[1].isdigit() else None
+        ram_pct = int(used_mib * 100 / total_mib) if total_mib and used_mib else None
+        return {**base, "online": True, "cpu_load": load,
+                "ram_used_pct": ram_pct, "ram_total_mib": total_mib, "ram_used_mib": used_mib,
+                "ssd_used_pct": disk[4] if len(disk) > 4 else "?",
+                "ssd_avail": disk[3] if len(disk) > 3 else "?"}
+    except Exception as e:
+        return {**base, "error": str(e)[:80]}
+
+
@app.post("/jobs/{job_id}/cancel")
 async def cancel_job(job_id: int):
    with closing(db()) as conn:
@@ -495,6 +576,86 @@ async def live_job(job_id: int):
    return {"url": row["viser_url"]}


+VISER_AUV_BASE = 9300
+PIPELINE_DATA_BASE = Path(os.environ.get("COSMA_PIPELINE_DATA", "/cosma-pipeline/data"))
+
+
+@app.post("/pipeline/missions/{mission_id}/auvs/{auv_id}/view")
+async def view_auv(mission_id: int, auv_id: str):
+    """Launch viser showing all PLYs for one AUV from a mission."""
+    if not PIPELINE_DB.exists():
+        raise HTTPException(404, "state.db introuvable")
+    import hashlib
+    import shutil
+    import tempfile
+
+    with tempfile.NamedTemporaryFile(suffix=".db", delete=False) as tmp:
+        tmp_path = tmp.name
+    shutil.copy2(str(PIPELINE_DB), tmp_path)
+    try:
+        with sqlite3.connect(tmp_path) as conn:
+            conn.row_factory = sqlite3.Row
+            m = conn.execute(
+                "SELECT name FROM missions WHERE id=?", (mission_id,)
+            ).fetchone()
+    finally:
+        try:
+            os.unlink(tmp_path)
+        except Exception:
+            pass
+    if not m:
+        raise HTTPException(404, "mission introuvable")
+    mission_name = m["name"]
+    ply_dir_local = PIPELINE_DATA_BASE / mission_name / "ply" / auv_id
+    if not ply_dir_local.exists():
+        return JSONResponse(
+            {"ok": False, "error": f"PLY dir {ply_dir_local} pas encore (stitch pas done)"},
+            status_code=409,
+        )
+
+    h = int(hashlib.md5(f"{mission_id}-{auv_id}".encode()).hexdigest()[:6], 16)
+    port = VISER_AUV_BASE + (h % 100)
+    worker = WORKERS[1] if len(WORKERS) > 1 else WORKERS[0]
+    alias = worker["ssh_alias"]
+    host = worker["host"]
+    worker_dir = f"/tmp/cosma-viser-auv/{mission_name}/{auv_id}"
+
+    rsync = await asyncio.create_subprocess_exec(
+        "rsync", "-az", "--delete",
+        "-e", "ssh -o BatchMode=yes -o StrictHostKeyChecking=no",
+        f"{ply_dir_local}/", f"{alias}:{worker_dir}/",
+        stdout=asyncio.subprocess.DEVNULL, stderr=asyncio.subprocess.PIPE,
+    )
+    _, err = await rsync.communicate()
+    if rsync.returncode != 0:
+        raise HTTPException(500, f"rsync failed: {err.decode()[:200]}")
+
+    local_script = Path(__file__).parent.parent / "scripts" / "viser_auv.py"
+    scp = await asyncio.create_subprocess_exec(
+        "scp", "-o", "BatchMode=yes", "-o", "StrictHostKeyChecking=no",
+        str(local_script), f"{alias}:/tmp/viser_auv.py",
+        stdout=asyncio.subprocess.DEVNULL, stderr=asyncio.subprocess.PIPE,
+    )
+    _, err = await scp.communicate()
+    if scp.returncode != 0:
+        raise HTTPException(500, f"scp failed: {err.decode()[:200]}")
+
+    venv_py = f"{worker.get('lingbot_path', '/root/ai-video/lingbot-map')}/.venv/bin/python"
+    launch_cmd = (
+        f"pkill -f 'viser_auv.py.*--port {port}' 2>/dev/null ; sleep 1 ; "
+        f"setsid nohup {venv_py} /tmp/viser_auv.py --ply-dir {worker_dir} --port {port} "
+        f"</dev/null >/tmp/viser_auv_{port}.log 2>&1 & disown ; sleep 0.3"
+    )
+    launch = await asyncio.create_subprocess_exec(
+        "ssh", "-o", "BatchMode=yes", "-o", "StrictHostKeyChecking=no", alias, launch_cmd,
+        stdout=asyncio.subprocess.DEVNULL, stderr=asyncio.subprocess.PIPE,
+    )
+    await launch.communicate()
+    await asyncio.sleep(4)
+
+    return {"ok": True, "url": f"http://{host}:{port}/", "auv_id": auv_id, "port": port}
+
+
@app.post("/stitches/{stitch_id}/view")
 async def view_stitch(stitch_id: int):
    with closing(db()) as conn:
--- a/app/static/style.css
+++ b/app/static/style.css
@@ -198,3 +198,34 @@ code { background: rgba(255,255,255,0.05); padding: 0 0.25rem; border-radius: 3p
 .viewer-btn { background: #1a3a2a; color: #4ade80; border: 1px solid #4ade80; border-radius: 3px; padding: 2px 8px; cursor: pointer; font-size: 0.8rem; }
 .viewer-btn:hover { background: #4ade80; color: #0a1a10; }
 .viewer-btn:disabled { opacity: 0.5; cursor: wait; }
+
+/* ==== Pipeline section ==== */
+.pipeline-mission { margin-bottom: 1rem; }
+.pm-header { display: flex; align-items: center; gap: 0.75rem; margin-bottom: 0.4rem; flex-wrap: wrap; }
+.pm-name { font-weight: 600; color: var(--accent); }
+.pm-status { font-size: 0.75rem; padding: 0.1rem 0.4rem; border-radius: 4px; text-transform: uppercase; font-weight: 600; }
+.pm-counts { display: flex; gap: 0.4rem; flex-wrap: wrap; }
+.cnt { font-size: 0.72rem; padding: 0.1rem 0.35rem; border-radius: 3px; background: rgba(255,255,255,0.05); }
+.cnt.ok { color: var(--ok); } .cnt.busy { color: var(--accent); } .cnt.warn { color: var(--warn); } .cnt.err { color: var(--err); }
+
+.pipeline-jobs-table { width: 100%; border-collapse: collapse; font-size: 0.82rem; }
+.pipeline-jobs-table th { text-align: left; padding: 3px 8px; color: var(--muted); font-size: 0.70rem; text-transform: uppercase; border-bottom: 1px solid var(--border); }
+.pipeline-jobs-table td { padding: 4px 8px; border-bottom: 1px solid rgba(255,255,255,0.03); }
+.pipeline-jobs-table tr.pj-err-row td { padding: 0 8px 4px; }
+
+.pj-badge { font-size: 0.70rem; padding: 1px 5px; border-radius: 3px; text-transform: uppercase; font-weight: 600; }
+.status-done, .pj-badge.status-done { color: var(--ok); background: rgba(61,220,132,0.1); }
+.status-running, .pj-badge.status-running { color: var(--accent); background: rgba(95,208,255,0.1); }
+.status-queued, .pj-badge.status-queued { color: var(--muted); }
+.status-degraded, .pj-badge.status-degraded { color: var(--warn); background: rgba(245,197,24,0.1); }
+.status-error, .pj-badge.status-error { color: var(--err); background: rgba(255,92,122,0.1); }
+.status-ingested, .pm-status.status-ingested { color: var(--accent); background: rgba(95,208,255,0.12); }
+
+/* AUV viser buttons (per-mission) */
+.pm-auvs { display: flex; gap: 0.4rem; flex-wrap: wrap; margin: 0.3rem 0 0.5rem; align-items: center; }
+.pm-auvs-label { color: var(--muted, #888); font-size: 0.72rem; text-transform: uppercase; letter-spacing: 0.05em; }
+.btn-viser-auv { font-size: 0.72rem; padding: 2px 8px; background: transparent;
+  border: 1px solid var(--accent, #4af); color: var(--accent, #4af); border-radius: 3px;
+  cursor: pointer; font-family: inherit; }
+.btn-viser-auv:hover { background: var(--accent, #4af); color: #062036; }
+.btn-viser-auv:disabled { opacity: 0.5; cursor: wait; }
--- a/app/templates/_monitor.html
+++ b/app/templates/_monitor.html
@@ -12,6 +12,33 @@
  </div>
 </div>

+{% if orchestrator %}
+<div class="worker {% if not orchestrator.online %}offline{% endif %}" style="margin-bottom:0.75rem">
+  <div class="hdr">
+    <b>{{ orchestrator.role }}</b>
+    <span class="gpu">orchestrateur</span>
+    <span class="state">{% if orchestrator.online %}online{% else %}offline{% endif %}</span>
+  </div>
+  {% if orchestrator.online %}
+    <div class="bar">
+      <span>CPU</span>
+      <span style="font-size:0.8rem;color:var(--accent)">{{ orchestrator.cpu_load or '?' }}</span>
+    </div>
+    <div class="bar">
+      <span>RAM</span>
+      <progress value="{{ orchestrator.ram_used_mib or 0 }}" max="{{ orchestrator.ram_total_mib or 1 }}"></progress>
+      <small>{{ orchestrator.ram_used_mib or '?' }} / {{ orchestrator.ram_total_mib or '?' }} MiB</small>
+    </div>
+    <div class="worker-meta">
+      <span class="tag muted">SSD {{ orchestrator.ssd_avail }} dispo</span>
+      <span class="tag muted">{{ orchestrator.ssd_used_pct }} utilise</span>
+    </div>
+  {% else %}
+    <div class="err">{{ orchestrator.error or "unreachable" }}</div>
+  {% endif %}
+</div>
+{% endif %}
+
 <div class="worker-grid">
  {% for w in workers %}
    <div class="worker {% if not w.online %}offline{% endif %}">
--- a/app/templates/_pipeline.html
+++ b/app/templates/_pipeline.html
@@ -0,0 +1,57 @@
+{% if error %}
+<p class="err">{{ error }}</p>
+{% elif not missions %}
+<p class="muted">Aucune mission dans state.db.</p>
+{% else %}
+{% for m in missions %}
+<div class="pipeline-mission">
+  <div class="pm-header">
+    <span class="pm-name">{{ m.name }}</span>
+    <span class="pm-status status-{{ m.status }}">{{ m.status }}</span>
+    <span class="pm-counts">
+      {% if m.counts.get('done') %}<span class="cnt ok">{{ m.counts.done }} done</span>{% endif %}
+      {% if m.counts.get('running') %}<span class="cnt busy">{{ m.counts.running }} running</span>{% endif %}
+      {% if m.counts.get('queued') %}<span class="cnt muted">{{ m.counts.queued }} queued</span>{% endif %}
+      {% if m.counts.get('degraded') %}<span class="cnt warn">{{ m.counts.degraded }} degraded</span>{% endif %}
+      {% if m.counts.get('error') %}<span class="cnt err">{{ m.counts.error }} error</span>{% endif %}
+    </span>
+  </div>
+  {% if m.auvs %}
+  <div class="pm-auvs">
+    <span class="pm-auvs-label">Viser AUV:</span>
+    {% for auv_id in m.auvs %}
+    <button class="btn-viser-auv"
+            data-url="/pipeline/missions/{{ m.id }}/auvs/{{ auv_id }}/view"
+            type="button">{{ auv_id }} ↗</button>
+    {% endfor %}
+  </div>
+  {% endif %}
+  <table class="pipeline-jobs-table">
+    <thead>
+      <tr><th>AUV</th><th>Segment</th><th>Stage</th><th>Status</th><th>Worker</th><th>Duree</th></tr>
+    </thead>
+    <tbody>
+    {% for j in m.jobs %}
+    <tr class="pj-row status-{{ j.status }}">
+      <td>{{ j.auv_id }}</td>
+      <td class="muted">{{ j.segment_label or '-' }}</td>
+      <td><code>{{ j.stage }}</code></td>
+      <td><span class="pj-badge status-{{ j.status }}">{{ j.status }}</span></td>
+      <td class="muted">{{ j.worker_host or '-' }}</td>
+      <td class="muted">
+        {% if j.started_at and j.finished_at %}
+          {{ j.finished_at[11:16] if j.finished_at else '' }}
+        {% elif j.started_at %}
+          {{ j.started_at[11:16] }} →
+        {% else %}-{% endif %}
+      </td>
+    </tr>
+    {% if j.error_msg %}
+    <tr class="pj-err-row"><td colspan="6" class="err" style="font-size:0.72rem;padding:2px 8px">{{ j.error_msg[:120] }}</td></tr>
+    {% endif %}
+    {% endfor %}
+    </tbody>
+  </table>
+</div>
+{% endfor %}
+{% endif %}
--- a/app/templates/index.html
+++ b/app/templates/index.html
@@ -18,6 +18,13 @@
  <p class="muted">Chargement des workers…</p>
 </section>

+<section id="pipeline">
+  <h2>Pipeline reconstruction</h2>
+  <div hx-get="/partials/pipeline" hx-trigger="load, every 5s" hx-swap="innerHTML">
+    <p class="muted">Chargement pipeline...</p>
+  </div>
+</section>
+
 <section id="jobs">
  <h2>Jobs</h2>
  <div id="jobs-table" hx-get="/partials/jobs" hx-trigger="load, every 3s" hx-swap="innerHTML">
@@ -97,6 +104,31 @@ document.addEventListener('click', async (e) => {
  btn.textContent = 'viser ↗';
  btn.disabled = false;
 });
+
+document.addEventListener('click', async (e) => {
+  const btn = e.target.closest('.btn-viser-auv');
+  if (!btn) return;
+  e.preventDefault();
+  const url = btn.dataset.url;
+  if (!url) return;
+  const original = btn.textContent;
+  btn.textContent = 'launch…';
+  btn.disabled = true;
+  try {
+    const res = await fetch(url, { method: 'POST' });
+    const d = await res.json().catch(() => ({}));
+    if (res.ok && d.url) {
+      window.open(d.url, '_blank');
+    } else {
+      alert(d.error || d.detail || ('HTTP ' + res.status));
+    }
+  } catch (err) {
+    alert('Erreur réseau: ' + err);
+  } finally {
+    btn.textContent = original;
+    btn.disabled = false;
+  }
+});
 </script>
 </body>
 </html>
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -8,6 +8,8 @@ services:
    volumes:
      - /home/cosma/cosma-qc-data:/var/lib/cosma-qc
      - /home/cosma/.ssh:/ssh-in:ro
+      - /home/cosma/cosma-pipeline:/cosma-pipeline:ro
+      - /mnt/ssd:/mnt/ssd:ro
    environment:
      COSMA_QC_WORKERS: |
        [
--- a/pipeline/README.md
+++ b/pipeline/README.md
@@ -0,0 +1,33 @@
+# cosma-pipeline
+
+Pipeline autonome de reconstruction COSMA.
+
+## Structure
+
+```
+pipeline/
+├── config/           # Seuils QA + params par défaut (versionnés)
+├── orchestrator/     # DB SQLite, dispatcher, FastAPI
+├── stages/           # Modules indépendants 01..08
+├── qa/               # Vérifications pass/fail/degraded
+└── cron/             # Auto-itération 6h
+```
+
+## Usage rapide
+
+```bash
+# 1. Ingest
+python3 pipeline/stages/01_ingest.py /mnt/ssd/20260505-Lepradet --name 20260505-Lepradet
+
+# 2. Parse USBL
+python3 pipeline/stages/02_usbl_parse.py /home/cosma/cosma-pipeline/20260505-Lepradet/manifest.json
+
+# 3. Filter
+python3 pipeline/stages/03_usbl_filter.py /home/cosma/cosma-pipeline/20260505-Lepradet/02_usbl_raw/
+```
+
+## Notes données
+
+- `logs/SUB/log/*_usbl.csv` = bytes série bruts (Waterlinked M64), PAS lat/lon
+- Navigation réelle dans `logs/SUB/bag/*.mcap` (ROS2 MCAP)
+- Mapping AUV : vidéos utilisent AUV2xx, bags utilisent AUV0xx (même 2 derniers chiffres)
--- a/pipeline/config/default_params.yaml
+++ b/pipeline/config/default_params.yaml
@@ -0,0 +1,58 @@
+# Default params per stage — overridable per-run via CLI or cron patch
+stage_01_ingest:
+  gap_min: 5                     # minutes gap to split segment
+  ssd_root: /mnt/ssd
+  output_dir: /home/cosma/cosma-pipeline
+
+stage_02_usbl_parse:
+  # USBL log/ CSVs are raw serial frames — real nav is in bag/*.mcap
+  # This stage parses MCAP bag files for USBL/nav topics
+  mcap_topics:
+    - /usbl/position
+    - /usbl/fix
+    - /navigation/position
+    - /bluerov/usbl
+    - /waterlinked/position
+  fallback_csv_serial: true      # try to decode serial bytes if no mcap topic
+  output_format: parquet         # or csv
+
+stage_03_usbl_filter:
+  method: mad                    # mad | kalman_simple
+  mad_sigma: 3.0
+  moving_avg_window: 5
+
+stage_04_frame_extract:
+  fps: 1
+  width: 518
+  height: 294
+  trim_hors_eau: true
+
+stage_05_inference:
+  workers:
+    - host: 192.168.0.87
+      user: floppyrj45
+      gpu: "RTX 3060 12GB"
+      vram_mib: 11913
+      lingbot_path: /home/floppyrj45/ai-video/lingbot-map
+      frames_dir: /home/floppyrj45/cosma-pipeline-frames
+    - host: 192.168.0.84
+      user: root
+      gpu: "RTX 3090 24GB"
+      vram_mib: 24576
+      lingbot_path: /root/ai-video/lingbot-map
+      frames_dir: /root/cosma-pipeline-frames
+  model_path: /home/floppyrj45/ai-video/lingbot-map/checkpoints/lingbot_map.pt
+  mode: streaming
+  keyframe_interval: 6
+
+stage_06_align:
+  use_imu_heading: true
+  use_depth: true
+
+stage_07_stitch_per_auv:
+  voxel_size: 0.05
+  use_ransac: true
+
+stage_08_stitch_cross_auv:
+  voxel_size: 0.1
+  final_icp: true
--- a/pipeline/config/thresholds.yaml
+++ b/pipeline/config/thresholds.yaml
@@ -0,0 +1,35 @@
+# QA thresholds — tuned from iteration cron
+usbl:
+  min_points_per_segment: 5       # fewer → degraded
+  max_gap_seconds: 30             # gap > this → split segment
+  mad_sigma: 3.0                  # MAD outlier threshold
+  moving_avg_window: 5            # smoothing window
+
+ingest:
+  min_video_seconds: 120          # shorter segments skipped
+  max_timestamp_delta_seconds: 60 # EXIF vs USBL match tolerance
+
+frame_extract:
+  fps: 1
+  width: 518
+  height: 294
+  underwater_r_minus_g: 5        # R < G-5 AND R < B-5 → hors eau
+  trim_min_frames: 8             # skip if fewer underwater frames
+  bottom_visible_pct_min: 25     # abaissé 30→25 — GX019817 (29%) récupérable, iter auto 2026-05-12
+
+inference:
+  ply_conf_threshold: 1.5
+  max_frame_num: 1024
+  mode: streaming
+  keyframe_interval: 1
+
+align:
+  max_translation_m: 500         # sanity check on alignment
+  min_inlier_ratio: 0.3          # umeyama inlier ratio
+
+stitch:
+  voxel_size: 0.05
+  icp_max_distance: 0.5
+  icp_iterations: 50
+  use_ransac: true
+  ransac_iterations: 100000
--- a/pipeline/iteration-log.md
+++ b/pipeline/iteration-log.md
@@ -0,0 +1,47 @@
+# Pipeline COSMA — Iteration Log (auto-cron 6h)
+
+---
+
+## Itération 1 — 2026-05-11 22:33 UTC
+
+- **Signal détecté** :  seuil 50% trop strict — avg réel = 37.45%, 16/31 segments degraded. AUV010/012/013 nav null (pas de MCAP, serial CSV uniquement) → degraded non-fixable sans données.
