diff --git a/src/k1link/media_fragments.py b/src/k1link/media_fragments.py
new file mode 100644
index 0000000..8b0b4d7
--- /dev/null
+++ b/src/k1link/media_fragments.py
@@ -0,0 +1,408 @@
+"""Bounded ISO-BMFF timing shared by archive inspection and stream decode.
+
+Pure bytes only: no session store, source paths, materializer or decoder import.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from dataclasses import dataclass
+
+MAX_SAFE_INTEGER = (1 << 53) - 1
+MAX_MP4_BOXES = 100_000
+MAX_MP4_SAMPLES_PER_FRAGMENT = 1_000_000
+
+
+class Mp4IntegrityError(ValueError):
+    """An ISO-BMFF fragment has absent, ambiguous or unbounded timing."""
+
+
+@dataclass(frozen=True, slots=True)
+class VideoTiming:
+    track_id: int
+    timescale: int
+    default_sample_duration: int | None
+    default_sample_flags: int | None
+
+
+@dataclass(frozen=True, slots=True)
+class VideoFragmentTiming:
+    base_decode_time: int
+    duration_units: int
+    random_access: bool
+
+    @property
+    def end_decode_time(self) -> int:
+        return self.base_decode_time + self.duration_units
+
+
+@dataclass(slots=True)
+class ParseBudget:
+    boxes_remaining: int = MAX_MP4_BOXES
+    samples_remaining: int = MAX_MP4_SAMPLES_PER_FRAGMENT
+
+    def consume_box(self) -> None:
+        self.boxes_remaining -= 1
+        if self.boxes_remaining < 0:
+            raise Mp4IntegrityError("recorded media ISO-BMFF box budget was exceeded")
+
+    def consume_samples(self, count: int) -> None:
+        if count < 0 or count > self.samples_remaining:
+            raise Mp4IntegrityError("recorded media ISO-BMFF sample budget was exceeded")
+        self.samples_remaining -= count
+
+
+def video_timing(payload: bytes, budget: ParseBudget) -> VideoTiming:
+    moov_payloads = [
+        box_payload
+        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
+        if box_type == b"moov"
+    ]
+    if len(moov_payloads) != 1:
+        raise Mp4IntegrityError("recorded media init has no unique moov box")
+    moov = moov_payloads[0]
+    moov_boxes = tuple(_iter_mp4_boxes(moov, budget))
+    defaults: dict[int, tuple[int, int]] = {}
+    for box_type, box_payload in moov_boxes:
+        if box_type != b"mvex":
+            continue
+        for child_type, child_payload in _iter_mp4_boxes(box_payload, budget):
+            if child_type != b"trex":
+                continue
+            track_id, trex_default_duration, trex_default_flags = _parse_trex(child_payload)
+            if track_id in defaults:
+                raise Mp4IntegrityError("recorded media init repeats a trex track")
+            defaults[track_id] = (trex_default_duration, trex_default_flags)
+
+    video_tracks: list[tuple[int, int]] = []
+    for box_type, trak_payload in moov_boxes:
+        if box_type != b"trak":
+            continue
+        track_id = _trak_track_id(trak_payload, budget)
+        media = _trak_media_timing(trak_payload, budget)
+        if media is not None:
+            video_tracks.append((track_id, media))
+    if len(video_tracks) != 1:
+        raise Mp4IntegrityError("recorded media init has no unique video track")
+    track_id, timescale = video_tracks[0]
+    default_sample = defaults.get(track_id)
+    default_duration = None if default_sample is None else default_sample[0]
+    return VideoTiming(
+        track_id=track_id,
+        timescale=timescale,
+        default_sample_duration=(
+            default_duration if default_duration is not None and default_duration > 0 else None
+        ),
+        default_sample_flags=None if default_sample is None else default_sample[1],
+    )
+
+
+def _trak_track_id(payload: bytes, budget: ParseBudget) -> int:
+    track_ids = [
+        _parse_tkhd_track_id(box_payload)
+        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
+        if box_type == b"tkhd"
+    ]
+    if len(track_ids) != 1 or track_ids[0] <= 0:
+        raise Mp4IntegrityError("recorded media track id is invalid")
+    return track_ids[0]
+
+
+def _trak_media_timing(payload: bytes, budget: ParseBudget) -> int | None:
+    media_boxes = [
+        box_payload
+        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
+        if box_type == b"mdia"
+    ]
+    if len(media_boxes) != 1:
+        raise Mp4IntegrityError("recorded media track has no unique mdia box")
+    children = tuple(_iter_mp4_boxes(media_boxes[0], budget))
+    handlers = [
+        _parse_hdlr_type(box_payload) for box_type, box_payload in children if box_type == b"hdlr"
+    ]
+    if len(handlers) != 1:
+        raise Mp4IntegrityError("recorded media track handler is ambiguous")
+    if handlers[0] != b"vide":
+        return None
+    timescales = [
+        _parse_mdhd_timescale(box_payload)
+        for box_type, box_payload in children
+        if box_type == b"mdhd"
+    ]
+    if len(timescales) != 1 or timescales[0] <= 0:
+        raise Mp4IntegrityError("recorded media video timescale is invalid")
+    return timescales[0]
+
+
+def video_fragment_timing(
+    payload: bytes,
+    timing: VideoTiming,
+    budget: ParseBudget,
+) -> VideoFragmentTiming:
+    moof_payloads = [
+        box_payload
+        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
+        if box_type == b"moof"
+    ]
+    if len(moof_payloads) != 1:
+        raise Mp4IntegrityError("recorded media fragment has no unique moof box")
+    matching_timings: list[VideoFragmentTiming] = []
+    for box_type, traf_payload in _iter_mp4_boxes(moof_payloads[0], budget):
+        if box_type != b"traf":
+            continue
+        boxes = tuple(_iter_mp4_boxes(traf_payload, budget))
+        tfhd_payloads = [box for kind, box in boxes if kind == b"tfhd"]
+        if len(tfhd_payloads) != 1:
