# Phase 158 local completion evidence

Date: 2026-07-15

## Locally actionable gaps closed

### 158.7.10 — binary closed captions

`neith-vse-render::caption::write_cc` now emits RCWT v0.001 timed binary caption
data containing genuine CC1 EIA/CEA-608 pop-on commands, odd-parity character
pairs, bottom-row PAC placement, repeated control pairs, and exact cue-start/end
timestamps. It rejects negative/non-positive timing, overlapping CC1 cues, more
than 15 rows, more than 32 columns, and characters outside the implemented
EIA-608 basic set instead of performing a lossy conversion.

The byte-exact test covers the RCWT header, little-endian timing clusters,
caption-block counts, RCL/ENM/PAC/EOC/EDM commands, parity, and payload. The
implementation follows the public
[CCExtractor RCWT v0.001 specification](https://github.com/CCExtractor/ccextractor/blob/master/docs/BINARY_FILE_FORMAT.TXT)
and the interoperable
[FFmpeg RCWT muxer](https://github.com/FFmpeg/FFmpeg/blob/master/libavformat/rcwtenc.c).

### 158.9.2 / 158.11.4 — binary AAF round-trip

The phase checklist still described an old fail-loud binary seam. Current code
already writes and reads an MS-CFB v3 structured-storage `.aaf`, validates the
header/FAT/directory/payload bounds, and rejects corrupt containers. The
cross-crate fidelity gate now uses that binary path—not the JSON helper—and
proves exact video and audio record placement/source-in values at both 24 fps
and 30000/1001.

### 158.11.3 — Blender VSE parity

The host carries Blender 4.0.2. The new ignored release test launches Blender
with a hermetic Python fixture, has Blender construct and inspect the canonical
three-strip cut, converts Blender's one-based record frames to Neith's
zero-based timeline, and compares names, starts, durations, and source-in values
with the Rust `reference_cut`. It passed through the real Blender executable.

### 158.11.8 — live DCC scene-strip update

A second Blender-backed release test creates a VSE scene strip, mutates the
source scene after the strip is linked, updates Blender's view layer, and reads
the changed transform and content revision back through the existing strip's
scene reference. It also verifies that the observed Blender scene/camera
identity maps exactly to `neith_vse_ingest::scene_strip`. It passed through
Blender 4.0.2 without rebuilding or relinking the strip.

### 158.11.7 — executable wall-clock soak gate

`run_timed_soak` now repeatedly creates, stresses, and drops bounded timelines
for a requested wall-clock duration, checks structural invariants after every
edit, varies deterministic seeds between batches, and samples Linux resident
memory after allocator warm-up. The exact ignored release test requests eight
hours and fails if RSS grows by more than 64 MiB. A short default test exercises
the identical timed/RSS path. Multi-seed execution exposed and fixed a real
fixture defect where manual overwrite IDs could collide with IDs internally
allocated by three-point edits.

The checkbox remains open because the literal eight-hour command has not been
allowed to claim success without running for eight hours:

```bash
cargo test --manifest-path libs/neith/vse-testing/Cargo.toml \
  soak::tests::eight_hour_edit_session_has_no_resident_set_growth \
  -- --ignored --exact --nocapture
```

## Remaining non-local gates

- **158.11.1 / 158.11.2:** this 15 GiB / 4-vCPU host exposes a Red Hat Virtio
  GPU. `/dev/dri/renderD128` exists, but FFmpeg VAAPI device initialization
  returns an I/O error; no NVIDIA device exists. There is no real 4K/8K fixture
  in the repository. Consequently this host cannot supply accelerated 4K ≥60 fps
  or full-quality 8K ≥24 fps evidence, and both thresholds remain unchecked.
- **158.11.10:** cutting every Oshun marketing video in this VSE is an
  organizational content-production rollout, not a repository implementation or
  locally executable software gate. It remains unchecked.

## Verification

The authoritative command output and commit record show:

- all eleven phase-158 crate test suites pass;
- `neith-vse-render`: 70 tests pass;
- `neith-vse-conform`: 51 tests pass, including CFB binary round-trip and
  corruption rejection;
- `neith-vse-testing`: 18 default tests pass, with the literal eight-hour gate
  intentionally ignored;
- the Blender parity and live-scene ignored tests each pass when run explicitly;
- Clippy with `--all-targets -- -D warnings` passes for render, conform, and
  testing;
- targeted release builds pass for render, conform, and testing;
- formatting passes for every changed Rust source file.