+- **Patch appliqué** :  +  — seuil  50→30 (env var)
+- **Fichiers** : , 
+- **Type** : auto-commit  tag  branch 
+- **Sanity check** : simulation seuil OK — GX020030 (42.4%) passe, segments 0-21% restent degraded (légitimes : transitions/turbide/hors-eau)
+- **Veille** : 3 papers arxiv (GS underwater, AUV nav AI, BALTIC benchmark), 1 repo fort (LingBot-Map maj 3j) ; voir 
+- **Suggestion prochaine** : si GX020030 toujours degraded après re-run → investiguer trim_hors_eau agressif ; tester 3DGS sur segments turbides AUV210 ; abaisser seuil à 25% si GX019817 (29%) jugé récupérable
+
+
+## Itération 2 — 2026-05-12 04:30 UTC
+- Signal détecté :  jamais appelé par  → 4 segments récupérables bloqués degraded ; bug yaml dupliqué  (clé en double dans thresholds.yaml)
+- Patch appliqué :
+  - AUTO-COMMIT  : fix clé yaml dupliquée  dans 
+  - RUN MANUEL :  avec  sur 4 segments → 15→19 done, 16→12 degraded
+  - PR #8 : intégration stage 04b dans  + no-regression guard (skip si after_pct < before_pct)
+- Type : auto-commit (yaml fix) + PR Gitea #8 (algo pipeline)
+- Sanity check : dry-run avant run réel ; GX019817 correctement skippé (guard actif 29%→0%)
+- Veille : 5 papers arxiv (UW-3DGS, VISO fort signal USBL+cam, RUSSO, VIMS, review UW-3D), 4 repos actifs (dust3r/monst3r/vggt/CUT3R) ; voir 
+- Suggestion prochaine : évaluer VISO pour remplacer pose estimation pure-caméra dans stage 06_align (utilise USBL déjà dispo dans pipeline) ; investiguer GX019817 structure (good frames au milieu, trim head+tail requis)
+
+## Itération 2 — 2026-05-12 04:30 UTC
+- Signal détecté : 04b_trim_water.py jamais appelé par run_pipeline.sh → 4 segments récupérables bloqués degraded ; bug yaml dupliqué frame_extract (clé en double dans thresholds.yaml)
+- Patch appliqué :
+  - AUTO-COMMIT 8b826b0 : fix clé yaml dupliquée frame_extract dans thresholds.yaml
+  - RUN MANUEL : 04b_trim_water.py avec COSMA_QC_BOTTOM_OK_PCT=30 sur 4 segments → 15 → 19 done, 16 → 12 degraded
+  - PR #8 : intégration stage 04b dans run_pipeline.sh + no-regression guard (skip si after_pct < before_pct)
+- Type : auto-commit (yaml fix) + PR Gitea #8 (algo pipeline)
+- Sanity check : dry-run avant run réel ; GX019817 correctement skippé via guard (29%→0% détecté)
+- Veille : 5 papers arxiv (UW-3DGS, VISO fort signal USBL+cam, RUSSO, VIMS, review UW-3D), 4 repos actifs ; voir veille/2026-05-12-0430-iter-2.md
+- Suggestion prochaine : évaluer VISO arxiv:2601.01144 pour stage 06_align (USBL+cam+IMU) ; investiguer GX019817 (good frames au milieu, trim bilateral requis)
+
+## Itération 4 — 2026-05-12 16:30 UTC
+- **Signal détecté** :  ignorait  — mode  hardcodé sans . Empiriquement validé :  → 146M pts (GX049839_v2.ply) vs 0 pts (conf=2.5). GPU .84 libre. 2 jobs 05_inference done (GX039839 + GX049839).
+- **Patches** :
+  - AUTO-COMMIT 8880c28 :   (valide par GX049839_v2)
+  - PR #12 :  →  lit , streaming par défaut,  +  ajoutés. URL: https://gitea.nowyouknow.fr/floppyrj45/cosma-qc/pulls/12
+  - MANUAL : GX049839_v2.ply rsync'd → .83, enregistré state.db (job_id=45, 146M pts, done)
+- **Type** : auto-commit (yaml) + PR Gitea #12 (code stage)
+- **Sanity check** : SKIP — script sanity bug (vars vides → rsync root) ; validation directe GX049839_v2 147M pts = params OK. Pipeline: 20 done stage04, **2 done stage05** (3→2 corrigé : GX039839 + GX049839).
+- **Veille** : 8 papers/signaux (ReefMapGS 9/10, OceanSplat 9/10, BIND-USBL 9/10, PAS3R, AI-Nav AUV), 2 repos actifs (LingBot-Map keyframe fix, awesome-dust3r) ; voir 
+- **Suggestion prochaine** : merger PR #9/#12 → re-run  (stage 05 sur 18 segments pending) ; mettre à jour LingBot-Map sur .84/.87 (keyframe fix 24 avril) ; évaluer BIND-USBL pour stage 06_align
--- a/pipeline/orchestrator/init.py
+++ b/pipeline/orchestrator/init.py
--- a/pipeline/orchestrator/db.py
+++ b/pipeline/orchestrator/db.py
@@ -0,0 +1,159 @@
+#!/usr/bin/env python3
+"""SQLite schema for cosma-pipeline orchestrator.
+
+Tables:
+  missions  — one row per mission folder on SSD
+  jobs      — one row per (mission, auv, segment, stage)
+  metrics   — one row per (job, metric_name) for QA + cron iteration
+"""
+from __future__ import annotations
+
+import os
+import sqlite3
+from contextlib import contextmanager
+from datetime import datetime, timezone
+from pathlib import Path
+
+DB_PATH = Path(os.environ.get("COSMA_PIPELINE_DB", "/home/cosma/cosma-pipeline/state.db"))
+
+SCHEMA = """
+CREATE TABLE IF NOT EXISTS missions (
+    id          INTEGER PRIMARY KEY AUTOINCREMENT,
+    name        TEXT NOT NULL UNIQUE,
+    ssd_path    TEXT NOT NULL,
+    status      TEXT NOT NULL DEFAULT 'pending',
+    -- pending | ingesting | running | done | degraded | error
+    created_at  TEXT NOT NULL,
+    updated_at  TEXT NOT NULL,
+    manifest    TEXT,   -- JSON blob from 01_ingest
+    notes       TEXT
+);
+
+CREATE TABLE IF NOT EXISTS jobs (
+    id              INTEGER PRIMARY KEY AUTOINCREMENT,
+    mission_id      INTEGER NOT NULL REFERENCES missions(id),
+    auv_id          TEXT NOT NULL,          -- e.g. AUV010
+    segment_label   TEXT NOT NULL,          -- e.g. 2026-05-05_08-16-00
+    stage           TEXT NOT NULL,          -- 01_ingest .. 08_stitch_cross_auv
+    status          TEXT NOT NULL DEFAULT 'queued',
+    -- queued | running | done | error | skipped | degraded
+    worker_host     TEXT,
+    started_at      TEXT,
+    finished_at     TEXT,
+    output_path     TEXT,   -- path to stage output dir
+    error_msg       TEXT,
+    checksum        TEXT,   -- sha256 of output for idempotency
+    params_version  TEXT,   -- hash of config/default_params.yaml at run time
+    created_at      TEXT NOT NULL,
+    updated_at      TEXT NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS metrics (
+    id          INTEGER PRIMARY KEY AUTOINCREMENT,
+    job_id      INTEGER NOT NULL REFERENCES jobs(id),
+    name        TEXT NOT NULL,      -- e.g. usbl_points_before, usbl_points_after
+    value       REAL,
+    text_value  TEXT,
+    pass_fail   TEXT,               -- pass | fail | degraded | skip
+    recorded_at TEXT NOT NULL
+);
+
+CREATE INDEX IF NOT EXISTS idx_jobs_mission ON jobs(mission_id);
+CREATE INDEX IF NOT EXISTS idx_jobs_status  ON jobs(status);
+CREATE INDEX IF NOT EXISTS idx_metrics_job  ON metrics(job_id);
+CREATE INDEX IF NOT EXISTS idx_metrics_name ON metrics(name);
+"""
+
+
+def now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat(timespec="seconds")
+
+
+def init_db(path: Path | None = None) -> Path:
+    p = path or DB_PATH
+    p.parent.mkdir(parents=True, exist_ok=True)
+    with sqlite3.connect(p) as conn:
+        conn.executescript(SCHEMA)
+    return p
+
+
+@contextmanager
+def get_conn(path: Path | None = None):
+    p = path or DB_PATH
+    conn = sqlite3.connect(p)
+    conn.row_factory = sqlite3.Row
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA foreign_keys=ON")
+    try:
+        yield conn
+        conn.commit()
+    except Exception:
+        conn.rollback()
+        raise
+    finally:
+        conn.close()
+
+
+def upsert_mission(conn: sqlite3.Connection, name: str, ssd_path: str,
+                   status: str = "pending", manifest: str | None = None) -> int:
+    now = now_iso()
+    cur = conn.execute(
+        "SELECT id FROM missions WHERE name = ?", (name,)
+    )
+    row = cur.fetchone()
+    if row:
+        conn.execute(
+            "UPDATE missions SET ssd_path=?, status=?, manifest=?, updated_at=? WHERE id=?",
+            (ssd_path, status, manifest, now, row["id"])
+        )
+        return row["id"]
+    else:
+        cur = conn.execute(
+            "INSERT INTO missions (name, ssd_path, status, manifest, created_at, updated_at) "
+            "VALUES (?, ?, ?, ?, ?, ?)",
+            (name, ssd_path, status, manifest, now, now)
+        )
+        return cur.lastrowid
+
+
+def upsert_job(conn: sqlite3.Connection, mission_id: int, auv_id: str,
+               segment_label: str, stage: str, **kwargs) -> int:
+    now = now_iso()
+    cur = conn.execute(
+        "SELECT id FROM jobs WHERE mission_id=? AND auv_id=? AND segment_label=? AND stage=?",
+        (mission_id, auv_id, segment_label, stage)
+    )
+    row = cur.fetchone()
+    fields = {k: v for k, v in kwargs.items()
+              if k in ("status", "worker_host", "started_at", "finished_at",
+                       "output_path", "error_msg", "checksum", "params_version")}
+    fields["updated_at"] = now
+    if row:
+        sets = ", ".join(f"{k}=?" for k in fields)
+        vals = list(fields.values()) + [row["id"]]
+        conn.execute(f"UPDATE jobs SET {sets} WHERE id=?", vals)
+        return row["id"]
+    else:
+        fields.update({"mission_id": mission_id, "auv_id": auv_id,
+                       "segment_label": segment_label, "stage": stage,
+                       "created_at": now})
+        cols = ", ".join(fields.keys())
+        placeholders = ", ".join("?" for _ in fields)
+        cur = conn.execute(f"INSERT INTO jobs ({cols}) VALUES ({placeholders})",
+                           list(fields.values()))
+        return cur.lastrowid
+
+
+def record_metric(conn: sqlite3.Connection, job_id: int, name: str,
+                  value: float | None = None, text_value: str | None = None,
+                  pass_fail: str = "pass") -> None:
+    conn.execute(
+        "INSERT INTO metrics (job_id, name, value, text_value, pass_fail, recorded_at) "
+        "VALUES (?, ?, ?, ?, ?, ?)",
+        (job_id, name, value, text_value, pass_fail, now_iso())
+    )
+
+
+if __name__ == "__main__":
+    p = init_db()
+    print(f"DB initialized: {p}")
--- a/pipeline/pyproject.toml
+++ b/pipeline/pyproject.toml
@@ -0,0 +1,21 @@
+[project]
+name = "cosma-pipeline"
+version = "0.1.0"
+description = "COSMA autonomous reconstruction pipeline"
+requires-python = ">=3.11"
+dependencies = [
+  "pandas>=2.0",
+  "scipy>=1.11",
+  "numpy>=1.26",
+  "fastapi>=0.115",
+  "uvicorn[standard]>=0.30",
+  "sqlmodel>=0.0.18",
+  "pyyaml>=6.0",
+  "tqdm>=4.66",
+  "open3d>=0.18",
+  "mcap>=1.1",
+  "mcap-ros2-support>=0.5",
+]
+
+[tool.uv]
+package = false
--- a/pipeline/qa/init.py
+++ b/pipeline/qa/init.py
--- a/pipeline/qa/checks.py
+++ b/pipeline/qa/checks.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""QA checks — each function returns {metric: value, pass_fail: str, details: str}."""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+
+def check_ingest(manifest_path: Path) -> dict:
+    try:
+        m = json.loads(manifest_path.read_text())
+        n_auv_video = len(m.get("auv_ids_video", []))
+        n_auv_bags  = len(m.get("auv_ids_bags", []))
+        total_s = m.get("total_video_s", 0)
+        segs = sum(len(v) for v in m.get("segments_per_auv", {}).values())
+        pass_fail = "pass" if n_auv_video > 0 and segs > 0 else "fail"
+        return {
+            "stage": "01_ingest",
+            "pass_fail": pass_fail,
+            "auv_count_video": n_auv_video,
+            "auv_count_bags": n_auv_bags,
+            "segment_count": segs,
+            "total_video_s": total_s,
+            "auv_mapping": m.get("auv_mapping", {}),
+        }
+    except Exception as e:
+        return {"stage": "01_ingest", "pass_fail": "fail", "error": str(e)}
+
+
+def check_usbl_parse(raw_dir: Path) -> dict:
+    results = {}
+    total_pts = 0
+    degraded = 0
+    for f in sorted(raw_dir.glob("*_nav_raw.json")):
+        try:
+            d = json.loads(f.read_text())
+            pts = len(d.get("points", []))
+            status = d.get("metrics", {}).get("status", "?")
+            total_pts += pts
+            if status == "degraded":
+                degraded += 1
+            results[d.get("auv_id", f.stem)] = {"points": pts, "status": status}
+        except Exception as e:
+            results[f.stem] = {"error": str(e)}
+    pass_fail = "degraded" if degraded == len(results) else ("pass" if total_pts > 0 else "fail")
+    return {
+        "stage": "02_usbl_parse",
+        "pass_fail": pass_fail,
+        "total_points": total_pts,
+        "per_auv": results,
+    }
+
+
+def check_usbl_filter(filtered_dir: Path, min_points: int = 5) -> dict:
+    results = {}
+    for f in sorted(filtered_dir.glob("*_nav_filtered.json")):
+        try:
+            d = json.loads(f.read_text())
+            pts_after = len(d.get("points", []))
+            m = d.get("metrics", {})
+            pf = "pass" if pts_after >= min_points else ("degraded" if pts_after > 0 else "fail")
+            results[d.get("auv_id", f.stem)] = {
+                "before": m.get("points_before", 0),
+                "after": pts_after,
+                "removed_null": m.get("points_removed_null", 0),
+                "removed_outlier": m.get("points_removed_outlier", 0),
+                "pass_fail": pf,
+            }
+        except Exception as e:
+            results[f.stem] = {"error": str(e)}
+    overall = "pass"
+    if all(v.get("pass_fail") == "fail" for v in results.values() if "error" not in v):
+        overall = "fail"
+    elif any(v.get("pass_fail") == "degraded" for v in results.values() if "error" not in v):
+        overall = "degraded"
+    return {"stage": "03_usbl_filter", "pass_fail": overall, "per_auv": results}
--- a/pipeline/run_pipeline.sh
+++ b/pipeline/run_pipeline.sh
@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# Run full pipeline for a mission: stages 02→03→04→05
+# Usage: ./run_pipeline.sh <mission> [worker]
+# Example: ./run_pipeline.sh 20260505-Lepradet auto
+
+set -euo pipefail
+
+MISSION=${1:-20260505-Lepradet}
+WORKER=${2:-auto}
+MANIFEST="/home/cosma/cosma-pipeline/${MISSION}/manifest.json"
+PIPELINE_DIR="$(cd "$(dirname "$0")" && pwd)/stages"
+PIPELINE_BASE="/home/cosma/cosma-pipeline"
+NAV_DIR="${PIPELINE_BASE}/data/${MISSION}/nav"
+NAV_FILT_DIR="${PIPELINE_BASE}/data/${MISSION}/nav_filtered"
+FRAMES_DIR="${PIPELINE_BASE}/data/${MISSION}/frames"
+
+RUN_ID="$(date +%Y%m%d_%H%M%S)"
+RUN_LOG_DIR="${PIPELINE_BASE}/runs/${RUN_ID}"
+mkdir -p "${RUN_LOG_DIR}"
+
+echo "=== Pipeline run ${RUN_ID} mission=${MISSION} worker=${WORKER} ===" | tee "${RUN_LOG_DIR}/run.log"
+echo "Start: $(date -u +%Y-%m-%dT%H:%M:%SZ)" | tee -a "${RUN_LOG_DIR}/run.log"
+
+# Stage 02: nav parse
+echo "" | tee -a "${RUN_LOG_DIR}/run.log"
+echo "--- Stage 02: nav parse ---" | tee -a "${RUN_LOG_DIR}/run.log"
+python3 "${PIPELINE_DIR}/02_nav_parse.py" "${MANIFEST}" \
+    2>&1 | tee -a "${RUN_LOG_DIR}/stage02.log" "${RUN_LOG_DIR}/run.log"
+
+# Stage 03: nav filter
+echo "" | tee -a "${RUN_LOG_DIR}/run.log"
+echo "--- Stage 03: nav filter ---" | tee -a "${RUN_LOG_DIR}/run.log"
+python3 "${PIPELINE_DIR}/03_nav_filter.py" "${NAV_DIR}" \
+    2>&1 | tee -a "${RUN_LOG_DIR}/stage03.log" "${RUN_LOG_DIR}/run.log"
+
+# QC threshold: lowered from 50 to 30 (avg bottom_visible=37.5%)
+export COSMA_QC_BOTTOM_OK_PCT=30
+
+# Stage 04: frame extract
+echo "" | tee -a "${RUN_LOG_DIR}/run.log"
+echo "--- Stage 04: frame extract ---" | tee -a "${RUN_LOG_DIR}/run.log"
+python3 "${PIPELINE_DIR}/04_frame_extract.py" --mission "${MISSION}" \
+    2>&1 | tee -a "${RUN_LOG_DIR}/stage04.log" "${RUN_LOG_DIR}/run.log"
+
+# Stage 05: inference (sequential, one segment at a time)
+echo "" | tee -a "${RUN_LOG_DIR}/run.log"
+echo "--- Stage 05: inference ---" | tee -a "${RUN_LOG_DIR}/run.log"
+python3 "${PIPELINE_DIR}/05_inference.py" \
+    --frames-dir "${FRAMES_DIR}" \
+    --worker "${WORKER}" \
+    --mission "${MISSION}" \
+    2>&1 | tee -a "${RUN_LOG_DIR}/stage05.log" "${RUN_LOG_DIR}/run.log"
+
+echo "" | tee -a "${RUN_LOG_DIR}/run.log"
+echo "=== Pipeline DONE $(date -u +%Y-%m-%dT%H:%M:%SZ) ===" | tee -a "${RUN_LOG_DIR}/run.log"
+echo "Logs: ${RUN_LOG_DIR}/"
--- a/pipeline/stages/01_ingest.py
+++ b/pipeline/stages/01_ingest.py
@@ -0,0 +1,381 @@
+#!/usr/bin/env python3
+"""Stage 01 — Ingest mission folder from SSD.