+            raise Mp4IntegrityError("recorded media fragment tfhd is ambiguous")
+        track_id, fragment_default_duration, fragment_default_flags = _parse_tfhd(tfhd_payloads[0])
+        if track_id != timing.track_id:
+            continue
+        tfdt_payloads = [box for kind, box in boxes if kind == b"tfdt"]
+        if len(tfdt_payloads) != 1:
+            raise Mp4IntegrityError("recorded media fragment tfdt is ambiguous")
+        base_decode_time = _parse_tfdt(tfdt_payloads[0])
+        trun_payloads = [box for kind, box in boxes if kind == b"trun"]
+        if not trun_payloads:
+            raise Mp4IntegrityError("recorded media video fragment has no trun box")
+        default_duration = fragment_default_duration or timing.default_sample_duration
+        default_flags = (
+            fragment_default_flags
+            if fragment_default_flags is not None
+            else timing.default_sample_flags
+        )
+        trun_descriptors = tuple(
+            _parse_trun_descriptor(trun, default_duration, default_flags, budget)
+            for trun in trun_payloads
+        )
+        fragment_sample_count = sum(item[2] for item in trun_descriptors)
+        if fragment_sample_count != 1:
+            raise Mp4IntegrityError("recorded media video fragment must contain exactly one sample")
+        duration = sum(item[0] for item in trun_descriptors)
+        if base_decode_time > MAX_SAFE_INTEGER - duration:
+            raise Mp4IntegrityError("recorded media fragment decode time is outside bounds")
+        matching_timings.append(
+            VideoFragmentTiming(
+                base_decode_time=base_decode_time,
+                duration_units=duration,
+                random_access=(trun_descriptors[0][1] & 0x00010000) == 0,
+            )
+        )
+    if len(matching_timings) != 1:
+        raise Mp4IntegrityError("recorded media fragment video track is ambiguous")
+    return matching_timings[0]
+
+
+def _iter_mp4_boxes(
+    payload: bytes,
+    budget: ParseBudget,
+) -> Iterator[tuple[bytes, bytes]]:
+    offset = 0
+    payload_length = len(payload)
+    while offset < payload_length:
+        if payload_length - offset < 8:
+            raise Mp4IntegrityError("recorded media ISO-BMFF box is truncated")
+        size = int.from_bytes(payload[offset : offset + 4], "big")
+        box_type = payload[offset + 4 : offset + 8]
+        header_length = 8
+        if size == 1:
+            if payload_length - offset < 16:
+                raise Mp4IntegrityError("recorded media ISO-BMFF box is truncated")
+            size = int.from_bytes(payload[offset + 8 : offset + 16], "big")
+            header_length = 16
+        elif size == 0:
+            size = payload_length - offset
+        if size < header_length or size > payload_length - offset:
+            raise Mp4IntegrityError("recorded media ISO-BMFF box size is invalid")
+        budget.consume_box()
+        end = offset + size
+        yield box_type, payload[offset + header_length : end]
+        offset = end
+
+
+def single_sample_payload(payload: bytes, timing: VideoTiming) -> memoryview:
+    """Extract the reviewed low-delay fMP4 subset without opening a demuxer.
+
+    Exactly moof+mdat, one traf/trun/sample, explicit default sample size,
+    default-base-is-moof, no composition offset, no extra tracks or padding.
+    This intentionally accepts less than the archive timing inspector. Other
+    valid MP4 layouts require explicit qualification, never guessed offsets.
+    Call video_fragment_timing separately to validate continuity and duration.
+    """
+    budget = ParseBudget(128, 1)
+    top = tuple(_iter_mp4_boxes(payload, budget))
+    if [kind for kind, _ in top] != [b"moof", b"mdat"]:
+        raise Mp4IntegrityError("stream fragment requires exactly moof then mdat")
+    moof, mdat = top[0][1], top[1][1]
+    moof_size = int.from_bytes(payload[:4], "big")
+    if moof_size != len(moof) + 8 or moof_size < 8:
+        raise Mp4IntegrityError("stream fragment requires explicit 32-bit box sizes")
+    mdat_size = int.from_bytes(payload[moof_size : moof_size + 4], "big")
+    if mdat_size != len(mdat) + 8 or moof_size + mdat_size != len(payload):
+        raise Mp4IntegrityError("stream mdat size is ambiguous")
+    children = tuple(_iter_mp4_boxes(moof, budget))
+    if [kind for kind, _ in children] != [b"mfhd", b"traf"]:
+        raise Mp4IntegrityError("stream moof requires one mfhd and one traf")
+    boxes = tuple(_iter_mp4_boxes(children[1][1], budget))
+    if [kind for kind, _ in boxes] != [b"tfhd", b"tfdt", b"trun"]:
+        raise Mp4IntegrityError("stream traf layout is unsupported")
+    tfhd, trun = boxes[0][1], boxes[2][1]
+    if len(tfhd) != 20 or tfhd[:4] != b"\x00\x02\x00\x38":
+        raise Mp4IntegrityError("explicit moof-relative sample defaults required")
+    if _read_u32(tfhd, 4, "track id") != timing.track_id:
+        raise Mp4IntegrityError("stream sample track changed")
+    if not mdat or _read_u32(tfhd, 12, "sample size") != len(mdat):
+        raise Mp4IntegrityError("stream sample size does not match mdat")
+    flags = _full_box_flags(trun)
+    if trun[0] != 0 or flags not in (1, 5) or len(trun) != (16 if flags == 5 else 12):
+        raise Mp4IntegrityError("stream requires a single sample without composition offset")
+    if _read_u32(trun, 4, "sample count") != 1:
+        raise Mp4IntegrityError("stream requires exactly one sample")
+    if _read_u32(trun, 8, "data offset") != moof_size + 8:
+        raise Mp4IntegrityError("stream sample offset does not address mdat")
+    return memoryview(payload)[moof_size + 8 :]
+
+
+def _parse_tkhd_track_id(payload: bytes) -> int:
+    version = _full_box_version(payload)
+    offset = 20 if version == 1 else 12 if version == 0 else -1