+
+Scans /mnt/ssd/<mission>/raw_data/ and builds a manifest:
+- Videos per AUV+GoPro segment (from medias/videos/)
+- USBL/bag sessions per AUV (from logs/SUB/bag/*.mcap)
+- Auto-detects AUV ID mapping (AUV010↔AUV210 etc.) by timestamp proximity
+
+Usage:
+    python3 01_ingest.py /mnt/ssd/20260505-Lepradet --name 20260505-Lepradet
+    python3 01_ingest.py /mnt/ssd/20260505-Lepradet --name 20260505-Lepradet --out /home/cosma/cosma-pipeline
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+import re
+import subprocess
+import sys
+from datetime import datetime, timedelta
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+from orchestrator.db import init_db, get_conn, upsert_mission, now_iso
+
+LOG_DIR = Path(os.environ.get("COSMA_PIPELINE_LOGS", "/home/cosma/cosma-pipeline/logs"))
+LOG_DIR.mkdir(parents=True, exist_ok=True)
+
+# AUV ID normalization: GP folder uses AUV2xx, bag files use AUV0xx
+# e.g. AUV210 <-> AUV010, AUV213 <-> AUV013, AUV212 <-> AUV012
+AUV_FOLDER_RE = re.compile(r"GP(\d+)[_-]AUV(\d+)", re.I)
+AUV_BAG_RE    = re.compile(r"auv(\d+)", re.I)
+AUV_CSV_RE    = re.compile(r"AUV(\d+)", re.I)
+
+
+def normalize_auv(raw_id: str) -> str:
+    """Normalize AUV IDs: strip leading zeros or map 0xx -> 2xx heuristic.
+    Returns canonical form like AUV010, AUV013, AUV210, AUV212 as-is.
+    We keep original for now and detect mapping via timestamp cross-correlation.
+    """
+    m = re.search(r"\d+", raw_id)
+    if not m:
+        return raw_id.upper()
+    n = int(m.group())
+    return f"AUV{n:03d}"
+
+
+def exif_create_date(path: Path) -> datetime | None:
+    try:
+        out = subprocess.check_output(
+            ["exiftool", "-s3", "-CreateDate", "-api", "QuickTimeUTC=1", str(path)],
+            stderr=subprocess.DEVNULL, text=True, timeout=10
+        ).strip()
+        if not out:
+            return None
+        out = re.sub(r'[+-]\d{2}:\d{2}$', '', out).strip()
+        return datetime.strptime(out, "%Y:%m:%d %H:%M:%S")
+    except Exception:
+        return None
+
+
+def exif_duration_s(path: Path) -> float | None:
+    try:
+        out = subprocess.check_output(
+            ["exiftool", "-s3", "-Duration#", str(path)],
+            stderr=subprocess.DEVNULL, text=True, timeout=10
+        ).strip()
+        return float(out) if out else None
+    except Exception:
+        return None
+
+
+def scan_videos(raw_data: Path) -> dict[str, list[dict]]:
+    """Scan medias/videos/ and return dict {auv_id: [video_info, ...]}.
+    Handles both GP1-AUV210 and GP1_AUV210 naming conventions.
+    """
+    videos_dir = raw_data / "medias" / "videos"
+    if not videos_dir.exists():
+        return {}
+
+    result: dict[str, list[dict]] = {}
+    for folder in sorted(videos_dir.iterdir()):
+        if not folder.is_dir():
+            continue
+        m = AUV_FOLDER_RE.search(folder.name)
+        if not m:
+            continue
+        gopro_n = int(m.group(1))
+        auv_id = normalize_auv(m.group(2))
+
+        mp4_files = sorted(folder.glob("*.MP4")) + sorted(folder.glob("*.mp4"))
+        for mp4 in mp4_files:
+            create_date = exif_create_date(mp4)
+            duration = exif_duration_s(mp4)
+            info = {
+                "path": str(mp4),
+                "gopro": gopro_n,
+                "auv_id": auv_id,
+                "filename": mp4.name,
+                "create_date": create_date.isoformat() if create_date else None,
+                "duration_s": duration,
+                "size_mb": round(mp4.stat().st_size / 1e6, 1),
+            }
+            result.setdefault(auv_id, []).append(info)
+
+    return result
+
+
+def scan_bags(raw_data: Path) -> dict[str, list[dict]]:
+    """Scan logs/SUB/bag/ for MCAP files grouped by session+AUV."""
+    bag_dir = raw_data / "logs" / "SUB" / "bag"
+    if not bag_dir.exists():
+        return {}
+
+    result: dict[str, list[dict]] = {}
+    for session_dir in sorted(bag_dir.iterdir()):
+        if not session_dir.is_dir():
+            continue
+        # dir name: 20260505_074718_AUV013
+        m = re.match(r"(\d{8}_\d{6})_AUV(\d+)", session_dir.name)
+        if not m:
+            continue
+        ts_str = m.group(1)
+        auv_id = normalize_auv(m.group(2))
+        try:
+            ts = datetime.strptime(ts_str, "%Y%m%d_%H%M%S")
+        except ValueError:
+            ts = None
+
+        mcap_files = sorted(session_dir.glob("*.mcap"))
+        total_size = sum(f.stat().st_size for f in mcap_files)
+        non_empty = [str(f) for f in mcap_files if f.stat().st_size > 0]
+
+        if not non_empty:
+            continue
+
+        session_info = {
+            "session": session_dir.name,
+            "auv_id": auv_id,
+            "timestamp": ts.isoformat() if ts else None,
+            "mcap_files": non_empty,
+            "total_mb": round(total_size / 1e6, 1),
+        }
+        result.setdefault(auv_id, []).append(session_info)
+
+    return result
+
+
+def scan_usbl_csv(raw_data: Path) -> dict[str, list[dict]]:
+    """Scan logs/SUB/log/ for *_usbl.csv files.
+    Note: these are raw serial byte logs, not lat/lon CSV.
+    We record them for reference; actual nav comes from MCAP bags.
+    """
+    log_dir = raw_data / "logs" / "SUB" / "log"
+    if not log_dir.exists():
+        return {}
+
+    result: dict[str, list[dict]] = {}
+    for f in sorted(log_dir.glob("*_usbl.csv")):
+        m = AUV_CSV_RE.search(f.name)
+        if not m:
+            continue
+        auv_id = normalize_auv(m.group(1))
+        # parse timestamp from filename: 2026-05-05_08-16-00_AUV010_usbl.csv
+        ts_m = re.match(r"(\d{4}-\d{2}-\d{2}_\d{2}-\d{2}-\d{2})", f.name)
+        ts = None
+        if ts_m:
+            try:
+                ts = datetime.strptime(ts_m.group(1), "%Y-%m-%d_%H-%M-%S")
+            except ValueError:
+                pass
+        result.setdefault(auv_id, []).append({
+            "path": str(f),
+            "auv_id": auv_id,
+            "timestamp": ts.isoformat() if ts else None,
+            "size_kb": round(f.stat().st_size / 1e3, 1),
+            "format": "raw_serial",  # NOT lat/lon CSV
+        })
+
+    return result
+
+
+def detect_auv_mapping(videos: dict, bags: dict) -> dict[str, str]:
+    """Detect AUV ID mapping between video folders (AUV2xx) and bag sessions (AUV0xx).
+
+    Heuristic: if video AUV2xx and bag AUV0xx share same last 2 digits and
+    timestamps are within 60s → they are the same physical AUV.
+
+    Returns: {video_auv_id: bag_auv_id} e.g. {"AUV210": "AUV010"}
+    """
+    mapping: dict[str, str] = {}
+
+    for vid_auv in videos:
+        vid_ts_list = []
+        for v in videos[vid_auv]:
+            if v.get("create_date"):
+                try:
+                    vid_ts_list.append(datetime.fromisoformat(v["create_date"]))
+                except Exception:
+                    pass
+        if not vid_ts_list:
+            continue
+
+        vid_digits = vid_auv[-2:]  # last 2 digits of AUV2xx
+
+        best_bag_auv = None
+        best_delta = timedelta(seconds=999)
+
+        for bag_auv in bags:
+            # check same last 2 digits
+            if bag_auv[-2:] != vid_digits:
+                continue
+            for sess in bags[bag_auv]:
+                if sess.get("timestamp"):
+                    try:
+                        bag_ts = datetime.fromisoformat(sess["timestamp"])
+                    except Exception:
+                        continue
+                    for vt in vid_ts_list:
+                        delta = abs(vt - bag_ts)
+                        if delta < best_delta:
+                            best_delta = delta
+                            best_bag_auv = bag_auv
+
+        if best_bag_auv and best_delta < timedelta(seconds=3600):
+            mapping[vid_auv] = best_bag_auv
+
+    return mapping
+
+
+def group_video_segments(video_list: list[dict], gap_min: int = 5) -> list[dict]:
+    """Group consecutive videos into segments by timestamp gap."""
+    sorted_vids = sorted(
+        [v for v in video_list if v.get("create_date")],
+        key=lambda x: x["create_date"]
+    )
+    if not sorted_vids:
+        return [{"videos": video_list, "label": "seg_unknown", "total_s": None}]
+
+    segments = []
+    current: list[dict] = [sorted_vids[0]]
+
+    for vid in sorted_vids[1:]:
+        prev = current[-1]
+        try:
+            prev_end = datetime.fromisoformat(prev["create_date"])
+            if prev.get("duration_s"):
+                prev_end += timedelta(seconds=prev["duration_s"])
+            cur_start = datetime.fromisoformat(vid["create_date"])
+            gap = (cur_start - prev_end).total_seconds() / 60
+            if gap > gap_min:
+                segments.append(_finalize_segment(current))
+                current = [vid]
+            else:
+                current.append(vid)
+        except Exception:
+            current.append(vid)
+
+    if current:
+        segments.append(_finalize_segment(current))
+
+    return segments
+
+
+def _finalize_segment(videos: list[dict]) -> dict:
+    label = videos[0]["create_date"][:19].replace(":", "-").replace("T", "_") if videos[0].get("create_date") else "seg_unknown"
+    total_s = sum(v["duration_s"] or 0 for v in videos)
+    return {
+        "label": label,
+        "videos": videos,
+        "total_s": total_s,
+        "start": videos[0].get("create_date"),
+        "end": videos[-1].get("create_date"),
+    }
+
+
+def build_manifest(mission_path: Path, gap_min: int = 5) -> dict:
+    raw_data = mission_path / "raw_data"
+    if not raw_data.exists():
+        # try direct
+        raw_data = mission_path
+
+    print(f"[01_ingest] scanning {raw_data} ...")
+
+    videos = scan_videos(raw_data)
+    bags   = scan_bags(raw_data)
+    csvs   = scan_usbl_csv(raw_data)
+    auv_map = detect_auv_mapping(videos, bags)
+
+    # Build per-AUV segments
+    auv_segments: dict[str, list[dict]] = {}
+    for auv_id, vid_list in videos.items():
+        segs = group_video_segments(vid_list, gap_min=gap_min)
+        auv_segments[auv_id] = segs
+
+    # Compute AUVs with real data
+    auv_ids_with_video = sorted(videos.keys())
+    auv_ids_with_bags  = sorted(bags.keys())
+
+    total_video_s = sum(
+        seg["total_s"] or 0
+        for segs in auv_segments.values()
+        for seg in segs
+    )
+
+    manifest = {
+        "mission": mission_path.name,
+        "ssd_path": str(mission_path),
+        "generated_at": now_iso(),
+        "auv_ids_video": auv_ids_with_video,
+        "auv_ids_bags": auv_ids_with_bags,
+        "auv_mapping": auv_map,
+        "total_video_s": round(total_video_s),
+        "segments_per_auv": auv_segments,
+        "bag_sessions_per_auv": bags,
+        "usbl_csv_per_auv": csvs,
+        "notes": {
+            "usbl_csv_format": "raw_serial_bytes",
+            "nav_source": "mcap_bags",
+        },
+    }
+
+    return manifest
+
+
+def ingest(mission_path: Path, mission_name: str, out_dir: Path,
+           gap_min: int = 5) -> dict:
+    out_dir.mkdir(parents=True, exist_ok=True)
+    manifest_path = out_dir / mission_name / "manifest.json"
+
+    # Idempotency check
+    if manifest_path.exists():
+        existing = json.loads(manifest_path.read_text())
+        chk = hashlib.sha256(mission_path.name.encode()).hexdigest()[:8]
+        print(f"[01_ingest] manifest exists (checksum {chk}), skipping scan")
+        return existing
+
+    manifest = build_manifest(mission_path, gap_min=gap_min)
+
+    # Save manifest
+    manifest_path.parent.mkdir(parents=True, exist_ok=True)
+    manifest_path.write_text(json.dumps(manifest, indent=2))
+    print(f"[01_ingest] manifest saved: {manifest_path}")
+
+    # Write to DB
+    init_db()
+    with get_conn() as conn:
+        upsert_mission(conn, mission_name, str(mission_path),
+                       status="ingested", manifest=json.dumps(manifest))
+
+    return manifest
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 01 — Ingest mission from SSD")
+    ap.add_argument("mission_path", type=Path, help="Path to mission folder (e.g. /mnt/ssd/20260505-Lepradet)")
+    ap.add_argument("--name", type=str, default=None, help="Mission name (defaults to folder name)")
+    ap.add_argument("--out", type=Path, default=Path("/home/cosma/cosma-pipeline"), help="Output base dir")
+    ap.add_argument("--gap-min", type=int, default=5, help="Gap in minutes to split video segments")
+    args = ap.parse_args()
+
+    mission_name = args.name or args.mission_path.name
+    manifest = ingest(args.mission_path, mission_name, args.out, args.gap_min)
+
+    print(f"\n=== Ingest summary for {mission_name} ===")
+    print(f"AUVs with video: {manifest['auv_ids_video']}")
+    print(f"AUVs with bags:  {manifest['auv_ids_bags']}")
+    print(f"AUV mapping:     {manifest['auv_mapping']}")
+    print(f"Total video:     {manifest['total_video_s']}s")
+    print(f"Segments:")
+    for auv, segs in manifest["segments_per_auv"].items():
+        for seg in segs:
+            n_vids = len(seg["videos"])
+            dur = f"{seg['total_s']:.0f}s" if seg["total_s"] else "?"
+            print(f"  {auv} / {seg['label']}  {n_vids} videos  {dur}")
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/02_nav_parse.py
+++ b/pipeline/stages/02_nav_parse.py
@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+"""Stage 02 — Parse navigation from ROS2 MCAP bag files.
+
+Extracts per-AUV trajectories from MCAP bags using mcap_ros2:
+- /mavros/global_position/global  → NavSatFix (lat, lon, alt)
+- /mavros/imu/data               → Imu (qx, qy, qz, qw)
+- /mavros/imu/static_pressure    → FluidPressure (pressure_pa)
+
+Joins on nearest timestamp (tolerance 100ms).
+Saves parquet: ~/cosma-pipeline/data/<mission>/nav/<AUV>_<segment>.parquet
+
+Fallback: if no MCAP GPS data, marks as degraded=True (GPS=0 under water is normal).
+
+Usage:
+    python3 02_nav_parse.py /home/cosma/cosma-pipeline/20260505-Lepradet/manifest.json
+    python3 02_nav_parse.py /path/manifest.json --auv AUV013
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
+
+PIPELINE_BASE = Path(os.environ.get("COSMA_PIPELINE_BASE", "/home/cosma/cosma-pipeline"))
+NAV_TOPICS = [
+    "/mavros/global_position/global",
+    "/mavros/imu/data",
+    "/mavros/imu/static_pressure",
+]
+
+
+def parse_mcap_segment(mcap_files: list[Path]) -> dict[str, list]:
+    """Extract raw topic data from a list of MCAP files (one session/segment).
+    Returns dict keyed by topic -> list of (ts_ns, data_dict).
+    """
+    from mcap_ros2.reader import read_ros2_messages
+
+    topic_data: dict[str, list] = {t: [] for t in NAV_TOPICS}
+
+    for mcap_path in mcap_files:
+        if not mcap_path.exists():
+            continue
+        try:
+            for msg in read_ros2_messages(str(mcap_path), topics=NAV_TOPICS):
+                topic = msg.channel.topic
+                m = msg.ros_msg
+                ts_ns = int(msg.log_time.timestamp() * 1e9)
+
+                if topic == "/mavros/global_position/global":
+                    topic_data[topic].append((ts_ns, {
+                        "lat": float(m.latitude),
+                        "lon": float(m.longitude),
+                        "alt": float(m.altitude),
+                    }))
+                elif topic == "/mavros/imu/data":
+                    topic_data[topic].append((ts_ns, {
+                        "qx": float(m.orientation.x),
+                        "qy": float(m.orientation.y),
+                        "qz": float(m.orientation.z),
+                        "qw": float(m.orientation.w),
+                    }))
+                elif topic == "/mavros/imu/static_pressure":
+                    topic_data[topic].append((ts_ns, {
+                        "pressure_pa": float(m.fluid_pressure),
+                    }))
+        except Exception as e:
+            print(f"  [02] Error reading {mcap_path.name}: {e}")
+
+    return topic_data
+
+
+def join_topics(topic_data: dict[str, list], tol_ns: int = 100_000_000) -> list[dict]:
+    """Join NavSatFix + Imu + FluidPressure on nearest timestamp (100ms tol).