+    return _read_u32(payload, offset, "tkhd track id")
+
+
+def _parse_mdhd_timescale(payload: bytes) -> int:
+    version = _full_box_version(payload)
+    offset = 20 if version == 1 else 12 if version == 0 else -1
+    return _read_u32(payload, offset, "mdhd timescale")
+
+
+def _parse_hdlr_type(payload: bytes) -> bytes:
+    _full_box_version(payload)
+    if len(payload) < 12:
+        raise Mp4IntegrityError("recorded media hdlr box is truncated")
+    return payload[8:12]
+
+
+def _parse_trex(payload: bytes) -> tuple[int, int, int]:
+    _full_box_version(payload)
+    return (
+        _read_u32(payload, 4, "trex track id"),
+        _read_u32(payload, 12, "trex default sample duration"),
+        _read_u32(payload, 20, "trex default sample flags"),
+    )
+
+
+def _parse_tfhd(payload: bytes) -> tuple[int, int | None, int | None]:
+    flags = _full_box_flags(payload)
+    track_id = _read_u32(payload, 4, "tfhd track id")
+    cursor = 8
+    for flag, width in ((0x000001, 8), (0x000002, 4)):
+        if flags & flag:
+            cursor = _advance_box_cursor(payload, cursor, width, "tfhd optional field")
+    default_duration: int | None = None
+    if flags & 0x000008:
+        default_duration = _read_u32(payload, cursor, "tfhd default sample duration")
+        cursor += 4
+    if flags & 0x000010:
+        cursor = _advance_box_cursor(payload, cursor, 4, "tfhd default sample size")
+    default_flags: int | None = None
+    if flags & 0x000020:
+        default_flags = _read_u32(payload, cursor, "tfhd default sample flags")
+    return (
+        track_id,
+        default_duration if default_duration and default_duration > 0 else None,
+        default_flags,
+    )
+
+
+def _parse_tfdt(payload: bytes) -> int:
+    version = _full_box_version(payload)
+    if version == 0:
+        return _read_u32(payload, 4, "tfdt base decode time")
+    if version == 1:
+        if len(payload) < 12:
+            raise Mp4IntegrityError("recorded media tfdt base decode time is truncated")
+        return int.from_bytes(payload[4:12], "big")
+    raise Mp4IntegrityError("recorded media tfdt version is unsupported")
+
+
+def _parse_trun_descriptor(
+    payload: bytes,
+    default_duration: int | None,
+    default_sample_flags: int | None,
+    budget: ParseBudget,
+) -> tuple[int, int, int]:
+    flags = _full_box_flags(payload)
+    sample_count = _read_u32(payload, 4, "trun sample count")
+    if sample_count < 1:
+        raise Mp4IntegrityError("recorded media trun has no samples")
+    budget.consume_samples(sample_count)
+    cursor = 8
+    if flags & 0x000001:
+        cursor = _advance_box_cursor(payload, cursor, 4, "trun data offset")
+    first_sample_flags: int | None = None
+    if flags & 0x000004:
+        first_sample_flags = _read_u32(payload, cursor, "trun first sample flags")
+        cursor += 4
+    if first_sample_flags is not None and flags & 0x000400:
+        raise Mp4IntegrityError("recorded media trun sample flags are ambiguous")
+
+    duration = 0
+    first_per_sample_flags: int | None = None
+    for sample_index in range(sample_count):
+        if flags & 0x000100:
+            sample_duration = _read_u32(payload, cursor, "trun sample duration")
+            if sample_duration <= 0:
+                raise Mp4IntegrityError("recorded media sample duration is invalid")
+            duration += sample_duration
+            cursor += 4
+        elif default_duration is None or default_duration <= 0:
+            raise Mp4IntegrityError("recorded media sample duration is unavailable")
+        else:
+            duration += default_duration
+        if flags & 0x000200:
+            cursor = _advance_box_cursor(payload, cursor, 4, "trun sample size")
+        if flags & 0x000400:
+            sample_flags = _read_u32(payload, cursor, "trun sample flags")
+            if sample_index == 0:
+                first_per_sample_flags = sample_flags
+            cursor += 4
+        if flags & 0x000800:
+            cursor = _advance_box_cursor(
+                payload,
+                cursor,
+                4,
+                "trun sample composition time offset",
+            )
+
+    effective_flags = (
+        first_per_sample_flags
+        if first_per_sample_flags is not None
+        else first_sample_flags
+        if first_sample_flags is not None
+        else default_sample_flags
+    )
+    if effective_flags is None:
+        raise Mp4IntegrityError("recorded media sample flags are unavailable")
+    return duration, effective_flags, sample_count
+
+
+def _full_box_version(payload: bytes) -> int:
+    if len(payload) < 4:
+        raise Mp4IntegrityError("recorded media full box is truncated")
+    return payload[0]
+
+
+def _full_box_flags(payload: bytes) -> int:
+    _full_box_version(payload)
+    return int.from_bytes(payload[1:4], "big")
+
+
+def _read_u32(payload: bytes, offset: int, description: str) -> int:
+    if offset < 0 or offset + 4 > len(payload):
+        raise Mp4IntegrityError(f"recorded media {description} is truncated")
+    return int.from_bytes(payload[offset : offset + 4], "big")
+
+
+def _advance_box_cursor(payload: bytes, cursor: int, width: int, description: str) -> int:
+    if cursor < 0 or width < 0 or cursor + width > len(payload):
+        raise Mp4IntegrityError(f"recorded media {description} is truncated")
+    return cursor + width
diff --git a/src/k1link/perception/streaming_decoder.py b/src/k1link/perception/streaming_decoder.py
new file mode 100644
index 0000000..bf8bbe4
--- /dev/null
+++ b/src/k1link/perception/streaming_decoder.py
@@ -0,0 +1,110 @@
+"""One-sample fMP4 → persistent H.264 decoder, no recording/file dependency.
+
+Reuses the archive's bounded ISO-BMFF timing validator, not its materializer.
+Only the supplied init and current complete fragment are inspected. Codec state
+persists across fragments; no later fragment or source EOF is supplied to make
+the current frame appear. Other codecs/reordered/multi-sample streams fail
+closed and require a different explicit profile. Run inside a supervised,
+memory-limited CPU child; native decoder allocations are not Python byte counts.
+"""
+
+from __future__ import annotations
+
+import importlib
+import io
+from fractions import Fraction
+from typing import Any
+
+from k1link.media_fragments import (
+    ParseBudget,
+    single_sample_payload,
+    video_fragment_timing,
+    video_timing,
+)
+
+
+class StreamingDecodeError(ValueError):
+    pass
+
+
+class FragmentDecoder:
+    def __init__(self) -> None:
+        self._av = importlib.import_module("av")
+        if self._av.__version__ != "18.0.0":
+            raise StreamingDecodeError("unqualified PyAV version")
+        # Import before readiness, not on the first source frame's deadline.
+        importlib.import_module("numpy")
+        self._codec: Any = None
+        self._init: bytes | None = None
+        self._timing: Any = None
+        self._next_dts: int | None = None
+        self.frames = 0
+        self.failed = False
+
+    def configure(self, payload: bytes) -> None:
+        if self._init is not None or self.failed or not 0 < len(payload) <= 65536:
+            self.failed = True
+            raise StreamingDecodeError("invalid or repeated codec initialization")
+        try:
+            timing = video_timing(payload, ParseBudget(128, 1))
+            with self._av.open(io.BytesIO(payload), format="mp4", mode="r") as container:
+                if len(container.streams) != 1 or len(container.streams.video) != 1:
+                    raise StreamingDecodeError("one video track is required")
+                description = container.streams.video[0].codec_context
+                if description.name != "h264" or (description.width, description.height) != (
+                    800,
+                    600,
+                ):
+                    raise StreamingDecodeError("codec or calibrated native raster changed")
+                extra = description.extradata
+                if not extra or len(extra) > 65536:
+                    raise StreamingDecodeError("bounded H264 codec configuration required")
+                codec = self._av.CodecContext.create("h264", "r")
+                codec.extradata = extra
+                codec.thread_count = 1
+                codec.thread_type = "SLICE"
+                codec.open()
+            self._codec, self._init, self._timing = codec, bytes(payload), timing
+        except Exception:
+            self.failed = True
+            raise
+
+    def decode(self, payload: bytes) -> Any:
+        if self.failed or self._init is None or not 0 < len(payload) <= 1024 * 1024:
+            self.failed = True
+            raise StreamingDecodeError("uninitialized, failed or oversized fragment")
+        try:
+            timing = video_fragment_timing(payload, self._timing, ParseBudget(128, 1))
+            if self.frames == 0 and not timing.random_access:
+                raise StreamingDecodeError("epoch must start at a random-access fragment")
+            if self._next_dts is not None and timing.base_decode_time != self._next_dts:
+                raise StreamingDecodeError("fragment codec timeline is discontinuous")
+            # Opening a new demuxer per fragment repeats FFmpeg stream probing.
+            # Validate the narrow explicit layout, then feed the one AVCC sample
+            # to the persistent codec. No future packet or flush is supplied.
+            packet = self._av.Packet(single_sample_payload(payload, self._timing))
+            packet.pts = packet.dts = timing.base_decode_time
+            packet.duration = timing.duration_units
+            packet.time_base = Fraction(1, self._timing.timescale)
+            frames = self._codec.decode(packet)
+            if len(frames) != 1 or self._codec.has_b_frames:
+                raise StreamingDecodeError("current frame requires future input or reordering")
+            frame = frames[0]
+            if (frame.width, frame.height) != (800, 600) or frame.pts != packet.pts:
+                raise StreamingDecodeError("decoded raster or frame identity changed")
+            if self.frames == 0 and not frame.key_frame:
+                raise StreamingDecodeError("decoder did not confirm the initial keyframe")
+            image = frame.to_ndarray(format="bgr24")
+            if image.shape != (600, 800, 3) or image.nbytes != 1_440_000:
+                raise StreamingDecodeError("invalid native decoder output")
+            image.setflags(write=False)
+            self._next_dts = timing.base_decode_time + timing.duration_units
+            self.frames += 1
+            return image
+        except Exception:
+            self.failed = True
+            raise
+
+    def close(self) -> None:
+        self.failed = True
+        self._codec = self._init = self._timing = None
diff --git a/src/k1link/sessions/media.py b/src/k1link/sessions/media.py
index 26424fb..6f5a0e3 100644
--- a/src/k1link/sessions/media.py
+++ b/src/k1link/sessions/media.py
@@ -8,12 +8,14 @@ import re
 import secrets
 import stat
 import threading
-from collections.abc import Iterable, Iterator
+from collections.abc import Iterable
 from contextlib import suppress
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, TypeGuard
 
+from k1link import media_fragments as _fragment_timing
+
 from .models import RecordedMediaArtifact, ReplayCommand, SessionIntegrityError
 
 CAMERA_ARCHIVE_SCHEMA = "missioncore.camera-recording/v1"
@@ -25,8 +27,8 @@ MAX_MEDIA_INDEX_LINE_BYTES = 64 * 1024
 MAX_INIT_BYTES = 8 * 1024 * 1024
 MAX_MEDIA_SEGMENT_BYTES = 64 * 1024 * 1024