+    Base timeline = NavSatFix if available, else Imu.
+    """
+    import pandas as pd
+
+    nav_pts = topic_data.get("/mavros/global_position/global", [])
+    imu_pts = topic_data.get("/mavros/imu/data", [])
+    pres_pts = topic_data.get("/mavros/imu/static_pressure", [])
+
+    if not nav_pts and not imu_pts:
+        return []
+
+    # Build DataFrames
+    if nav_pts:
+        df_nav = pd.DataFrame([{"ts_ns": ts, **d} for ts, d in nav_pts])
+    else:
+        df_nav = pd.DataFrame(columns=["ts_ns", "lat", "lon", "alt"])
+
+    if imu_pts:
+        df_imu = pd.DataFrame([{"ts_ns": ts, **d} for ts, d in imu_pts])
+    else:
+        df_imu = pd.DataFrame(columns=["ts_ns", "qx", "qy", "qz", "qw"])
+
+    if pres_pts:
+        df_pres = pd.DataFrame([{"ts_ns": ts, **d} for ts, d in pres_pts])
+    else:
+        df_pres = pd.DataFrame(columns=["ts_ns", "pressure_pa"])
+
+    # Use nav as base if it has data, else imu
+    base_df = df_nav if len(df_nav) > 0 else df_imu
+    base_df = base_df.sort_values("ts_ns").reset_index(drop=True)
+
+    # Merge-as-of for IMU
+    result = base_df.copy()
+    if len(df_imu) > 0:
+        df_imu_s = df_imu.sort_values("ts_ns").reset_index(drop=True)
+        # Simple nearest-neighbor join
+        imu_ts = df_imu_s["ts_ns"].values
+        for col in ["qx", "qy", "qz", "qw"]:
+            result[col] = np.nan
+        for i, row_ts in enumerate(result["ts_ns"].values):
+            idx = np.argmin(np.abs(imu_ts - row_ts))
+            if abs(imu_ts[idx] - row_ts) <= tol_ns:
+                for col in ["qx", "qy", "qz", "qw"]:
+                    result.at[i, col] = float(df_imu_s.at[idx, col])
+
+    # Merge pressure
+    if len(df_pres) > 0:
+        df_pres_s = df_pres.sort_values("ts_ns").reset_index(drop=True)
+        pres_ts = df_pres_s["ts_ns"].values
+        result["pressure_pa"] = np.nan
+        for i, row_ts in enumerate(result["ts_ns"].values):
+            idx = np.argmin(np.abs(pres_ts - row_ts))
+            if abs(pres_ts[idx] - row_ts) <= tol_ns:
+                result.at[i, "pressure_pa"] = float(df_pres_s.at[idx, "pressure_pa"])
+
+    # Ensure all columns exist
+    for col in ["lat", "lon", "alt", "qx", "qy", "qz", "qw", "pressure_pa"]:
+        if col not in result.columns:
+            result[col] = np.nan
+
+    return result.to_dict("records")
+
+
+def parse_auv(manifest: dict, auv_id: str, out_dir: Path) -> dict:
+    """Parse all MCAP sessions for one AUV. Returns metrics."""
+    from pathlib import Path as P
+
+    metrics = {
+        "auv_id": auv_id,
+        "segments": [],
+        "total_points": 0,
+        "degraded": False,
+        "status": "ok",
+    }
+
+    bag_sessions = manifest.get("bag_sessions_per_auv", {}).get(auv_id, [])
+    if not bag_sessions:
+        auv_map = manifest.get("auv_mapping", {})
+        bag_auv = auv_map.get(auv_id)
+        if bag_auv:
+            bag_sessions = manifest.get("bag_sessions_per_auv", {}).get(bag_auv, [])
+
+    if not bag_sessions:
+        # Build from raw SSD structure
+        ssd_path = P(manifest.get("ssd_path", "/mnt/ssd") + "/" + manifest["mission"].split("-")[0] + "-" + manifest["mission"].split("-")[1] if "-" in manifest["mission"] else manifest.get("ssd_path", "/mnt/ssd"))
+        auv_num = auv_id.replace("AUV", "0")  # AUV013 -> 0013? No: AUV013 -> AUV013
+        bag_root = P(manifest.get("ssd_path", "/mnt/ssd")) / "raw_data/logs/SUB/bag"
+        sessions = sorted(bag_root.glob(f"*_{auv_id}"))
+        bag_sessions = [{"label": s.name, "mcap_files": [str(f) for f in sorted(s.glob("*.mcap"))]} for s in sessions]
+
+    import pandas as pd
+
+    all_points_total = 0
+    for sess in bag_sessions:
+        label = sess.get("session", sess.get("label", "unknown"))
+        mcap_files = [P(f) for f in sess.get("mcap_files", [])]
+        if not mcap_files:
+            continue
+
+        out_parquet = out_dir / f"{auv_id}_{label}.parquet"
+        if out_parquet.exists():
+            df_ex = pd.read_parquet(out_parquet)
+            n = len(df_ex)
+            print(f"  [02] {auv_id}/{label}: cached ({n} pts)")
+            all_points_total += n
+            metrics["segments"].append({"label": label, "points": n, "cached": True})
+            continue
+
+        print(f"  [02] {auv_id}/{label}: parsing {len(mcap_files)} MCAP files...")
+        topic_data = parse_mcap_segment(mcap_files)
+        points = join_topics(topic_data)
+
+        if not points:
+            print(f"  [02] {auv_id}/{label}: no data")
+            metrics["segments"].append({"label": label, "points": 0, "degraded": True})
+            metrics["degraded"] = True
+            continue
+
+        df = pd.DataFrame(points)
+        n = len(df)
+        # Check GPS quality
+        has_gps = df["lat"].notna().any() and (df["lat"] != 0).any()
+        if not has_gps:
+            print(f"  [02] {auv_id}/{label}: {n} pts, GPS=0 (degraded — AUV underwater)")
+            metrics["degraded"] = True
+        else:
+            print(f"  [02] {auv_id}/{label}: {n} pts, GPS OK")
+
+        df.to_parquet(out_parquet, index=False)
+        all_points_total += n
+        metrics["segments"].append({"label": label, "points": n, "degraded": not has_gps})
+
+    metrics["total_points"] = all_points_total
+    if all_points_total == 0:
+        metrics["status"] = "degraded"
+    return metrics
+
+
+def parse_mission(manifest_path: Path, auv_filter: str | None = None) -> list[dict]:
+    manifest = json.loads(manifest_path.read_text())
+    mission_name = manifest["mission"]
+
+    out_dir = PIPELINE_BASE / "data" / mission_name / "nav"
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    auv_ids = list(set(
+        manifest.get("auv_ids_bags", []) +
+        list(manifest.get("auv_mapping", {}).keys())
+    ))
+    if not auv_ids:
+        auv_ids = manifest.get("auv_ids_video", [])
+    if auv_filter:
+        auv_ids = [a for a in auv_ids if a == auv_filter]
+
+    all_metrics = []
+    init_db()
+
+    for auv_id in sorted(auv_ids):
+        print(f"[02] === {auv_id} ===")
+        m = parse_auv(manifest, auv_id, out_dir)
+        all_metrics.append(m)
+
+        with get_conn() as conn:
+            mission_row = conn.execute("SELECT id FROM missions WHERE name=?", (mission_name,)).fetchone()
+            if mission_row:
+                job_id = upsert_job(conn, mission_row["id"], auv_id, "all", "02_nav_parse",
+                                    status="done" if m["status"] == "ok" else m["status"],
+                                    output_path=str(out_dir))
+                record_metric(conn, job_id, "nav_points_total", value=m["total_points"],
+                              pass_fail="pass" if m["total_points"] > 0 else "warn")
+
+    return all_metrics
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 02 — Parse nav from MCAP bags")
+    ap.add_argument("manifest", type=Path)
+    ap.add_argument("--auv", type=str, default=None)
+    args = ap.parse_args()
+
+    metrics = parse_mission(args.manifest, auv_filter=args.auv)
+
+    print("\n=== Stage 02 summary ===")
+    for m in metrics:
+        segs = m.get("segments", [])
+        total = m.get("total_points", 0)
+        deg = "DEGRADED" if m.get("degraded") else "OK"
+        print(f"  {m['auv_id']}: {total} pts across {len(segs)} segments [{deg}]")
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/02_usbl_parse.py
+++ b/pipeline/stages/02_usbl_parse.py
@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+"""Stage 02 — Parse USBL/navigation from MCAP bag files.
+
+The USBL CSV logs in logs/SUB/log/ contain raw serial bytes, NOT lat/lon.
+Real navigation data is in MCAP bags (logs/SUB/bag/).
+
+This stage:
+1. Reads MCAP files per AUV session
+2. Extracts position topics (configurable in default_params.yaml)
+3. Falls back to parsing serial bytes if no nav topic found (best-effort)
+4. Outputs Parquet per AUV with columns: timestamp, lat, lon, depth, heading
+
+Usage:
+    python3 02_usbl_parse.py /home/cosma/cosma-pipeline/20260505-Lepradet/manifest.json
+    python3 02_usbl_parse.py /home/cosma/cosma-pipeline/20260505-Lepradet/manifest.json --auv AUV010
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+import numpy as np
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
+
+LOG_DIR = Path(os.environ.get("COSMA_PIPELINE_LOGS", "/home/cosma/cosma-pipeline/logs"))
+LOG_DIR.mkdir(parents=True, exist_ok=True)
+
+# Known nav topics to try in MCAP files
+NAV_TOPICS = [
+    "/usbl/position",
+    "/usbl/fix",
+    "/navigation/position",
+    "/bluerov/usbl",
+    "/waterlinked/position",
+    "/fix",
+    "/gps/fix",
+    "/mavros/global_position/global",
+    "/mavros/global_position/local",
+]
+
+
+def try_parse_mcap(mcap_path: Path, topics: list[str] | None = None) -> list[dict]:
+    """Try to extract nav points from MCAP file. Returns list of {ts, lat, lon, depth}."""
+    try:
+        from mcap.reader import make_reader
+    except ImportError:
+        print(f"  [02] mcap not installed, skipping {mcap_path.name}")
+        return []
+
+    points = []
+    try:
+        with open(mcap_path, "rb") as f:
+            reader = make_reader(f)
+            for schema, channel, message in reader.iter_messages(topics=topics):
+                # Try to deserialize — support ROS2 JSON-encoded or raw
+                try:
+                    import json as _json
+                    data = _json.loads(message.data)
+                    lat = data.get("latitude") or data.get("lat")
+                    lon = data.get("longitude") or data.get("lon")
+                    depth = data.get("depth") or data.get("altitude")
+                    heading = data.get("heading") or data.get("yaw")
+                    if lat is not None and lon is not None:
+                        ts_ns = message.log_time
+                        ts = ts_ns / 1e9
+                        points.append({
+                            "timestamp": ts,
+                            "lat": float(lat),
+                            "lon": float(lon),
+                            "depth": float(depth) if depth is not None else None,
+                            "heading": float(heading) if heading is not None else None,
+                            "source": channel.topic,
+                        })
+                except Exception:
+                    pass
+    except Exception as e:
+        print(f"  [02] MCAP read error {mcap_path.name}: {e}")
+
+    return points
+
+
+def try_parse_serial_csv(csv_path: Path) -> list[dict]:
+    """Best-effort: parse raw serial byte log for USBL range/bearing frames.
+    Waterlinked USBL M64 protocol: 0xBB 0x55 frame header.
+    This is a rough attempt — actual lat/lon requires ship GPS + range+bearing.
+    Returns relative-only positions (range, bearing) if decoded.
+    """
+    points = []
+    try:
+        with open(csv_path, "r") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                parts = line.split(",", 2)
+                if len(parts) < 3:
+                    continue
+                ts_str, direction, raw = parts[0], parts[1], parts[2]
+                if direction.strip() != "RECEIVED":
+                    continue
+                # Extract bytes from repr string b'\xbb...'
+                try:
+                    raw_clean = raw.strip().strip('"')
+                    # Parse Python bytes repr
+                    data = eval(raw_clean)  # safe: only used on known CSV
+                    if len(data) >= 4 and data[0] == 0xBB and data[1] == 0x55:
+                        # Waterlinked M64 position frame: len byte at [2], payload follows
+                        payload_len = data[2]
+                        if len(data) >= payload_len + 3:
+                            # Best effort: look for float32 values in payload
+                            # Actual protocol decoding would need WL M64 spec
+                            ts = datetime.fromisoformat(ts_str)
+                            points.append({
+                                "timestamp": ts.timestamp(),
+                                "lat": None,
+                                "lon": None,
+                                "depth": None,
+                                "heading": None,
+                                "source": "serial_raw",
+                                "raw_bytes": data.hex()[:32],
+                            })
+                except Exception:
+                    pass
+    except Exception as e:
+        print(f"  [02] serial CSV parse error {csv_path.name}: {e}")
+
+    return points
+
+
+def parse_auv_sessions(manifest: dict, auv_id: str, out_dir: Path,
+                       topics: list[str] | None = None) -> dict:
+    """Parse all sessions for one AUV. Returns metrics dict."""
+    metrics = {"auv_id": auv_id, "points_raw": 0, "sources": [], "status": "ok"}
+    all_points: list[dict] = []
+
+    # Try MCAP bags first
+    bag_sessions = manifest.get("bag_sessions_per_auv", {}).get(auv_id, [])
+    if not bag_sessions:
+        # Try mapping: maybe bags use AUV0xx while videos use AUV2xx
+        auv_map = manifest.get("auv_mapping", {})
+        bag_auv = auv_map.get(auv_id)
+        if bag_auv:
+            bag_sessions = manifest.get("bag_sessions_per_auv", {}).get(bag_auv, [])
+
+    for sess in bag_sessions:
+        for mcap_path_str in sess.get("mcap_files", []):
+            mcap_path = Path(mcap_path_str)
+            if not mcap_path.exists():
+                continue
+            pts = try_parse_mcap(mcap_path, topics=topics or NAV_TOPICS)
+            if pts:
+                all_points.extend(pts)
+                if "mcap" not in metrics["sources"]:
+                    metrics["sources"].append("mcap")
+                print(f"  [02] {auv_id} {mcap_path.name}: {len(pts)} nav points")
+
+    # Fallback: serial CSV if no MCAP nav
+    if not all_points:
+        csv_entries = manifest.get("usbl_csv_per_auv", {}).get(auv_id, [])
+        for entry in csv_entries:
+            csv_path = Path(entry["path"])
+            if not csv_path.exists():
+                continue
+            pts = try_parse_serial_csv(csv_path)
+            if pts:
+                all_points.extend(pts)
+                if "serial_csv" not in metrics["sources"]:
+                    metrics["sources"].append("serial_csv")
+
+    metrics["points_raw"] = len(all_points)
+
+    if not all_points:
+        metrics["status"] = "degraded"
+        metrics["note"] = "no nav points found in MCAP or serial CSV"
+        print(f"  [02] {auv_id}: NO nav data found — degraded")
+    else:
+        print(f"  [02] {auv_id}: {len(all_points)} raw nav points from {metrics['sources']}")
+
+    # Save output
+    out_file = out_dir / f"{auv_id}_nav_raw.json"
+    out_file.write_text(json.dumps({
+        "auv_id": auv_id,
+        "generated_at": now_iso(),
+        "metrics": metrics,
+        "points": all_points,
+    }, indent=2, default=str))
+
+    # Also save as simple CSV for downstream
+    csv_out = out_dir / f"{auv_id}_nav_raw.csv"
+    with open(csv_out, "w") as f:
+        f.write("timestamp,lat,lon,depth,heading,source\n")
+        for p in all_points:
+            f.write(f"{p['timestamp']},{p.get('lat','')},{p.get('lon','')},{p.get('depth','')},{p.get('heading','')},{p.get('source','')}\n")
+
+    return metrics
+
+
+def parse_mission(manifest_path: Path, auv_filter: str | None = None,
+                  out_dir: Path | None = None) -> list[dict]:
+    manifest = json.loads(manifest_path.read_text())
+    mission_name = manifest["mission"]
+
+    if out_dir is None:
+        out_dir = manifest_path.parent / "02_usbl_raw"
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    # Idempotency: check if all AUV outputs exist
+    auv_ids = manifest.get("auv_ids_bags", []) or manifest.get("auv_ids_video", [])
+    if auv_filter:
+        auv_ids = [a for a in auv_ids if a == auv_filter]
+
+    all_metrics = []
+    init_db()
+
+    for auv_id in auv_ids:
+        out_file = out_dir / f"{auv_id}_nav_raw.json"
+        if out_file.exists():
+            print(f"[02] {auv_id}: output exists, skipping")
+            existing = json.loads(out_file.read_text())
+            all_metrics.append(existing.get("metrics", {"auv_id": auv_id, "status": "cached"}))
+            continue
+
+        print(f"[02] Parsing {auv_id} ...")
+        m = parse_auv_sessions(manifest, auv_id, out_dir)
+        all_metrics.append(m)
+
+        # Record in DB
+        with get_conn() as conn:
+            from orchestrator.db import upsert_mission
+            mission_id_row = conn.execute(
+                "SELECT id FROM missions WHERE name=?", (mission_name,)
+            ).fetchone()
+            if mission_id_row:
+                mission_id = mission_id_row["id"]
+                job_id = upsert_job(conn, mission_id, auv_id, "all", "02_usbl_parse",
+                                    status="done" if m["status"] == "ok" else m["status"],
+                                    output_path=str(out_dir))
+                record_metric(conn, job_id, "usbl_points_raw",
+                              value=m["points_raw"],
+                              pass_fail="pass" if m["points_raw"] > 0 else "fail")
+
+    return all_metrics
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 02 — Parse USBL/nav from MCAP bags")
+    ap.add_argument("manifest", type=Path, help="manifest.json from stage 01")
+    ap.add_argument("--auv", type=str, default=None, help="Filter to single AUV ID")
+    ap.add_argument("--out", type=Path, default=None, help="Output directory")
+    args = ap.parse_args()
+
+    metrics = parse_mission(args.manifest, auv_filter=args.auv, out_dir=args.out)
+
+    print("\n=== Stage 02 summary ===")
+    total_pts = sum(m.get("points_raw", 0) for m in metrics)
+    for m in metrics:
+        status = m.get("status", "?")