 MAX_SAFE_INTEGER = (1 << 53) - 1
-MAX_MP4_BOXES = 100_000
-MAX_MP4_SAMPLES_PER_FRAGMENT = 1_000_000
+MAX_MP4_BOXES = _fragment_timing.MAX_MP4_BOXES
+MAX_MP4_SAMPLES_PER_FRAGMENT = _fragment_timing.MAX_MP4_SAMPLES_PER_FRAGMENT
 MAX_MP4_FRAGMENT_DURATION_SECONDS = 3_600.0
 MEDIA_RECORDING_TIMELINE_TOLERANCE_SECONDS = 0.05
 MAX_MEDIA_EPOCHS = 4_096
@@ -88,39 +90,9 @@ class _CachedManifest:
     manifest: RecordedMediaManifest
 
 
-@dataclass(frozen=True, slots=True)
-class _Mp4VideoTiming:
-    track_id: int
-    timescale: int
-    default_sample_duration: int | None
-    default_sample_flags: int | None
-
-
-@dataclass(frozen=True, slots=True)
-class _Mp4VideoFragmentTiming:
-    base_decode_time: int
-    duration_units: int
-    random_access: bool
-
-    @property
-    def end_decode_time(self) -> int:
-        return self.base_decode_time + self.duration_units
-
-
-@dataclass(slots=True)
-class _Mp4ParseBudget:
-    boxes_remaining: int = MAX_MP4_BOXES
-    samples_remaining: int = MAX_MP4_SAMPLES_PER_FRAGMENT
-
-    def consume_box(self) -> None:
-        self.boxes_remaining -= 1
-        if self.boxes_remaining < 0:
-            raise SessionIntegrityError("recorded media ISO-BMFF box budget was exceeded")
-
-    def consume_samples(self, count: int) -> None:
-        if count < 0 or count > self.samples_remaining:
-            raise SessionIntegrityError("recorded media ISO-BMFF sample budget was exceeded")
-        self.samples_remaining -= count
+_Mp4VideoTiming = _fragment_timing.VideoTiming
+_Mp4VideoFragmentTiming = _fragment_timing.VideoFragmentTiming
+_Mp4ParseBudget = _fragment_timing.ParseBudget
 
 
 class RecordedMediaInspector:
@@ -1155,85 +1127,10 @@ def _checked_fragment_duration_seconds(duration_units: int, timescale: int) -> f
 
 
 def _mp4_video_timing(payload: bytes, budget: _Mp4ParseBudget) -> _Mp4VideoTiming:
-    moov_payloads = [
-        box_payload
-        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
-        if box_type == b"moov"
-    ]
-    if len(moov_payloads) != 1:
-        raise SessionIntegrityError("recorded media init has no unique moov box")
-    moov = moov_payloads[0]
-    moov_boxes = tuple(_iter_mp4_boxes(moov, budget))
-    defaults: dict[int, tuple[int, int]] = {}
-    for box_type, box_payload in moov_boxes:
-        if box_type != b"mvex":
-            continue
-        for child_type, child_payload in _iter_mp4_boxes(box_payload, budget):
-            if child_type != b"trex":
-                continue
-            track_id, trex_default_duration, trex_default_flags = _parse_trex(child_payload)
-            if track_id in defaults:
-                raise SessionIntegrityError("recorded media init repeats a trex track")
-            defaults[track_id] = (trex_default_duration, trex_default_flags)
-
-    video_tracks: list[tuple[int, int]] = []
-    for box_type, trak_payload in moov_boxes:
-        if box_type != b"trak":
-            continue
-        track_id = _trak_track_id(trak_payload, budget)
-        media = _trak_media_timing(trak_payload, budget)
-        if media is not None:
-            video_tracks.append((track_id, media))
-    if len(video_tracks) != 1:
-        raise SessionIntegrityError("recorded media init has no unique video track")
-    track_id, timescale = video_tracks[0]
-    default_sample = defaults.get(track_id)
-    default_duration = None if default_sample is None else default_sample[0]
-    return _Mp4VideoTiming(
-        track_id=track_id,
-        timescale=timescale,
-        default_sample_duration=(
-            default_duration if default_duration is not None and default_duration > 0 else None
-        ),
-        default_sample_flags=None if default_sample is None else default_sample[1],
-    )
-
-
-def _trak_track_id(payload: bytes, budget: _Mp4ParseBudget) -> int:
-    track_ids = [
-        _parse_tkhd_track_id(box_payload)
-        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
-        if box_type == b"tkhd"
-    ]
-    if len(track_ids) != 1 or track_ids[0] <= 0:
-        raise SessionIntegrityError("recorded media track id is invalid")
-    return track_ids[0]
-
-
-def _trak_media_timing(payload: bytes, budget: _Mp4ParseBudget) -> int | None:
-    media_boxes = [
-        box_payload
-        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
-        if box_type == b"mdia"
-    ]
-    if len(media_boxes) != 1:
-        raise SessionIntegrityError("recorded media track has no unique mdia box")
-    children = tuple(_iter_mp4_boxes(media_boxes[0], budget))
-    handlers = [
-        _parse_hdlr_type(box_payload) for box_type, box_payload in children if box_type == b"hdlr"
-    ]
-    if len(handlers) != 1:
-        raise SessionIntegrityError("recorded media track handler is ambiguous")
-    if handlers[0] != b"vide":
-        return None
-    timescales = [
-        _parse_mdhd_timescale(box_payload)
-        for box_type, box_payload in children
-        if box_type == b"mdhd"
-    ]
-    if len(timescales) != 1 or timescales[0] <= 0:
-        raise SessionIntegrityError("recorded media video timescale is invalid")
-    return timescales[0]
+    try:
+        return _fragment_timing.video_timing(payload, budget)
+    except _fragment_timing.Mp4IntegrityError as exc:
+        raise SessionIntegrityError(str(exc)) from exc
 
 
 def _mp4_video_fragment_timing(
@@ -1241,232 +1138,10 @@ def _mp4_video_fragment_timing(
     timing: _Mp4VideoTiming,
     budget: _Mp4ParseBudget,
 ) -> _Mp4VideoFragmentTiming:
-    moof_payloads = [
-        box_payload
-        for box_type, box_payload in _iter_mp4_boxes(payload, budget)
-        if box_type == b"moof"
-    ]
-    if len(moof_payloads) != 1:
-        raise SessionIntegrityError("recorded media fragment has no unique moof box")
-    matching_timings: list[_Mp4VideoFragmentTiming] = []
-    for box_type, traf_payload in _iter_mp4_boxes(moof_payloads[0], budget):
-        if box_type != b"traf":
-            continue
-        boxes = tuple(_iter_mp4_boxes(traf_payload, budget))
-        tfhd_payloads = [box for kind, box in boxes if kind == b"tfhd"]
-        if len(tfhd_payloads) != 1:
-            raise SessionIntegrityError("recorded media fragment tfhd is ambiguous")
-        track_id, fragment_default_duration, fragment_default_flags = _parse_tfhd(tfhd_payloads[0])
-        if track_id != timing.track_id:
-            continue
-        tfdt_payloads = [box for kind, box in boxes if kind == b"tfdt"]
-        if len(tfdt_payloads) != 1:
-            raise SessionIntegrityError("recorded media fragment tfdt is ambiguous")
-        base_decode_time = _parse_tfdt(tfdt_payloads[0])
-        trun_payloads = [box for kind, box in boxes if kind == b"trun"]
-        if not trun_payloads:
-            raise SessionIntegrityError("recorded media video fragment has no trun box")
-        default_duration = fragment_default_duration or timing.default_sample_duration
-        default_flags = (
-            fragment_default_flags
-            if fragment_default_flags is not None
-            else timing.default_sample_flags
-        )
-        trun_descriptors = tuple(
-            _parse_trun_descriptor(trun, default_duration, default_flags, budget)
-            for trun in trun_payloads
-        )
-        fragment_sample_count = sum(item[2] for item in trun_descriptors)
-        if fragment_sample_count != 1:
-            raise SessionIntegrityError(
-                "recorded media video fragment must contain exactly one sample"
-            )
-        duration = sum(item[0] for item in trun_descriptors)
-        if base_decode_time > MAX_SAFE_INTEGER - duration:
-            raise SessionIntegrityError("recorded media fragment decode time is outside bounds")
-        matching_timings.append(
-            _Mp4VideoFragmentTiming(
-                base_decode_time=base_decode_time,
-                duration_units=duration,
-                random_access=(trun_descriptors[0][1] & 0x00010000) == 0,
-            )
-        )
-    if len(matching_timings) != 1:
-        raise SessionIntegrityError("recorded media fragment video track is ambiguous")
-    return matching_timings[0]
-
-
-def _iter_mp4_boxes(
-    payload: bytes,
-    budget: _Mp4ParseBudget,
-) -> Iterator[tuple[bytes, bytes]]:
-    offset = 0
-    payload_length = len(payload)
-    while offset < payload_length:
-        if payload_length - offset < 8:
-            raise SessionIntegrityError("recorded media ISO-BMFF box is truncated")
-        size = int.from_bytes(payload[offset : offset + 4], "big")
-        box_type = payload[offset + 4 : offset + 8]
-        header_length = 8
-        if size == 1:
-            if payload_length - offset < 16:
-                raise SessionIntegrityError("recorded media ISO-BMFF box is truncated")
-            size = int.from_bytes(payload[offset + 8 : offset + 16], "big")
-            header_length = 16
-        elif size == 0:
-            size = payload_length - offset
-        if size < header_length or size > payload_length - offset:
-            raise SessionIntegrityError("recorded media ISO-BMFF box size is invalid")
-        budget.consume_box()
-        end = offset + size
-        yield box_type, payload[offset + header_length : end]
-        offset = end
-
-
-def _parse_tkhd_track_id(payload: bytes) -> int:
-    version = _full_box_version(payload)
-    offset = 20 if version == 1 else 12 if version == 0 else -1
-    return _read_u32(payload, offset, "tkhd track id")
-
-
-def _parse_mdhd_timescale(payload: bytes) -> int:
-    version = _full_box_version(payload)
-    offset = 20 if version == 1 else 12 if version == 0 else -1
-    return _read_u32(payload, offset, "mdhd timescale")
-
-
-def _parse_hdlr_type(payload: bytes) -> bytes:
-    _full_box_version(payload)
-    if len(payload) < 12:
-        raise SessionIntegrityError("recorded media hdlr box is truncated")
-    return payload[8:12]
-
-
-def _parse_trex(payload: bytes) -> tuple[int, int, int]:
-    _full_box_version(payload)
-    return (
-        _read_u32(payload, 4, "trex track id"),
-        _read_u32(payload, 12, "trex default sample duration"),
-        _read_u32(payload, 20, "trex default sample flags"),
-    )
-
-
-def _parse_tfhd(payload: bytes) -> tuple[int, int | None, int | None]:
-    flags = _full_box_flags(payload)
-    track_id = _read_u32(payload, 4, "tfhd track id")
-    cursor = 8
-    for flag, width in ((0x000001, 8), (0x000002, 4)):
-        if flags & flag:
-            cursor = _advance_box_cursor(payload, cursor, width, "tfhd optional field")
-    default_duration: int | None = None
-    if flags & 0x000008:
-        default_duration = _read_u32(payload, cursor, "tfhd default sample duration")
-        cursor += 4
-    if flags & 0x000010:
-        cursor = _advance_box_cursor(payload, cursor, 4, "tfhd default sample size")
-    default_flags: int | None = None
-    if flags & 0x000020:
-        default_flags = _read_u32(payload, cursor, "tfhd default sample flags")
-    return (
-        track_id,
-        default_duration if default_duration and default_duration > 0 else None,
-        default_flags,
-    )
-
-
-def _parse_tfdt(payload: bytes) -> int:
-    version = _full_box_version(payload)
-    if version == 0:
-        return _read_u32(payload, 4, "tfdt base decode time")
-    if version == 1:
-        if len(payload) < 12:
-            raise SessionIntegrityError("recorded media tfdt base decode time is truncated")
-        return int.from_bytes(payload[4:12], "big")
-    raise SessionIntegrityError("recorded media tfdt version is unsupported")
-
-
-def _parse_trun_descriptor(
-    payload: bytes,