+        pts = m.get("points_raw", 0)
+        src = m.get("sources", [])
+        print(f"  {m['auv_id']}: {pts} pts  {src}  [{status}]")
+    print(f"Total nav points: {total_pts}")
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/03_nav_filter.py
+++ b/pipeline/stages/03_nav_filter.py
@@ -0,0 +1,185 @@
+#!/usr/bin/env python3
+"""Stage 03 — Filter and smooth navigation trajectories.
+
+Input:  ~/cosma-pipeline/data/<mission>/nav/<AUV>_<segment>.parquet
+Output: ~/cosma-pipeline/data/<mission>/nav_filtered/<AUV>_<segment>.parquet
+
+Steps:
+1. Drop rows with null lat/lon OR lat==0 AND lon==0 (no GPS lock)
+2. MAD-3σ outlier removal on lat, lon
+3. Moving average smoothing (window 5s, KISS)
+4. Depth from pressure: depth_m = (pressure_pa - 101325) / (1025 * 9.81)
+5. Output: same columns + lat_smooth, lon_smooth, depth_m
+
+Usage:
+    python3 03_nav_filter.py /home/cosma/cosma-pipeline/data/20260505-Lepradet/nav/
+    python3 03_nav_filter.py /path/nav/ --auv AUV013 --sigma 2.5
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
+
+PIPELINE_BASE = Path(os.environ.get("COSMA_PIPELINE_BASE", "/home/cosma/cosma-pipeline"))
+RHO_SEA = 1025.0  # kg/m3
+G = 9.81          # m/s2
+P_ATM = 101325.0  # Pa
+
+
+def mad_mask(arr: np.ndarray, sigma: float = 3.0) -> np.ndarray:
+    """True = keep."""
+    if len(arr) < 4:
+        return np.ones(len(arr), dtype=bool)
+    med = np.median(arr)
+    mad = np.median(np.abs(arr - med))
+    if mad == 0:
+        return np.ones(len(arr), dtype=bool)
+    return np.abs(0.6745 * (arr - med) / mad) < sigma
+
+
+def moving_average(arr: np.ndarray, window: int = 5) -> np.ndarray:
+    if len(arr) < window:
+        return arr.copy()
+    pad = window // 2
+    padded = np.pad(arr, (pad, pad), mode="edge")
+    return np.convolve(padded, np.ones(window) / window, mode="valid")[:len(arr)]
+
+
+def filter_parquet(src: Path, dst_dir: Path, sigma: float = 3.0, window: int = 5) -> dict:
+    import pandas as pd
+
+    df = pd.read_parquet(src)
+    auv_seg = src.stem
+    metrics = {
+        "file": src.name,
+        "points_in": len(df),
+        "points_out": 0,
+        "status": "ok",
+    }
+
+    # Step 1: drop null/zero GPS
+    has_lat = "lat" in df.columns and df["lat"].notna().any()
+    if has_lat:
+        mask_valid = df["lat"].notna() & df["lon"].notna() & (df["lat"] != 0) & (df["lon"] != 0)
+        df_valid = df[mask_valid].copy()
+    else:
+        # No GPS — keep all rows for IMU/pressure
+        df_valid = df.copy()
+        metrics["degraded"] = True
+
+    if len(df_valid) == 0:
+        metrics["status"] = "degraded"
+        metrics["note"] = "no valid GPS points"
+        print(f"  [03] {auv_seg}: no valid GPS — saving as-is with depth calc only")
+        df_out = df.copy()
+    else:
+        # Step 2: MAD outlier removal on lat/lon
+        if has_lat and len(df_valid) >= 4:
+            lats = df_valid["lat"].values
+            lons = df_valid["lon"].values
+            mask = mad_mask(lats, sigma) & mad_mask(lons, sigma)
+            n_removed = int((~mask).sum())
+            df_valid = df_valid[mask].copy()
+            metrics["points_removed_outlier"] = n_removed
+        else:
+            metrics["points_removed_outlier"] = 0
+
+        # Step 3: sort by timestamp
+        if "ts_ns" in df_valid.columns:
+            df_valid = df_valid.sort_values("ts_ns").reset_index(drop=True)
+
+        # Step 4: smooth lat/lon
+        if has_lat and len(df_valid) >= window:
+            df_valid["lat_smooth"] = moving_average(df_valid["lat"].values, window)
+            df_valid["lon_smooth"] = moving_average(df_valid["lon"].values, window)
+        elif has_lat and len(df_valid) > 0:
+            df_valid["lat_smooth"] = df_valid["lat"]
+            df_valid["lon_smooth"] = df_valid["lon"]
+        else:
+            df_valid["lat_smooth"] = np.nan
+            df_valid["lon_smooth"] = np.nan
+
+        df_out = df_valid
+
+    # Step 5: depth from pressure
+    if "pressure_pa" in df_out.columns and df_out["pressure_pa"].notna().any():
+        df_out["depth_m"] = (df_out["pressure_pa"] - P_ATM) / (RHO_SEA * G)
+        df_out["depth_m"] = df_out["depth_m"].abs()  # negative when underwater (P < Patm)  # surface = 0
+    else:
+        df_out["depth_m"] = np.nan
+
+    dst_dir.mkdir(parents=True, exist_ok=True)
+    out_path = dst_dir / src.name
+    df_out.to_parquet(out_path, index=False)
+
+    metrics["points_out"] = len(df_out)
+    removed_null = metrics["points_in"] - len(df_out) - metrics.get("points_removed_outlier", 0)
+    metrics["points_removed_null"] = max(0, removed_null)
+    print(f"  [03] {auv_seg}: {metrics['points_in']} → {metrics['points_out']} pts, "
+          f"depth_m range=[{df_out['depth_m'].min():.1f}, {df_out['depth_m'].max():.1f}]"
+          if df_out["depth_m"].notna().any() else
+          f"  [03] {auv_seg}: {metrics['points_in']} → {metrics['points_out']} pts, no pressure")
+    return metrics
+
+
+def filter_mission(nav_dir: Path, auv_filter: str | None = None,
+                   sigma: float = 3.0, window: int = 5) -> list[dict]:
+    out_dir = nav_dir.parent / "nav_filtered"
+
+    parquet_files = sorted(nav_dir.glob("*.parquet"))
+    if auv_filter:
+        parquet_files = [f for f in parquet_files if auv_filter in f.name]
+
+    all_metrics = []
+    init_db()
+
+    for pf in parquet_files:
+        out_file = out_dir / pf.name
+        if out_file.exists():
+            print(f"[03] {pf.stem}: cached")
+            continue
+
+        print(f"[03] Filtering {pf.name}...")
+        m = filter_parquet(pf, out_dir, sigma=sigma, window=window)
+        all_metrics.append(m)
+
+        with get_conn() as conn:
+            mission_name = nav_dir.parent.name
+            mission_row = conn.execute("SELECT id FROM missions WHERE name=?", (mission_name,)).fetchone()
+            if mission_row:
+                auv_id = pf.stem.split("_")[0]
+                job_id = upsert_job(conn, mission_row["id"], auv_id, "all", "03_nav_filter",
+                                    status="done" if m.get("status") == "ok" else m.get("status", "done"),
+                                    output_path=str(out_dir))
+                record_metric(conn, job_id, "nav_points_filtered", value=m.get("points_out", 0),
+                              pass_fail="pass" if m.get("points_out", 0) > 0 else "warn")
+
+    return all_metrics
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 03 — Filter nav trajectories")
+    ap.add_argument("nav_dir", type=Path, help="Directory with *.parquet from stage 02")
+    ap.add_argument("--auv", type=str, default=None)
+    ap.add_argument("--sigma", type=float, default=3.0)
+    ap.add_argument("--window", type=int, default=5)
+    args = ap.parse_args()
+
+    metrics = filter_mission(args.nav_dir, auv_filter=args.auv,
+                             sigma=args.sigma, window=args.window)
+
+    print("\n=== Stage 03 summary ===")
+    for m in metrics:
+        print(f"  {m.get('file','?')}: {m.get('points_in',0)} → {m.get('points_out',0)} "
+              f"[{m.get('status','?')}]")
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/03_usbl_filter.py
+++ b/pipeline/stages/03_usbl_filter.py
@@ -0,0 +1,248 @@
+#!/usr/bin/env python3
+"""Stage 03 — Filter and smooth USBL navigation trajectory.
+
+Input:  02_usbl_raw/<AUV>_nav_raw.json (or .csv)
+Output: 03_usbl_filtered/<AUV>_nav_filtered.json + .csv
+
+Steps:
+1. Drop points with null lat/lon
+2. MAD-3σ outlier removal on lat, lon, depth independently
+3. Moving-average smoothing (window=5 by default)
+4. Optional: simple 1D Kalman on each axis (KISS — no cross-covariance)
+
+Usage:
+    python3 03_usbl_filter.py /home/cosma/cosma-pipeline/20260505-Lepradet/02_usbl_raw/
+    python3 03_usbl_filter.py /path/to/02_usbl_raw/ --auv AUV010 --sigma 2.5
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
+
+LOG_DIR = Path(os.environ.get("COSMA_PIPELINE_LOGS", "/home/cosma/cosma-pipeline/logs"))
+LOG_DIR.mkdir(parents=True, exist_ok=True)
+
+
+def mad_outlier_mask(arr: np.ndarray, sigma: float = 3.0) -> np.ndarray:
+    """Returns boolean mask: True = keep (inlier). Uses MAD-based sigma."""
+    if len(arr) < 4:
+        return np.ones(len(arr), dtype=bool)
+    median = np.median(arr)
+    mad = np.median(np.abs(arr - median))
+    if mad == 0:
+        return np.ones(len(arr), dtype=bool)
+    modified_z = 0.6745 * (arr - median) / mad
+    return np.abs(modified_z) < sigma
+
+
+def moving_average(arr: np.ndarray, window: int = 5) -> np.ndarray:
+    """Centered moving average with edge padding."""
+    if len(arr) < window:
+        return arr.copy()
+    pad = window // 2
+    padded = np.pad(arr, (pad, pad), mode="edge")
+    kernel = np.ones(window) / window
+    return np.convolve(padded, kernel, mode="valid")[:len(arr)]
+
+
+def simple_kalman_1d(measurements: np.ndarray,
+                     process_noise: float = 1e-4,
+                     measurement_noise: float = 1e-2) -> np.ndarray:
+    """Very simple 1D Kalman filter (scalar, no velocity state).
+    KISS: just smooths, no cross-axis coupling.
+    """
+    n = len(measurements)
+    filtered = np.zeros(n)
+    x_est = measurements[0]
+    p_est = 1.0
+
+    for i, z in enumerate(measurements):
+        # Predict
+        p_pred = p_est + process_noise
+        # Update
+        K = p_pred / (p_pred + measurement_noise)
+        x_est = x_est + K * (z - x_est)
+        p_est = (1 - K) * p_pred
+        filtered[i] = x_est
+
+    return filtered
+
+
+def filter_auv_nav(raw_path: Path, out_path: Path,
+                   sigma: float = 3.0, window: int = 5,
+                   use_kalman: bool = False) -> dict:
+    """Filter nav for one AUV. Returns metrics dict."""
+    data = json.loads(raw_path.read_text())
+    points = data.get("points", [])
+    auv_id = data.get("auv_id", raw_path.stem.replace("_nav_raw", ""))
+
+    metrics = {
+        "auv_id": auv_id,
+        "points_before": len(points),
+        "points_after": 0,
+        "points_removed_null": 0,
+        "points_removed_outlier": 0,
+        "status": "ok",
+    }
+
+    if not points:
+        metrics["status"] = "degraded"
+        metrics["note"] = "no points to filter"
+        _save_output(auv_id, [], out_path, metrics)
+        return metrics
+
+    # Step 1: Drop null lat/lon
+    valid = [p for p in points if p.get("lat") is not None and p.get("lon") is not None]
+    metrics["points_removed_null"] = len(points) - len(valid)
+
+    if not valid:
+        metrics["status"] = "degraded"
+        metrics["note"] = "all points have null lat/lon (serial-only data)"
+        print(f"  [03] {auv_id}: all null lat/lon — degraded (serial CSV source, no MCAP nav)")
+        _save_output(auv_id, [], out_path, metrics)
+        return metrics
+
+    # Step 2: MAD outlier removal
+    lats = np.array([p["lat"] for p in valid])
+    lons = np.array([p["lon"] for p in valid])
+
+    mask_lat = mad_outlier_mask(lats, sigma)
+    mask_lon = mad_outlier_mask(lons, sigma)
+    mask = mask_lat & mask_lon
+
+    # Also filter depth if present
+    depths = np.array([p.get("depth") or np.nan for p in valid])
+    if not np.all(np.isnan(depths)):
+        mask_depth = mad_outlier_mask(depths[~np.isnan(depths)], sigma)
+        # Map back — only filter where we have depth
+        depth_idx = np.where(~np.isnan(depths))[0]
+        for i, keep in zip(depth_idx, mask_depth):
+            if not keep:
+                mask[i] = False
+
+    filtered_points = [p for p, keep in zip(valid, mask) if keep]
+    metrics["points_removed_outlier"] = int(np.sum(~mask))
+
+    # Step 3: Sort by timestamp
+    filtered_points.sort(key=lambda p: p["timestamp"])
+
+    # Step 4: Smooth
+    if len(filtered_points) >= window:
+        filt_lats = moving_average(np.array([p["lat"] for p in filtered_points]), window)
+        filt_lons = moving_average(np.array([p["lon"] for p in filtered_points]), window)
+        for i, p in enumerate(filtered_points):
+            p = dict(p)
+            p["lat"] = float(filt_lats[i])
+            p["lon"] = float(filt_lons[i])
+            filtered_points[i] = p
+
+    # Step 5: Optional Kalman
+    if use_kalman and len(filtered_points) > 4:
+        k_lats = simple_kalman_1d(np.array([p["lat"] for p in filtered_points]))
+        k_lons = simple_kalman_1d(np.array([p["lon"] for p in filtered_points]))
+        for i, p in enumerate(filtered_points):
+            p = dict(p)
+            p["lat"] = float(k_lats[i])
+            p["lon"] = float(k_lons[i])
+            filtered_points[i] = p
+
+    metrics["points_after"] = len(filtered_points)
+    if metrics["points_after"] < 5:
+        metrics["status"] = "degraded"
+        metrics["note"] = f"too few points after filter: {metrics['points_after']}"
+
+    print(f"  [03] {auv_id}: {metrics['points_before']} → {metrics['points_after']} "
+          f"(removed {metrics['points_removed_null']} null, {metrics['points_removed_outlier']} outliers)")
+
+    _save_output(auv_id, filtered_points, out_path, metrics)
+    return metrics
+
+
+def _save_output(auv_id: str, points: list[dict], out_dir: Path, metrics: dict) -> None:
+    out_dir.mkdir(parents=True, exist_ok=True)
+    json_out = out_dir / f"{auv_id}_nav_filtered.json"
+    json_out.write_text(json.dumps({
+        "auv_id": auv_id,
+        "generated_at": now_iso(),
+        "metrics": metrics,
+        "points": points,
+    }, indent=2, default=str))
+
+    csv_out = out_dir / f"{auv_id}_nav_filtered.csv"
+    with open(csv_out, "w") as f:
+        f.write("timestamp,lat,lon,depth,heading,source\n")
+        for p in points:
+            f.write(f"{p['timestamp']},{p.get('lat','')},{p.get('lon','')},{p.get('depth','')},{p.get('heading','')},{p.get('source','')}\n")
+
+
+def filter_mission(raw_dir: Path, out_dir: Path | None = None,
+                   auv_filter: str | None = None,
+                   sigma: float = 3.0, window: int = 5,
+                   use_kalman: bool = False) -> list[dict]:
+    if out_dir is None:
+        out_dir = raw_dir.parent / "03_usbl_filtered"
+
+    raw_files = sorted(raw_dir.glob("*_nav_raw.json"))
+    if auv_filter:
+        raw_files = [f for f in raw_files if auv_filter in f.name]
+
+    all_metrics = []
+    init_db()
+
+    for raw_file in raw_files:
+        out_file = out_dir / raw_file.name.replace("_raw", "_filtered")
+        if out_file.exists():
+            print(f"[03] {raw_file.stem}: output exists, skipping")
+            existing = json.loads(out_file.read_text())
+            all_metrics.append(existing.get("metrics", {}))
+            continue
+
+        print(f"[03] Filtering {raw_file.name} ...")
+        m = filter_auv_nav(raw_file, out_dir, sigma=sigma, window=window, use_kalman=use_kalman)
+        all_metrics.append(m)
+
+        # DB record
+        with get_conn() as conn:
+            mission_name = raw_dir.parent.name
+            mission_row = conn.execute("SELECT id FROM missions WHERE name=?", (mission_name,)).fetchone()
+            if mission_row:
+                job_id = upsert_job(conn, mission_row["id"], m["auv_id"], "all", "03_usbl_filter",
+                                    status="done" if m["status"] == "ok" else m["status"],
+                                    output_path=str(out_dir))
+                record_metric(conn, job_id, "usbl_points_before", value=m.get("points_before", 0))
+                record_metric(conn, job_id, "usbl_points_after", value=m.get("points_after", 0),
+                              pass_fail="pass" if m.get("points_after", 0) >= 5 else "fail")
+                record_metric(conn, job_id, "usbl_points_removed_outlier",
+                              value=m.get("points_removed_outlier", 0))
+
+    return all_metrics
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 03 — Filter USBL navigation")
+    ap.add_argument("raw_dir", type=Path, help="Directory with *_nav_raw.json files")
+    ap.add_argument("--out", type=Path, default=None)
+    ap.add_argument("--auv", type=str, default=None)
+    ap.add_argument("--sigma", type=float, default=3.0, help="MAD sigma threshold")
+    ap.add_argument("--window", type=int, default=5, help="Moving average window")
+    ap.add_argument("--kalman", action="store_true", help="Apply simple Kalman smoothing")
+    args = ap.parse_args()
+
+    metrics = filter_mission(args.raw_dir, out_dir=args.out, auv_filter=args.auv,
+                             sigma=args.sigma, window=args.window, use_kalman=args.kalman)
+
+    print("\n=== Stage 03 summary ===")
+    for m in metrics:
+        print(f"  {m.get('auv_id','?')}: {m.get('points_before',0)} → {m.get('points_after',0)} [{m.get('status','?')}]")
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/04_frame_extract.py
+++ b/pipeline/stages/04_frame_extract.py
@@ -0,0 +1,341 @@
+#!/usr/bin/env python3
+"""Stage 04 — Extract frames from GoPro videos.
+
+For each MP4 in /mnt/ssd/<mission>/raw_data/medias/videos/GP*-AUV*/
+- Skip files < 2MB (placeholders)
+- Auto-trim hors-eau: sample frames at start/end, detect non-blue/green pixels
+- ffmpeg fps=1, scale=518:294, q:v=3
+- Output: ~/cosma-pipeline/data/<mission>/frames/<AUV>/<segment>/frame_XXXXX.jpg
+- Skip if output dir exists and has >= expected frames
+- Log to SQLite state.db
+
+Usage:
+    python3 04_frame_extract.py --mission 20260505-Lepradet
+    python3 04_frame_extract.py --video /mnt/ssd/.../GP1-AUV210/GX019837.MP4
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+sys.path.insert(0, str(Path(__file__).parent))
+from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
+from lib_frame_qc import score_image_file, aggregate as qc_aggregate
+
+QC_SAMPLE_RATE = int(os.environ.get("COSMA_QC_SAMPLE_RATE", "5"))
+QC_BOTTOM_OK_PCT = float(os.environ.get("COSMA_QC_BOTTOM_OK_PCT", "50"))
+
+PIPELINE_BASE = Path(os.environ.get("COSMA_PIPELINE_BASE", "/home/cosma/cosma-pipeline"))
+SSD_BASE = Path(os.environ.get("COSMA_SSD_BASE", "/mnt/ssd"))
+MIN_VIDEO_SIZE_MB = 2.0
+
+
+def is_underwater_frame(frame_bgr: np.ndarray, threshold: float = 0.6) -> bool:
+    """Return True if frame looks like underwater footage (dominant blue/green).
+    Hors-eau: R > G-5 AND R > B-5 (dry/air dominant).
+    Underwater: blue or green channel dominant.
+    """
+    b, g, r = cv2.split(frame_bgr.astype(np.float32))
+    mean_r = float(np.mean(r))
+    mean_g = float(np.mean(g))
+    mean_b = float(np.mean(b))
+    # Not underwater: red dominates
+    if mean_r > mean_g - 5 and mean_r > mean_b - 5:
+        return False
+    return True
+
+
+def detect_water_range(video_path: Path, sample_count: int = 10) -> tuple[float, float]:
+    """Detect start/end times of underwater portion by sampling frames.
+    Returns (start_s, end_s). Returns (0, duration) if uncertain.
+    """
+    cap = cv2.VideoCapture(str(video_path))
+    if not cap.isOpened():
+        return 0.0, -1.0
+
+    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    duration_s = total_frames / fps if fps > 0 else 0
+
+    # Sample frames: first 20% and last 20%
+    probe_times_start = [duration_s * i / (sample_count * 5) for i in range(sample_count)]
+    probe_times_end = [duration_s * (1 - i / (sample_count * 5)) for i in range(sample_count)]
+
+    # Find first underwater frame from start
+    start_s = 0.0
+    for t in probe_times_start:
+        cap.set(cv2.CAP_PROP_POS_MSEC, t * 1000)
+        ret, frame = cap.read()
+        if ret and is_underwater_frame(frame):
+            start_s = max(0.0, t - 2.0)
+            break
+
+    # Find last underwater frame from end
+    end_s = duration_s
+    for t in sorted(probe_times_end, reverse=True):
+        cap.set(cv2.CAP_PROP_POS_MSEC, t * 1000)
+        ret, frame = cap.read()
+        if ret and is_underwater_frame(frame):
+            end_s = min(duration_s, t + 2.0)
+            break
+
+    cap.release()
+    return start_s, end_s
+
+
+def get_video_duration(video_path: Path) -> float:
+    """Get video duration in seconds via ffprobe."""
+    cmd = [
+        "ffprobe", "-v", "quiet", "-print_format", "json",
+        "-show_streams", str(video_path)
+    ]
+    try:
+        r = subprocess.run(cmd, capture_output=True, text=True, timeout=30)
+        data = json.loads(r.stdout)
+        for stream in data.get("streams", []):
+            dur = float(stream.get("duration", 0))
+            if dur > 0:
+                return dur
+    except Exception:
+        pass
+    return 0.0
+
+
+def extract_frames(video_path: Path, out_dir: Path, fps: int = 1,
+                   scale: str = "518:294", quality: int = 3,
+                   start_s: float = 0.0, end_s: float = -1.0) -> dict:
+    """Run ffmpeg to extract frames. Returns metrics dict."""
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    # Build ffmpeg args
+    cmd = ["ffmpeg", "-y", "-loglevel", "error"]
+    cmd += ["-ss", str(start_s), "-i", str(video_path)]
+    if end_s > 0 and end_s > start_s:
+        cmd += ["-t", str(end_s - start_s)]
+    cmd += [
+        "-vf", f"fps={fps},scale={scale}",
+        "-q:v", str(quality),
+        str(out_dir / "frame_%05d.jpg"),
+    ]
+
+    t0 = time.time()
+    result = subprocess.run(cmd, capture_output=True, text=True, timeout=3600)
+    elapsed = time.time() - t0
+
+    frames = sorted(out_dir.glob("frame_*.jpg"))
+    n_frames = len(frames)
+
+    metrics = {
+        "video": str(video_path),
+        "out_dir": str(out_dir),
+        "n_frames": n_frames,
+        "elapsed_s": round(elapsed, 1),
+        "returncode": result.returncode,
+        "start_s": start_s,
+        "end_s": end_s,
+    }
+    if result.returncode != 0:
+        metrics["error"] = result.stderr[-500:]
+        print(f"  [04] ffmpeg error for {video_path.name}: {result.stderr[-200:]}")
+    else:
+        print(f"  [04] {video_path.name}: {n_frames} frames in {elapsed:.1f}s")
+        # Score a subsample for QC
+        qc = qc_segment(out_dir, sample_rate=QC_SAMPLE_RATE)
+        if qc:
+            metrics.update(qc)
+
+    return metrics
+
+
+def qc_segment(frames_dir: Path, sample_rate: int = 5) -> dict | None:
+    """Sample 1/sample_rate frames, score each, write qc.json, return aggregate."""
+    frames = sorted(frames_dir.glob("frame_*.jpg"))
+    if not frames:
+        return None
+    sampled = frames[::max(1, sample_rate)]
+    per_frame = []
+    for f in sampled:
+        s = score_image_file(f)
+        if s is not None:
+            per_frame.append(s)
+    if not per_frame:
+        return None
+    agg = qc_aggregate(per_frame)
+    qc_payload = {
+        "frames_in_dir": len(frames),
+        "frames_sampled": len(per_frame),
+        "sample_rate": sample_rate,
+        **agg,
+        "per_frame": per_frame,
+    }
+    try:
+        (frames_dir / "qc.json").write_text(json.dumps(qc_payload, indent=2))
+    except Exception as e:
+        print(f"  [04] qc.json write failed: {e}")
+    print(
+        f"  [04]   QC: bottom_visible={agg['bottom_visible_pct']}% "
+        f"(b={agg['frames_bottom_visible']} ooo={agg['frames_out_of_water']} "
+        f"turb={agg['frames_turbid']} nob={agg['frames_water_no_bottom']})"
+    )
+    return agg
+
+
+def process_video(video_path: Path, auv_id: str, mission_name: str) -> dict:
+    """Process one video file. Returns metrics."""
+    size_mb = video_path.stat().st_size / (1024 * 1024)
+    if size_mb < MIN_VIDEO_SIZE_MB:
+        print(f"  [04] Skip {video_path.name} ({size_mb:.1f}MB < {MIN_VIDEO_SIZE_MB}MB)")
+        return {"video": str(video_path), "skipped": True, "reason": "placeholder"}
+
+    segment = video_path.stem
+    out_dir = PIPELINE_BASE / "data" / mission_name / "frames" / auv_id / segment
+
+    # Check if already done
+    if out_dir.exists():
+        existing = list(out_dir.glob("frame_*.jpg"))
+        duration_s = get_video_duration(video_path)
+        expected = max(1, int(duration_s) - 10) if duration_s > 0 else 1
+        if len(existing) >= expected:
+            print(f"  [04] {video_path.name}: already done ({len(existing)} frames), skip")
+            cached_m: dict = {"video": str(video_path), "n_frames": len(existing), "cached": True,
+                              "out_dir": str(out_dir)}
+            # Re-run QC if qc.json is missing (idempotent enrichment)
+            if not (out_dir / "qc.json").exists():
+                qc = qc_segment(out_dir, sample_rate=QC_SAMPLE_RATE)
+                if qc:
+                    cached_m.update(qc)
+            else:
+                try:
+                    cached_qc = json.loads((out_dir / "qc.json").read_text())
+                    for k in (
+                        "frames_total", "frames_bottom_visible", "frames_out_of_water",
+                        "frames_turbid", "frames_water_no_bottom", "bottom_visible_pct",
+                    ):
+                        if k in cached_qc:
+                            cached_m[k] = cached_qc[k]
+                except Exception:
+                    pass
+            return cached_m
+
+    print(f"  [04] {video_path.name} ({size_mb:.0f}MB): detecting water range...")
+    start_s, end_s = detect_water_range(video_path)
+    print(f"  [04]   water range: {start_s:.1f}s → {end_s:.1f}s")
+
+    return extract_frames(video_path, out_dir, start_s=start_s, end_s=end_s)
+
+
+def find_auv_videos(mission_path: Path) -> dict[str, list[Path]]:
+    """Find all MP4 files per AUV in medias/videos/GP*-AUV*/."""
+    videos_root = mission_path / "raw_data/medias/videos"
+    result: dict[str, list[Path]] = {}
+
+    for gopro_dir in sorted(videos_root.glob("GP*-AUV*")):
+        # Extract AUV ID from dir name: GP1-AUV210 -> AUV210
+        parts = gopro_dir.name.split("-")
+        if len(parts) >= 2:
+            auv_id = parts[1]
+            mp4_files = [f for f in sorted(gopro_dir.glob("GX*.MP4"))
+                         if f.stat().st_size / (1024 * 1024) >= MIN_VIDEO_SIZE_MB]
+            if mp4_files:
+                if auv_id not in result:
+                    result[auv_id] = []
+                result[auv_id].extend(mp4_files)
+
+    return result
+
+
+def process_mission(mission_name: str) -> list[dict]:
+    mission_path = SSD_BASE / mission_name
+    auv_videos = find_auv_videos(mission_path)
+    print(f"[04] Found AUVs: {list(auv_videos.keys())}")
+
+    all_metrics = []
+    init_db()
+
+    for auv_id, videos in sorted(auv_videos.items()):
+        print(f"[04] === {auv_id}: {len(videos)} videos ===")
+        for video_path in videos:
+            t0 = time.time()
+            m = process_video(video_path, auv_id, mission_name)
+            m["auv_id"] = auv_id
+            all_metrics.append(m)
+
+            if not m.get("skipped"):
+                with get_conn() as conn:
+                    mission_row = conn.execute(
+                        "SELECT id FROM missions WHERE name=?", (mission_name,)
+                    ).fetchone()
+                    if mission_row:
+                        bottom_pct = m.get("bottom_visible_pct")
+                        if m.get("returncode", 0) != 0:
+                            job_status = "error"
+                        elif bottom_pct is not None and bottom_pct < QC_BOTTOM_OK_PCT:
+                            job_status = "degraded"
+                        else:
+                            job_status = "done"
+                        job_id = upsert_job(
+                            conn, mission_row["id"], auv_id,
+                            video_path.stem, "04_frame_extract",
+                            status=job_status,
+                            output_path=m.get("out_dir", ""),
+                            error_msg=(
+                                f"bottom_visible_pct={bottom_pct}% <{QC_BOTTOM_OK_PCT}%"
+                                if job_status == "degraded" else None
+                            ),
+                        )
+                        if not m.get("cached"):
+                            record_metric(conn, job_id, "frames_extracted",
+                                          value=m.get("n_frames", 0),
+                                          pass_fail="pass" if m.get("n_frames", 0) > 0 else "fail")
+                            record_metric(conn, job_id, "extract_time_s",
+                                          value=m.get("elapsed_s", 0))
+                        # Always record QC metrics (so cached frames also get scored history)
+                        for k in (
+                            "frames_total", "frames_bottom_visible", "frames_out_of_water",
+                            "frames_turbid", "frames_water_no_bottom",
+                        ):
+                            if k in m:
+                                record_metric(conn, job_id, k, value=float(m[k]))
+                        if bottom_pct is not None:
+                            record_metric(
+                                conn, job_id, "bottom_visible_pct",
+                                value=float(bottom_pct),
+                                pass_fail="pass" if bottom_pct >= QC_BOTTOM_OK_PCT else "degraded",
+                            )
+
+    return all_metrics
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 04 — Extract frames from GoPro videos")
+    ap.add_argument("--mission", type=str, help="Mission name (e.g. 20260505-Lepradet)")
+    ap.add_argument("--video", type=Path, help="Single video path")
+    ap.add_argument("--auv", type=str, default="UNKNOWN", help="AUV ID for single video mode")
+    args = ap.parse_args()
+
+    if args.video:
+        mission_name = args.mission or "unknown"
+        m = process_video(args.video, args.auv, mission_name)
+        print(f"\nResult: {m}")
+    elif args.mission:
+        metrics = process_mission(args.mission)
+        print("\n=== Stage 04 summary ===")
+        total_frames = sum(m.get("n_frames", 0) for m in metrics if not m.get("skipped"))
+        skipped = sum(1 for m in metrics if m.get("skipped"))
+        print(f"Total frames: {total_frames}, skipped: {skipped}")
+    else:
+        ap.print_help()
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/04b_trim_water.py
+++ b/pipeline/stages/04b_trim_water.py
@@ -0,0 +1,428 @@
+#!/usr/bin/env python3
+"""Stage 04b — Trim out-of-water (hors-eau) head/tail frames from already-extracted segments.
+
+Ports the sustained-run trim logic from cosma-qc/scripts/dispatcher.py (_AUTO_TRIM_SCRIPT,
+trim_above_water_prefix) into the new cosma-pipeline pipeline. Re-runs frame QC scoring
+on the trimmed set and updates state.db (jobs.status + metrics).
+
+Usage:
+    python3 04b_trim_water.py --mission 20260505-Lepradet
+    python3 04b_trim_water.py --mission 20260505-Lepradet --auv AUV210 --segment GX019837
+    python3 04b_trim_water.py --mission 20260505-Lepradet --dry-run
+
+Safety:
+    - Skips segments where ffmpeg is still running on the frames dir (extraction in progress).
+    - Skips segments with a queued/running 05_inference job in state.db.
+    - Skips segments whose frame count is not stable over a 5s window.
+    - Never deletes all frames (sanity floor: keep everything if trim would empty the dir).
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+import cv2
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+sys.path.insert(0, str(Path(__file__).parent))
+from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
+from lib_frame_qc import score_image_file, aggregate as qc_aggregate
+
+PIPELINE_BASE = Path(os.environ.get("COSMA_PIPELINE_BASE", "/home/cosma/cosma-pipeline"))
+QC_SAMPLE_RATE = int(os.environ.get("COSMA_QC_SAMPLE_RATE", "5"))
+QC_BOTTOM_OK_PCT = float(os.environ.get("COSMA_QC_BOTTOM_OK_PCT", "50"))
+NEED_STREAK = 10  # consecutive underwater frames required to lock start/end
+
+
+# -----------------------------------------------------------------------------
+# Trim logic (ported verbatim from dispatcher._AUTO_TRIM_SCRIPT)
+# -----------------------------------------------------------------------------
+
+def is_underwater(path: Path) -> bool | None:
+    img = cv2.imread(str(path), cv2.IMREAD_REDUCED_COLOR_4)
+    if img is None:
+        return None
+    b, g, r = [float(c) for c in cv2.mean(img)[:3]]
+    # Red is absorbed by water → R < G AND R < B on underwater shots.
+    return r < g - 5 and r < b - 5
+
+
+def trim_segment(frames_dir: Path, dry_run: bool = False) -> tuple[int, int, int]:
+    """Delete leading and trailing out-of-water frames.
+    Returns (head_removed, tail_removed, remaining).
+    """
+    paths = sorted(frames_dir.glob("frame_*.jpg"))
+    if not paths:
+        return (0, 0, 0)
+
+    # Scan from start
+    start = 0
+    streak = 0
+    for i, p in enumerate(paths):
+        uw = is_underwater(p)
+        if uw is None:
+            continue
+        if uw:
+            streak += 1
+            if streak >= NEED_STREAK:
+                start = i - NEED_STREAK + 1
+                break
+        else:
+            streak = 0
+
+    # Scan from end
+    end = len(paths)
+    streak = 0
+    for j in range(len(paths) - 1, -1, -1):
+        uw = is_underwater(paths[j])
+        if uw is None:
+            continue
+        if uw:
+            streak += 1
+            if streak >= NEED_STREAK:
+                end = j + NEED_STREAK  # exclusive
+                break
+        else:
+            streak = 0
+
+    if end <= start:
+        # Sanity: never delete everything.
+        start = 0
+        end = len(paths)
+
+    removed_head = start
+    removed_tail = len(paths) - end
+
+    if not dry_run:
+        for p in paths[:start]:
+            try:
+                p.unlink()
+            except OSError:
+                pass
+        for p in paths[end:]:
+            try:
+                p.unlink()
+            except OSError:
+                pass
+
+    return (removed_head, removed_tail, end - start)
+
+
+# -----------------------------------------------------------------------------
+# Safety: is this segment currently being touched?
+# -----------------------------------------------------------------------------
+
+def has_ffmpeg_running_on(frames_dir: Path) -> bool:
+    """Check if any ffmpeg process is writing into frames_dir."""
+    try:
+        r = subprocess.run(
+            ["pgrep", "-af", "ffmpeg"], capture_output=True, text=True, timeout=5
+        )
+        for line in r.stdout.splitlines():
+            if str(frames_dir) in line:
+                return True
+    except Exception:
+        pass
+    return False
+
+
+def has_inference_running_on(frames_dir: Path) -> bool:
+    """Check if any 05_inference.py process is running on frames_dir."""
+    try:
+        r = subprocess.run(
+            ["pgrep", "-af", "05_inference"], capture_output=True, text=True, timeout=5
+        )
+        for line in r.stdout.splitlines():
+            if str(frames_dir) in line:
+                return True
+    except Exception:
+        pass
+    return False
+
+
+def has_pending_inference_job(conn, mission_id: int, auv_id: str, segment: str) -> bool:
+    """Check state.db for queued/running 05_inference job on this segment."""
+    row = conn.execute(
+        "SELECT status FROM jobs WHERE mission_id=? AND auv_id=? "
+        "AND segment_label=? AND stage='05_inference'",
+        (mission_id, auv_id, segment),
+    ).fetchone()
+    if row is None:
+        return False
+    return row["status"] in ("queued", "running")
+
+
+def frame_count_is_stable(frames_dir: Path, wait_s: float = 5.0) -> bool:
+    """Return True if the frame count doesn't change over wait_s."""