-    default_duration: int | None,
-    default_sample_flags: int | None,
-    budget: _Mp4ParseBudget,
-) -> tuple[int, int, int]:
-    flags = _full_box_flags(payload)
-    sample_count = _read_u32(payload, 4, "trun sample count")
-    if sample_count < 1:
-        raise SessionIntegrityError("recorded media trun has no samples")
-    budget.consume_samples(sample_count)
-    cursor = 8
-    if flags & 0x000001:
-        cursor = _advance_box_cursor(payload, cursor, 4, "trun data offset")
-    first_sample_flags: int | None = None
-    if flags & 0x000004:
-        first_sample_flags = _read_u32(payload, cursor, "trun first sample flags")
-        cursor += 4
-    if first_sample_flags is not None and flags & 0x000400:
-        raise SessionIntegrityError("recorded media trun sample flags are ambiguous")
-
-    duration = 0
-    first_per_sample_flags: int | None = None
-    for sample_index in range(sample_count):
-        if flags & 0x000100:
-            sample_duration = _read_u32(payload, cursor, "trun sample duration")
-            if sample_duration <= 0:
-                raise SessionIntegrityError("recorded media sample duration is invalid")
-            duration += sample_duration
-            cursor += 4
-        elif default_duration is None or default_duration <= 0:
-            raise SessionIntegrityError("recorded media sample duration is unavailable")
-        else:
-            duration += default_duration
-        if flags & 0x000200:
-            cursor = _advance_box_cursor(payload, cursor, 4, "trun sample size")
-        if flags & 0x000400:
-            sample_flags = _read_u32(payload, cursor, "trun sample flags")
-            if sample_index == 0:
-                first_per_sample_flags = sample_flags
-            cursor += 4
-        if flags & 0x000800:
-            cursor = _advance_box_cursor(
-                payload,
-                cursor,
-                4,
-                "trun sample composition time offset",
-            )
-
-    effective_flags = (
-        first_per_sample_flags
-        if first_per_sample_flags is not None
-        else first_sample_flags
-        if first_sample_flags is not None
-        else default_sample_flags
-    )
-    if effective_flags is None:
-        raise SessionIntegrityError("recorded media sample flags are unavailable")
-    return duration, effective_flags, sample_count
-
-
-def _full_box_version(payload: bytes) -> int:
-    if len(payload) < 4:
-        raise SessionIntegrityError("recorded media full box is truncated")
-    return payload[0]
-
-
-def _full_box_flags(payload: bytes) -> int:
-    _full_box_version(payload)
-    return int.from_bytes(payload[1:4], "big")
-
-
-def _read_u32(payload: bytes, offset: int, description: str) -> int:
-    if offset < 0 or offset + 4 > len(payload):
-        raise SessionIntegrityError(f"recorded media {description} is truncated")
-    return int.from_bytes(payload[offset : offset + 4], "big")
-
-
-def _advance_box_cursor(payload: bytes, cursor: int, width: int, description: str) -> int:
-    if cursor < 0 or width < 0 or cursor + width > len(payload):
-        raise SessionIntegrityError(f"recorded media {description} is truncated")
-    return cursor + width
+    try:
+        return _fragment_timing.video_fragment_timing(payload, timing, budget)
+    except _fragment_timing.Mp4IntegrityError as exc:
+        raise SessionIntegrityError(str(exc)) from exc
 
 
 def _mp4_media_type(init_payload: bytes) -> str:
diff --git a/tests/test_perception_streaming_decoder.py b/tests/test_perception_streaming_decoder.py
new file mode 100644
index 0000000..5bc0c3e
--- /dev/null
+++ b/tests/test_perception_streaming_decoder.py
@@ -0,0 +1,231 @@
+"""Synthetic boxes and fake codec only; real decoding belongs on Worker 006."""
+
+import subprocess
+import sys
+from types import SimpleNamespace
+
+import pytest
+
+from k1link import media_fragments as mp4
+from k1link.perception import streaming_decoder as decoder_module
+
+
+def box(kind, data=b""):
+    return (len(data) + 8).to_bytes(4, "big") + kind + data
+
+
+def u32(value):
+    return value.to_bytes(4, "big")
+
+
+def fixture(*, dts=0, sync=True, size=5, offset_delta=0, count=1, trun_flags=1):
+    tkhd = box(b"tkhd", bytes(12) + u32(1))
+    mdhd = box(b"mdhd", bytes(12) + u32(1000))
+    hdlr = box(b"hdlr", bytes(8) + b"vide")
+    init = box(b"moov", box(b"trak", tkhd + box(b"mdia", mdhd + hdlr)))
+    tfhd = box(
+        b"tfhd", b"\x00\x02\x00\x38" + u32(1) + u32(100) + u32(size) + u32(0 if sync else 0x10000)
+    )
+    tfdt = box(b"tfdt", bytes(4) + u32(dts))
+    mfhd = box(b"mfhd", bytes(8))
+
+    def moof(offset):
+        trun = box(b"trun", u32(trun_flags) + u32(count) + u32(offset))
+        return box(b"moof", mfhd + box(b"traf", tfhd + tfdt + trun))
+
+    return init, moof(len(moof(0)) + 8 + offset_delta) + box(b"mdat", b"frame")
+
+
+def test_shared_parser_and_direct_sample_are_source_neutral():
+    init, fragment = fixture()
+    timing = mp4.video_timing(init, mp4.ParseBudget(128, 1))
+    frame = mp4.video_fragment_timing(fragment, timing, mp4.ParseBudget(128, 1))
+    assert (timing.track_id, timing.timescale) == (1, 1000)
+    assert (frame.base_decode_time, frame.duration_units, frame.random_access) == (0, 100, True)
+    sample = mp4.single_sample_payload(fragment, timing)
+    assert sample == b"frame" and sample.obj is fragment
+    script = (
+        "import sys; import k1link.perception.streaming_decoder; "
+        "assert 'k1link.sessions' not in sys.modules; "
+        "assert 'k1link.compute' not in sys.modules; assert 'av' not in sys.modules"
+    )
+    subprocess.run([sys.executable, "-c", script], check=True, timeout=5)
+
+
+@pytest.mark.parametrize(
+    "changes,match",
+    [
+        ({"size": 4}, "size"),
+        ({"offset_delta": 1}, "offset"),