+    n1 = sum(1 for _ in frames_dir.glob("frame_*.jpg"))
+    time.sleep(wait_s)
+    n2 = sum(1 for _ in frames_dir.glob("frame_*.jpg"))
+    return n1 == n2
+
+
+# -----------------------------------------------------------------------------
+# QC re-scoring (mirrors stage 04 qc_segment)
+# -----------------------------------------------------------------------------
+
+def qc_segment(frames_dir: Path, sample_rate: int = QC_SAMPLE_RATE) -> dict | None:
+    frames = sorted(frames_dir.glob("frame_*.jpg"))
+    if not frames:
+        return None
+    sampled = frames[::max(1, sample_rate)]
+    per_frame = []
+    for f in sampled:
+        s = score_image_file(f)
+        if s is not None:
+            per_frame.append(s)
+    if not per_frame:
+        return None
+    agg = qc_aggregate(per_frame)
+    qc_payload = {
+        "frames_in_dir": len(frames),
+        "frames_sampled": len(per_frame),
+        "sample_rate": sample_rate,
+        **agg,
+        "per_frame": per_frame,
+    }
+    try:
+        (frames_dir / "qc.json").write_text(json.dumps(qc_payload, indent=2))
+    except Exception as e:
+        print(f"  [04b] qc.json write failed: {e}")
+    return agg
+
+
+# -----------------------------------------------------------------------------
+# Main per-segment driver
+# -----------------------------------------------------------------------------
+
+def process_segment(mission_name: str, auv_id: str, segment: str,
+                    frames_dir: Path, dry_run: bool, conn) -> dict:
+    result = {
+        "auv_id": auv_id,
+        "segment": segment,
+        "frames_dir": str(frames_dir),
+        "skipped": False,
+        "head_removed": 0,
+        "tail_removed": 0,
+        "remaining": 0,
+        "before_total": 0,
+        "before_bottom_pct": None,
+        "after_bottom_pct": None,
+        "status_before": None,
+        "status_after": None,
+    }
+
+    if not frames_dir.is_dir():
+        result["skipped"] = True
+        result["reason"] = "no_frames_dir"
+        return result
+
+    # Safety checks
+    if has_ffmpeg_running_on(frames_dir):
+        result["skipped"] = True
+        result["reason"] = "ffmpeg_running"
+        print(f"  [04b] SKIP {auv_id}/{segment}: ffmpeg still extracting")
+        return result
+
+    if has_inference_running_on(frames_dir):
+        result["skipped"] = True
+        result["reason"] = "inference_running_proc"
+        print(f"  [04b] SKIP {auv_id}/{segment}: 05_inference process running")
+        return result
+
+    # Look up mission_id + current 04 job
+    mission_row = conn.execute(
+        "SELECT id FROM missions WHERE name=?", (mission_name,)
+    ).fetchone()
+    if not mission_row:
+        result["skipped"] = True
+        result["reason"] = "mission_not_in_db"
+        return result
+    mission_id = mission_row["id"]
+
+    if has_pending_inference_job(conn, mission_id, auv_id, segment):
+        result["skipped"] = True
+        result["reason"] = "inference_job_pending"
+        print(f"  [04b] SKIP {auv_id}/{segment}: 05_inference queued/running in DB")
+        return result
+
+    if not frame_count_is_stable(frames_dir, wait_s=5.0):
+        result["skipped"] = True
+        result["reason"] = "frame_count_unstable"
+        print(f"  [04b] SKIP {auv_id}/{segment}: frame count not stable")
+        return result
+
+    # Snapshot before
+    before_paths = sorted(frames_dir.glob("frame_*.jpg"))
+    result["before_total"] = len(before_paths)
+    job04_row = conn.execute(
+        "SELECT id, status FROM jobs WHERE mission_id=? AND auv_id=? "
+        "AND segment_label=? AND stage='04_frame_extract'",
+        (mission_id, auv_id, segment),
+    ).fetchone()
+    if job04_row is None:
+        result["skipped"] = True
+        result["reason"] = "no_04_job_in_db"
+        print(f"  [04b] SKIP {auv_id}/{segment}: no 04 job row")
+        return result
+    result["status_before"] = job04_row["status"]
+
+    # Read current QC if available
+    qc_path = frames_dir / "qc.json"
+    if qc_path.exists():
+        try:
+            result["before_bottom_pct"] = json.loads(qc_path.read_text()).get("bottom_visible_pct")
+        except Exception:
+            pass
+
+    # Trim
+    head, tail, remaining = trim_segment(frames_dir, dry_run=dry_run)
+    result["head_removed"] = head
+    result["tail_removed"] = tail
+    result["remaining"] = remaining
+
+    # Re-QC if not dry-run and something was trimmed (or always to keep metrics fresh)
+    after_agg = None
+    if not dry_run and (head > 0 or tail > 0):
+        after_agg = qc_segment(frames_dir)
+    elif dry_run:
+        # In dry-run, don't touch qc.json; compute aggregate from remaining slice in-memory
+        remaining_paths = sorted(frames_dir.glob("frame_*.jpg"))[head: len(before_paths) - tail]
+        sampled = remaining_paths[::max(1, QC_SAMPLE_RATE)]
+        per_frame = [s for s in (score_image_file(f) for f in sampled) if s is not None]
+        if per_frame:
+            after_agg = qc_aggregate(per_frame)
+
+    if after_agg is not None:
+        result["after_bottom_pct"] = after_agg.get("bottom_visible_pct")
+
+    if dry_run:
+        print(
+            f"  [04b] DRY {auv_id}/{segment}: head={head} tail={tail} "
+            f"remaining={remaining} (before={len(before_paths)}, "
+            f"bottom_pct {result['before_bottom_pct']}→{result['after_bottom_pct']})"
+        )
+        return result
+
+    # Update DB: job row + metrics
+    job_id = job04_row["id"]
+    bottom_pct = after_agg.get("bottom_visible_pct") if after_agg else None
+
+    if bottom_pct is not None and bottom_pct >= QC_BOTTOM_OK_PCT:
+        new_status = "done"
+        err_msg = None
+    elif bottom_pct is not None:
+        new_status = "degraded"
+        err_msg = f"bottom_visible_pct={bottom_pct}% <{QC_BOTTOM_OK_PCT}% (after trim)"
+    else:
+        new_status = job04_row["status"]
+        err_msg = None
+
+    upsert_job(
+        conn, mission_id, auv_id, segment, "04_frame_extract",
+        status=new_status,
+        output_path=str(frames_dir),
+        error_msg=err_msg,
+    )
+    record_metric(conn, job_id, "trimmed_head", value=float(head))
+    record_metric(conn, job_id, "trimmed_tail", value=float(tail))
+    record_metric(conn, job_id, "frames_after_trim", value=float(remaining))
+    if after_agg:
+        for k in (
+            "frames_total", "frames_bottom_visible", "frames_out_of_water",
+            "frames_turbid", "frames_water_no_bottom",
+        ):
+            if k in after_agg:
+                record_metric(conn, job_id, k, value=float(after_agg[k]))
+        if bottom_pct is not None:
+            record_metric(
+                conn, job_id, "bottom_visible_pct",
+                value=float(bottom_pct),
+                pass_fail="pass" if bottom_pct >= QC_BOTTOM_OK_PCT else "degraded",
+            )
+
+    result["status_after"] = new_status
+    print(
+        f"  [04b] {auv_id}/{segment}: trimmed head={head} tail={tail} "
+        f"remaining={remaining}, bottom_pct={bottom_pct}% ({result['status_before']}→{new_status})"
+    )
+    return result
+
+
+# -----------------------------------------------------------------------------
+# Discovery + CLI
+# -----------------------------------------------------------------------------
+
+def find_segments(mission_name: str, auv_filter: str | None,
+                  segment_filter: str | None) -> list[tuple[str, str, Path]]:
+    base = PIPELINE_BASE / "data" / mission_name / "frames"
+    out: list[tuple[str, str, Path]] = []
+    if not base.is_dir():
+        return out
+    for auv_dir in sorted(base.iterdir()):
+        if not auv_dir.is_dir():
+            continue
+        if auv_filter and auv_dir.name != auv_filter:
+            continue
+        for seg_dir in sorted(auv_dir.iterdir()):
+            if not seg_dir.is_dir():
+                continue
+            if segment_filter and seg_dir.name != segment_filter:
+                continue
+            out.append((auv_dir.name, seg_dir.name, seg_dir))
+    return out
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 04b — Trim hors-eau head/tail frames")
+    ap.add_argument("--mission", default="20260505-Lepradet")
+    ap.add_argument("--auv")
+    ap.add_argument("--segment")
+    ap.add_argument("--dry-run", action="store_true")
+    args = ap.parse_args()
+
+    init_db()
+
+    segments = find_segments(args.mission, args.auv, args.segment)
+    if not segments:
+        print(f"[04b] No segments found under {args.mission}")
+        sys.exit(1)
+
+    print(f"[04b] Mission={args.mission}  segments={len(segments)}  dry_run={args.dry_run}")
+    results: list[dict] = []
+    with get_conn() as conn:
+        for auv_id, segment, frames_dir in segments:
+            try:
+                r = process_segment(args.mission, auv_id, segment, frames_dir,
+                                    args.dry_run, conn)
+            except Exception as e:
+                r = {"auv_id": auv_id, "segment": segment, "error": str(e),
+                     "skipped": True}
+                print(f"  [04b] ERR {auv_id}/{segment}: {e}")
+            results.append(r)
+
+    # Summary
+    print("\n=== Stage 04b summary ===")
+    upgraded = [r for r in results
+                if r.get("status_before") == "degraded" and r.get("status_after") == "done"]
+    still_degraded = [r for r in results
+                      if r.get("status_after") == "degraded"]
+    skipped = [r for r in results if r.get("skipped")]
+    print(f"Upgraded degraded→done : {len(upgraded)}")
+    for r in upgraded:
+        print(f"  + {r['auv_id']}/{r['segment']} "
+              f"({r['before_bottom_pct']}%→{r['after_bottom_pct']}%, "
+              f"trim head={r['head_removed']} tail={r['tail_removed']})")
+    print(f"Still degraded         : {len(still_degraded)}")
+    print(f"Skipped                : {len(skipped)}")
+    for r in skipped:
+        print(f"  - {r['auv_id']}/{r['segment']}: {r.get('reason', 'unknown')}")
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/05_inference.py
+++ b/pipeline/stages/05_inference.py
@@ -0,0 +1,319 @@
+#!/usr/bin/env python3
+"""Stage 05 — Run lingbot-map inference on extracted frames.
+
+Args:
+    --frames-dir <path>   Directory with frame_*.jpg (or parent with AUV subdirs)
+    --worker <auto|.84|.87>   GPU worker selection
+    --mission <name>      Mission name for output paths
+
+Workers:
+  .84: /root/ai-video/lingbot-map/.venv/bin/python demo.py ...
+  .87: /home/floppyrj45/ai-video/lingbot-map/.venv/bin/python demo.py ...
+
+Auto: pick by lowest GPU memory usage (nvidia-smi via SSH).
+
+Flow:
+1. rsync frames .83 → worker /root/cosma-frames-tmp/ (or /home/floppyrj45/)
+2. SSH launch demo.py with windowed mode (window=64, overlap=16)
+3. Retrieve PLY + NPZ → .83 ~/cosma-pipeline/data/<mission>/ply/<AUV>/<segment>.{ply,npz}
+4. Cleanup worker temp dir
+5. Log to SQLite: duration, GPU peak mem, nb points in PLY
+
+Usage:
+    python3 05_inference.py --frames-dir ~/cosma-pipeline/data/20260505-Lepradet/frames/AUV210/GX019837 --worker auto --mission 20260505-Lepradet
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+import yaml
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+from orchestrator.db import init_db, get_conn, upsert_job, record_metric, now_iso
+
+PIPELINE_BASE = Path(os.environ.get("COSMA_PIPELINE_BASE", "/home/cosma/cosma-pipeline"))
+
+def _load_inference_cfg() -> dict:
+    """Load inference params from thresholds.yaml, with sane defaults."""
+    cfg_path = Path(__file__).parent.parent / "config" / "thresholds.yaml"
+    try:
+        data = yaml.safe_load(cfg_path.read_text())
+        return data.get("inference", {})
+    except Exception:
+        return {}
+
+_INF_CFG = _load_inference_cfg()
+
+WORKERS = {
+    ".84": {
+        "host": "192.168.0.84",
+        "user": "root",
+        "ai_dir": "/root/ai-video/lingbot-map",
+        "venv": "/root/ai-video/lingbot-map/.venv/bin/python",
+        "tmp_dir": "/root/cosma-frames-tmp",
+    },
+    ".87": {
+        "host": "192.168.0.87",
+        "user": "floppyrj45",
+        "ai_dir": "/home/floppyrj45/ai-video/lingbot-map",
+        "venv": "/home/floppyrj45/ai-video/lingbot-map/.venv/bin/python",
+        "tmp_dir": "/home/floppyrj45/cosma-frames-tmp",
+    },
+}
+
+
+def get_gpu_mem_used(worker_key: str) -> int:
+    """Return GPU memory used in MB via SSH nvidia-smi. Returns 99999 on error."""
+    w = WORKERS[worker_key]
+    cmd = [
+        "ssh", "-o", "StrictHostKeyChecking=no", "-o", "ConnectTimeout=5",
+        f"{w['user']}@{w['host']}",
+        "nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits 2>/dev/null | head -1"
+    ]
+    try:
+        r = subprocess.run(cmd, capture_output=True, text=True, timeout=10)
+        return int(r.stdout.strip())
+    except Exception:
+        return 99999
+
+
+def pick_worker() -> str:
+    """Auto-select worker with lowest GPU memory usage."""
+    best = None
+    best_mem = 99999
+    for key in WORKERS:
+        mem = get_gpu_mem_used(key)
+        print(f"  [05] Worker {key}: GPU mem={mem}MB")
+        if mem < best_mem:
+            best_mem = mem
+            best = key
+    if best is None:
+        raise RuntimeError("No GPU worker available")
+    print(f"  [05] Selected worker {best}")
+    return best
+
+
+def count_ply_points(ply_path: Path) -> int:
+    """Count vertex count in PLY file header."""
+    try:
+        with open(ply_path, "rb") as f:
+            for _ in range(30):
+                line = f.readline().decode("ascii", errors="ignore").strip()
+                if line.startswith("element vertex"):
+                    return int(line.split()[-1])
+    except Exception:
+        pass
+    return 0
+
+
+def run_inference(frames_dir: Path, worker_key: str, mission_name: str,
+                  auv_id: str, segment: str) -> dict:
+    """Run lingbot-map on one segment. Returns metrics."""
+    w = WORKERS[worker_key]
+    host = w["host"]
+    user = w["user"]
+    ssh_target = f"{user}@{host}"
+    worker_frames = f"{w['tmp_dir']}/{mission_name}/{auv_id}/{segment}"
+    ply_remote = f"{w['tmp_dir']}/{mission_name}/{auv_id}/{segment}.ply"
+    npz_remote = f"{w['tmp_dir']}/{mission_name}/{auv_id}/{segment}.npz"
+
+    out_dir = PIPELINE_BASE / "data" / mission_name / "ply" / auv_id
+    out_dir.mkdir(parents=True, exist_ok=True)
+    out_ply = out_dir / f"{segment}.ply"
+    out_npz = out_dir / f"{segment}.npz"
+
+    if out_ply.exists() and out_ply.stat().st_size > 1000:
+        n_pts = count_ply_points(out_ply)
+        print(f"  [05] {auv_id}/{segment}: cached PLY ({n_pts} pts)")
+        return {"cached": True, "ply": str(out_ply), "n_points": n_pts}
+
+    metrics = {
+        "auv_id": auv_id,
+        "segment": segment,
+        "worker": worker_key,
+        "status": "ok",
+    }
+
+    # Step 1: create remote temp dir + rsync frames
+    print(f"  [05] rsync {frames_dir} → {ssh_target}:{worker_frames}...")
+    subprocess.run(
+        ["ssh", "-o", "StrictHostKeyChecking=no", ssh_target,
+         f"mkdir -p {worker_frames}"],
+        check=True, timeout=15,
+    )
+    r = subprocess.run(
+        ["rsync", "-az", "--delete",
+         str(frames_dir) + "/",
+         f"{ssh_target}:{worker_frames}/"],
+        capture_output=True, text=True, timeout=600,
+    )
+    if r.returncode != 0:
+        metrics["status"] = "error"
+        metrics["error"] = f"rsync failed: {r.stderr[-200:]}"
+        return metrics
+    print(f"  [05] rsync done")
+
+    # Step 2: build demo.py command -- params from thresholds.yaml[inference]
+    checkpoint = f"{w['ai_dir']}/checkpoints/lingbot-map/lingbot-map.pt"
+    inf_mode = _INF_CFG.get("mode", "streaming")
+    conf_thr = _INF_CFG.get("ply_conf_threshold", 1.5)
+    kf_interval = _INF_CFG.get("keyframe_interval", 1)
+    max_frames = _INF_CFG.get("max_frame_num", 1024)
+    if inf_mode == "windowed":
+        window_size = _INF_CFG.get("window_size", 64)
+        overlap_size = _INF_CFG.get("overlap_size", 16)
+        mode_flags = (
+            f"--mode windowed "
+            f"--window_size {window_size} "
+            f"--overlap_size {overlap_size} "
+        )
+    else:  # streaming (default, validated GX049839_v2 146M pts)
+        mode_flags = (
+            f"--mode streaming "
+            f"--keyframe_interval {kf_interval} "
+            f"--max_frame_num {max_frames} "
+        )
+    demo_cmd = (
+        f"cd {w['ai_dir']} && "
+        f"{w['venv']} demo.py "
+        f"--model_path {checkpoint} "
+        f"--image_folder {worker_frames} "
+        f"{mode_flags}"
+        f"--ply_conf_threshold {conf_thr} "
+        f"--save_ply {ply_remote} "
+        f"--save_poses {npz_remote} "
+        f"--use_sdpa "
+        f"--offload_to_cpu "
+        f"2>&1"
+    )
+
+    print(f"  [05] Launching inference on {host}...")
+    t0 = time.time()
+    r = subprocess.run(
+        ["ssh", "-o", "StrictHostKeyChecking=no", ssh_target, demo_cmd],
+        capture_output=True, text=True, timeout=7200,  # 2h max
+    )
+    elapsed = time.time() - t0
+    metrics["inference_s"] = round(elapsed, 1)
+
+    if r.returncode != 0:
+        metrics["status"] = "error"
+        metrics["error"] = r.stdout[-500:] + r.stderr[-200:]
+        print(f"  [05] inference error: {metrics['error'][-200:]}")
+        return metrics
+
+    print(f"  [05] Inference done in {elapsed:.1f}s")
+
+    # Step 3: GPU peak mem from nvidia-smi log (best-effort parse)
+    gpu_mem_line = [l for l in r.stdout.split("\n") if "MiB" in l]
+    metrics["gpu_peak_mb"] = get_gpu_mem_used(worker_key)
+
+    # Step 4: rsync PLY + NPZ back
+    print(f"  [05] Retrieving PLY + NPZ...")
+    for remote, local in [(ply_remote, out_ply), (npz_remote, out_npz)]:
+        r2 = subprocess.run(
+            ["rsync", "-az", f"{ssh_target}:{remote}", str(local)],
+            capture_output=True, text=True, timeout=120,
+        )
+        if r2.returncode != 0:
+            print(f"  [05] Warning: rsync back failed for {remote}: {r2.stderr[-100:]}")
+
+    # Step 5: cleanup worker
+    subprocess.run(
+        ["ssh", "-o", "StrictHostKeyChecking=no", ssh_target,
+         f"rm -rf {worker_frames} {ply_remote} {npz_remote}"],
+        timeout=30,
+    )
+
+    # Count PLY points
+    n_pts = count_ply_points(out_ply) if out_ply.exists() else 0
+    metrics["n_points"] = n_pts
+    metrics["ply"] = str(out_ply)
+    print(f"  [05] PLY: {n_pts} points → {out_ply}")
+
+    return metrics
+
+
+def process_frames_dir(frames_dir: Path, worker_key: str, mission_name: str) -> list[dict]:
+    """Process a directory of frames (single segment or AUV tree)."""
+    # Detect if frames_dir contains frame_*.jpg directly or subdirs
+    direct_frames = list(frames_dir.glob("frame_*.jpg"))
+
+    if direct_frames:
+        # Single segment
+        parts = frames_dir.parts
+        auv_id = frames_dir.parent.name if len(parts) >= 2 else "UNKNOWN"
+        segment = frames_dir.name
+        return [run_inference(frames_dir, worker_key, mission_name, auv_id, segment)]
+
+    # Tree: frames_dir/<AUV>/<segment>/frame_*.jpg
+    all_metrics = []
+    for auv_dir in sorted(frames_dir.iterdir()):
+        if not auv_dir.is_dir():
+            continue
+        auv_id = auv_dir.name
+        for seg_dir in sorted(auv_dir.iterdir()):
+            if not seg_dir.is_dir():
+                continue
+            frames = list(seg_dir.glob("frame_*.jpg"))
+            if not frames:
+                continue
+            print(f"\n[05] === {auv_id}/{seg_dir.name}: {len(frames)} frames ===")
+            m = run_inference(seg_dir, worker_key, mission_name, auv_id, seg_dir.name)
+            all_metrics.append(m)
+
+            init_db()
+            with get_conn() as conn:
+                mission_row = conn.execute(
+                    "SELECT id FROM missions WHERE name=?", (mission_name,)
+                ).fetchone()
+                if mission_row and not m.get("cached"):
+                    job_id = upsert_job(
+                        conn, mission_row["id"], auv_id, seg_dir.name, "05_inference",
+                        status="done" if m.get("status") == "ok" else m.get("status", "error"),
+                        output_path=m.get("ply", ""),
+                    )
+                    record_metric(conn, job_id, "ply_points", value=m.get("n_points", 0),
+                                  pass_fail="pass" if m.get("n_points", 0) > 100 else "fail")
+                    if "inference_s" in m:
+                        record_metric(conn, job_id, "inference_s", value=m["inference_s"])
+                    if "gpu_peak_mb" in m:
+                        record_metric(conn, job_id, "gpu_peak_mb", value=m["gpu_peak_mb"])
+
+    return all_metrics
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Stage 05 — lingbot-map inference")
+    ap.add_argument("--frames-dir", type=Path, required=True,
+                    help="Frames dir (single segment or AUV tree)")
+    ap.add_argument("--worker", type=str, default="auto",
+                    choices=["auto", ".84", ".87"])
+    ap.add_argument("--mission", type=str, required=True,
+                    help="Mission name (e.g. 20260505-Lepradet)")
+    args = ap.parse_args()
+
+    worker = args.worker
+    if worker == "auto":
+        worker = pick_worker()
+
+    metrics = process_frames_dir(args.frames_dir, worker, args.mission)
+
+    print("\n=== Stage 05 summary ===")
+    total_pts = sum(m.get("n_points", 0) for m in metrics)
+    ok = sum(1 for m in metrics if m.get("status") == "ok" or m.get("cached"))
+    print(f"  Segments OK: {ok}/{len(metrics)}, total PLY points: {total_pts}")
+    for m in metrics:
+        print(f"  {m.get('auv_id','?')}/{m.get('segment','?')}: "
+              f"{m.get('n_points',0)} pts "
+              f"[{m.get('status','cached' if m.get('cached') else '?')}]")
+
+
+if __name__ == "__main__":
+    main()
--- a/pipeline/stages/init.py
+++ b/pipeline/stages/init.py
--- a/pipeline/stages/lib_frame_qc.py
+++ b/pipeline/stages/lib_frame_qc.py
@@ -0,0 +1,83 @@
+"""Frame quality scoring for underwater footage.
+
+For each frame, compute:
+  - laplacian_var: focus/sharpness (cv2.Laplacian variance)
+  - contrast: stddev of grayscale
+  - blue_dominance: mean(B - R), positive = water dominant
+  - mean_r/g/b: per-channel means
+
+Classification (priority order):
+  - mean_r > mean_g + 5 AND mean_r > mean_b + 5 → 'out_of_water'
+  - laplacian_var < 50 AND contrast < 25       → 'turbid_water'
+  - laplacian_var >= 80 AND contrast >= 35
+    AND blue_dominance > -10                   → 'bottom_visible'
+  - else                                       → 'water_no_bottom'
+"""
+from __future__ import annotations
+
+from collections import Counter
+from typing import Iterable
+
+import cv2
+import numpy as np
+
+
+def score_frame(frame_bgr: np.ndarray) -> dict:
+    """Return per-frame QC metrics + class label."""
+    gray = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2GRAY)
+    lap_var = float(cv2.Laplacian(gray, cv2.CV_64F).var())
+    contrast = float(gray.std())
+    b, g, r = cv2.split(frame_bgr)
+    mean_r = float(r.mean())
+    mean_g = float(g.mean())
+    mean_b = float(b.mean())
+    blue_dom = float(mean_b - mean_r)
+
+    if mean_r > mean_g + 5 and mean_r > mean_b + 5:
+        klass = "out_of_water"
+    elif lap_var < 50 and contrast < 25:
+        klass = "turbid_water"
+    elif lap_var >= 80 and contrast >= 35 and blue_dom > -10:
+        klass = "bottom_visible"
+    else:
+        klass = "water_no_bottom"
+
+    return {
+        "laplacian_var": round(lap_var, 2),
+        "contrast": round(contrast, 2),
+        "blue_dominance": round(blue_dom, 2),
+        "mean_r": round(mean_r, 1),
+        "mean_g": round(mean_g, 1),
+        "mean_b": round(mean_b, 1),
+        "class": klass,
+        "score_ok": klass == "bottom_visible",
+    }
+
+
+def score_image_file(path) -> dict | None:
+    """Load image with OpenCV and score it. Returns None on failure."""
+    img = cv2.imread(str(path))
+    if img is None:
+        return None
+    res = score_frame(img)
+    res["file"] = str(path)
+    return res
+
+
+def aggregate(scores: Iterable[dict]) -> dict:
+    """Aggregate a sequence of score_frame() dicts."""
+    scores = list(scores)
+    total = len(scores)
+    counts = Counter(s["class"] for s in scores)
+    bottom = counts.get("bottom_visible", 0)
+    return {
+        "frames_total": total,
+        "frames_bottom_visible": bottom,
+        "frames_out_of_water": counts.get("out_of_water", 0),
+        "frames_turbid": counts.get("turbid_water", 0),
+        "frames_water_no_bottom": counts.get("water_no_bottom", 0),
+        "bottom_visible_pct": round(100.0 * bottom / total, 1) if total else 0.0,
+    }
+
+
+CLASS_ORDER = ("bottom_visible", "water_no_bottom", "turbid_water", "out_of_water")
--- a/pipeline/veille/2026-05-11-2233-iter-1.md
+++ b/pipeline/veille/2026-05-11-2233-iter-1.md
@@ -0,0 +1,14 @@
+# Veille — 2026-05-11 22:33 UTC — iter-1
+
+## ArXiv (signaux forts)
+- **[2605.04672]** AI-Aided Advancements in AUV Navigation — fusion caméra+DVL+IMU IA, pertinent nav AUV
+- **BALTIC benchmark** — cross-domain 3D recon air/eau illumination variable
+- **3D Gaussian Splatting underwater** — spatiotemporal degradation-aware GS scènes turbides
+
+## GitHub
+- **LingBot-Map** (maj 3j) — streaming 20FPS 518×378 10k+ frames drift correction attention paginée — fort signal
+- **DUSt3R** actif, suivi normal
+- MonST3R / VGGT / MoGe : pas de maj 7j
+
+## Action possible
+Tester 3DGS underwater sur frames AUV210 turbides (0% bottom visible) comme alternative à lingbot reconstruction
--- a/pipeline/veille/2026-05-12-0430-iter-2.md
+++ b/pipeline/veille/2026-05-12-0430-iter-2.md
@@ -0,0 +1,23 @@
+# Veille COSMA reconstruction — iter-2 — 2026-05-12 04:30 UTC
+
+## arxiv underwater 3D (7 derniers jours)
+- UW-3DGS: Underwater 3D Reconstruction, Physics-Aware Gaussian Splatting (arxiv 2508.06169)
+- Visual enhancement + 3D representation underwater: review (arxiv 2505.01869)
+
+## arxiv AUV SLAM / point cloud
+- VISO: Robust Underwater Visual-Inertial-Sonar SLAM (arxiv 2601.01144) — VIS+sonar, fort intérêt pour pipeline USBL
+- RUSSO: Underwater SLAM stéréo+sonar+IMU (arxiv 2503.01434)
+- VIMS: Visual-Inertial-Magnetic-Sonar SLAM (arxiv 2506.15126)
+
+## Repos GitHub actifs
+- naver/dust3r (7k★): actif, base pipeline lingbot-map
+- Junyi42/monst3r (ICLR 2025): géométrie vidéo dynamique
+- facebookresearch/vggt (CVPR 2025 Best Paper): reconstruction per-frame
+- CUT3R: Continuous 3D Perception, mise à jour mars 2026
+
+## HuggingFace
+- Video-Depth-Anything-Small: depth video temps-réel
+- StereoAdapter: adaptation profondeur stéréo sous-marine
+
+## Signal fort
+VISO (arxiv 2601.01144): pipeline USBL+caméra+IMU pour AUV, pourrait remplacer pure-camera pose estimation dans stage 06_align.
--- a/pipeline/veille/2026-05-12-1650-iter-4.md
+++ b/pipeline/veille/2026-05-12-1650-iter-4.md
@@ -0,0 +1,26 @@
+# Veille iter-4 — 2026-05-12 16:50 UTC
+
+## Top signaux (8-9/10)
+
+- **ReefMapGS** arxiv.org/abs/2604.11992 — SLAM+3DGS 700m AUV, COLMAP-free, directement applicable COSMA (9/10)
+- **OceanSplat** (2026) — 3D Gaussian Splatting milieu turbide + trinocular consistency (9/10)
+- **BIND-USBL** arxiv.org/abs/2604.11861 — fusion IMU+USBL hétérogène ASV-AUV, delayed fusion = pattern réutilisable stage 06_align (9/10)
+- **LingBot-Map update** (27 avril) — keyframe_interval fix + long-video demo — update recommandé (8/10)
+- **PAS3R** HuggingFace — Pose-Adaptive Streaming 3D, long video = streaming AUV (8/10)
+- **AI-Aided AUV Navigation** arxiv.org/abs/2605.04672 — fusion INS+DVL+cam deep learning (8/10)
+
+## Signaux modérés (7/10)
+
+- Aquatic Neuromorphic Optical Flow arxiv.org/abs/2605.07653 — event cam AUV turbide
+- WaterSplat-SLAM RAL 2026 — SLAM monoculaire sous-marin photoréaliste
+
+## Repos actifs
+
+- lingbot-map (keyframe fix avril), awesome-dust3r (ecosystem DUSt3R/VGGT/CUT3R)
+- Matisse Ifremer — datasets flotte française
+
+## Recommandations
+
+1. **BIND-USBL** : lire pour stage 06_align (pattern fusion USBL+IMU déjà dispo)
+2. **LingBot-Map update** : Already up to date. sur .84/.87 avant prochaine iter
+3. **ReefMapGS** : évaluer comme alternative stage 06_align si PR #9/#12 mergés
--- a/scripts/viser_auv.py
+++ b/scripts/viser_auv.py
@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+"""Open viser viewer with all PLYs from one AUV.
+
+Usage:
+    viser_auv.py --ply-dir /path/to/auv/ply --port 9210
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+import time
+from pathlib import Path
+
+import numpy as np
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--ply-dir", required=True)
+    ap.add_argument("--port", type=int, default=9210)
+    ap.add_argument("--point-size", type=float, default=0.01)
+    ap.add_argument("--max-points-per-ply", type=int, default=1_500_000)
+    args = ap.parse_args()
+
+    try:
+        import open3d as o3d
+        import viser
+    except ImportError as e:
+        sys.exit(f"missing dep: {e}")
+
+    ply_dir = Path(args.ply_dir)
+    plys = sorted(ply_dir.glob("**/*.ply"))
+    print(f"Found {len(plys)} PLY files in {ply_dir}", flush=True)
+    if not plys:
+        sys.exit("no PLY found")
+
+    server = viser.ViserServer(host="0.0.0.0", port=args.port)
+    palette = [
+        (1.0, 0.30, 0.30), (0.30, 1.0, 0.30), (0.30, 0.55, 1.0),
+        (1.0, 0.85, 0.20), (1.0, 0.30, 1.0), (0.30, 1.0, 1.0),
+        (1.0, 0.55, 0.20), (0.55, 0.30, 1.0),
+    ]
+
+    for i, p in enumerate(plys):
+        pcd = o3d.io.read_point_cloud(str(p))
+        pts = np.asarray(pcd.points, dtype=np.float32)
+        if len(pts) == 0:
+            print(f"  ! {p.name}: empty", flush=True)
+            continue
+        if pcd.has_colors():
+            cols = np.asarray(pcd.colors, dtype=np.float32)
+        else:
+            cols = np.tile(palette[i % len(palette)], (len(pts), 1)).astype(np.float32)
+        if len(pts) > args.max_points_per_ply:
+            idx = np.random.choice(len(pts), args.max_points_per_ply, replace=False)
+            pts = pts[idx]
+            cols = cols[idx]
+        # viser wants uint8 colors
+        cols_u8 = (cols * 255).clip(0, 255).astype(np.uint8)
+        name = f"/{p.parent.name}_{p.stem}"
+        server.scene.add_point_cloud(
+            name=name, points=pts, colors=cols_u8, point_size=args.point_size
+        )
+        print(f"  + {p.name}: {len(pts):,} pts", flush=True)
+
+    print(f"Viser ready on port {args.port}", flush=True)
+    while True:
+        time.sleep(60)
+
+
+if __name__ == "__main__":
+    main()
Author	SHA1	Message	Date
Ubuntu	81752163d2	Merge branch 'fix/05-inference-yaml-params' into feature/auto-pipeline	2026-05-12 22:46:30 +00:00
Poulpe	c06dd774ac	auto-iter 2026-05-12: log iter-4 + veille	2026-05-12 16:43:05 +00:00
Poulpe	3a6b058f0d	fix: 05_inference.py lit thresholds.yaml[inference] au lieu de windowed hardcodé - Ajoute _load_inference_cfg() qui lit config/thresholds.yaml - Mode/conf/keyframe_interval/max_frame_num depuis config (streaming par défaut) - Valide par GX049839_v2: streaming+conf=1.5+kf=1 → 146M pts vs 0 pts en windowed sans conf_threshold - Ajoute --offload_to_cpu (stable sur RTX 3090 .84)	2026-05-12 16:38:33 +00:00
Poulpe	8880c28af9	auto-iter 2026-05-12: keyframe_interval 6→1 (streaming, validé GX049839_v2 146M pts)	2026-05-12 16:37:06 +00:00
Poulpe	df45fd155d	auto-iter 2026-05-12: bottom_visible_pct_min 30→25 (GX019817 29% récupérable)	2026-05-12 10:34:19 +00:00
Poulpe	f0154d7ea5	auto-iter 2026-05-12: log iter-2 + veille	2026-05-12 04:40:56 +00:00
Poulpe	8b826b0827	auto-iter 2026-05-12: fix duplicate frame_extract key in thresholds.yaml	2026-05-12 04:39:03 +00:00
Poulpe	4f54d58cd3	auto-iter 2026-05-11: iteration-log + veille iter-1	2026-05-11 22:34:42 +00:00
Poulpe	06d4aa5d4d	auto-iter 2026-05-11: bottom_visible_pct seuil 50→30 (avg=37.5%)	2026-05-11 22:33:12 +00:00
Poulpe	e09ef7886b	feat(pipeline): stage 04b port trim_above_water from dispatcher	2026-05-11 14:08:30 +00:00
Ubuntu	82f71fcc96	feat: frame QC scoring + viser per-AUV button Stage 04 frame extract: - New lib_frame_qc.py: per-frame Laplacian/contrast/blue-dominance scoring - Classes: bottom_visible / water_no_bottom / turbid_water / out_of_water - Sample 1/5 frames after extraction, write qc.json per segment - Record metrics (frames_total, frames_bottom_visible, bottom_visible_pct) - Mark job degraded when bottom_visible_pct < 50% Per-AUV viser view: - scripts/viser_auv.py loads all PLYs of an AUV, color per file - POST /pipeline/missions/{id}/auvs/{auv}/view rsyncs ply -> worker - launches viser on hashed port 9300+, returns URL - _pipeline.html exposes AUV list, JS handler opens viser tab	2026-05-11 11:05:37 +00:00
Ubuntu	1a4fffd2c1	feat: pipeline monitor + orchestrator stats dashboard	2026-05-11 10:55:44 +00:00
Ubuntu	e597407ee5	feat(pipeline): jalon 1-3 — ingest, USBL parse, filter Stages 01-03 opérationnels sur 20260505-Lepradet: - 01_ingest: manifest auto, 3 AUVs vidéo, 3 AUVs bags, mapping AUV2xx↔AUV0xx - 02_usbl_parse: MCAP (format incompatible firmware) → fallback serial CSV, 213 pts bruts - 03_usbl_filter: MAD-3σ + moving-avg + Kalman optionnel, dégradé gracieux si null lat/lon - orchestrator/db.py: SQLite schema missions/jobs/metrics idempotent - config/: thresholds.yaml + default_params.yaml versionnés - qa/checks.py: vérifications pass/fail/degraded par étape Note: MCAP bags corrompus ou format non-standard firmware — lat/lon absent. Statut degraded (pas crash). Nécessite investigation format MCAP spécifique.	2026-05-11 10:25:27 +00:00