+        ({"count": 2}, "one sample"),
+        ({"trun_flags": 0x801}, "composition"),
+    ],
+)
+def test_direct_sample_rejects_guessed_offsets_or_sizes(changes, match):
+    init, fragment = fixture(**changes)
+    with pytest.raises(mp4.Mp4IntegrityError, match=match):
+        mp4.single_sample_payload(fragment, mp4.video_timing(init, mp4.ParseBudget()))
+
+
+def test_parser_bounds_and_truncation():
+    init, fragment = fixture(count=2)
+    timing = mp4.video_timing(init, mp4.ParseBudget())
+    with pytest.raises(mp4.Mp4IntegrityError, match="sample budget"):
+        mp4.video_fragment_timing(fragment, timing, mp4.ParseBudget(128, 1))
+    with pytest.raises(mp4.Mp4IntegrityError, match="box budget"):
+        mp4.video_timing(init, mp4.ParseBudget(1, 1))
+    with pytest.raises(mp4.Mp4IntegrityError, match="box size"):
+        mp4.single_sample_payload(fragment[:-1], timing)
+    with pytest.raises(mp4.Mp4IntegrityError, match="moof then mdat"):
+        mp4.single_sample_payload(fragment + box(b"free"), timing)
+
+
+@pytest.fixture
+def fake_av(monkeypatch):
+    class Image:
+        shape, nbytes = (600, 800, 3), 1_440_000
+
+        def setflags(self, *, write):
+            assert write is False
+
+    class Codec:
+        has_b_frames = False
+
+        def open(self):
+            pass
+
+        def decode(self, packet):
+            self.calls.append(packet)
+            return (
+                []
+                if self.future
+                else [
+                    SimpleNamespace(
+                        width=self.width,
+                        height=600,
+                        pts=packet.pts + self.pts_delta,
+                        key_frame=self.key_frame,
+                        to_ndarray=lambda **kwargs: Image(),
+                    )
+                ]
+            )
+
+    codec = Codec()
+    codec.calls, codec.future, codec.width, codec.pts_delta, codec.key_frame = (
+        [],
+        False,
+        800,
+        0,
+        True,
+    )
+
+    class Streams(list):
+        @property
+        def video(self):
+            return self
+
+    class Container:
+        def __enter__(self):
+            return SimpleNamespace(
+                streams=Streams(
+                    [
+                        SimpleNamespace(
+                            codec_context=SimpleNamespace(
+                                name="h264", width=800, height=600, extradata=b"config"
+                            )
+                        )
+                    ]
+                )
+            )
+
+        def __exit__(self, *args):
+            pass
+
+    opens = []
+
+    def open_container(stream, **kwargs):
+        opens.append(stream.getvalue())
+        return Container()
+
+    av = SimpleNamespace(
+        __version__="18.0.0",
+        open=open_container,
+        CodecContext=SimpleNamespace(create=lambda *args: codec),
+        Packet=lambda raw: SimpleNamespace(raw=bytes(raw)),
+    )
+    original = decoder_module.importlib.import_module
+    monkeypatch.setattr(
+        decoder_module.importlib,
+        "import_module",
+        lambda name: av if name == "av" else original(name),
+    )
+    return codec, av, opens
+
+
+def test_persistent_codec_emits_current_frame_without_demux_or_flush(fake_av):
+    codec, _, opens = fake_av
+    init, fragment = fixture()
+    decoder = decoder_module.FragmentDecoder()
+    decoder.configure(init)
+    for sequence in range(3):
+        decoder.decode(fixture(dts=sequence * 100)[1])
+    assert len(opens) == 1 and len(codec.calls) == decoder.frames == 3
+    assert codec.thread_count == 1 and codec.thread_type == "SLICE"
+    assert all(p.raw == b"frame" and p.pts == p.dts for p in codec.calls)
+    decoder.close()
+    with pytest.raises(decoder_module.StreamingDecodeError, match="failed"):
+        decoder.decode(fragment)
+
+
+@pytest.mark.parametrize(
+    "mutation,match",
+    [
+        (("future", True), "future"),
+        (("has_b_frames", True), "reordering"),
+        (("width", 801), "raster"),
+        (("pts_delta", 1), "identity"),
+        (("key_frame", False), "keyframe"),
+    ],
+)
+def test_ambiguous_native_output_poisons_decoder(fake_av, mutation, match):
+    codec, _, _ = fake_av
+    setattr(codec, *mutation)
+    decoder = decoder_module.FragmentDecoder()
+    init, fragment = fixture()
+    decoder.configure(init)
+    with pytest.raises(decoder_module.StreamingDecodeError, match=match):
+        decoder.decode(fragment)
+    assert decoder.failed and decoder.frames == 0
+
+
+def test_codec_gap_does_not_use_later_frame_as_current(fake_av):
+    codec, _, _ = fake_av
+    decoder = decoder_module.FragmentDecoder()
+    decoder.configure(fixture()[0])
+    decoder.decode(fixture()[1])
+    with pytest.raises(decoder_module.StreamingDecodeError, match="discontinuous"):
+        decoder.decode(fixture(dts=200)[1])
+    assert len(codec.calls) == 1 and decoder.failed
+
+
+@pytest.mark.parametrize("case", ["no-init", "reinit", "non-keyframe", "oversized", "version"])
+def test_invalid_stream_setup_fails_closed(fake_av, case):
+    _, av, _ = fake_av
+    if case == "version":
+        av.__version__ = "other"
+        with pytest.raises(decoder_module.StreamingDecodeError, match="version"):
+            decoder_module.FragmentDecoder()
+        return
+    decoder = decoder_module.FragmentDecoder()
+    if case != "no-init":
+        decoder.configure(fixture()[0])
+    with pytest.raises(decoder_module.StreamingDecodeError):
+        if case == "reinit":
+            decoder.configure(fixture()[0])
+        else:
+            decoder.decode(
+                b"x" * (1024 * 1024 + 1)
+                if case == "oversized"
+                else fixture(sync=case != "non-keyframe")[1]
+            )
+    assert decoder.failed