Iteration 1 closes with all five Phase 1 boolean-correctness criteria
green. Second codec passes — campaign scoreboard 1/5 → 2/5 (H.264
in T4, MPEG-2 in iter1). Loop terminates per
feedback_dev_process.md Phase 8.
What landed:
Code (libva-v4l2-request-fourier master 65969da..229d6d1):
e7dad7a iter1 Phase 6 commit A: config.c break for MPEG-2 cases
5fe873c iter1 Phase 6 commit B: rewrite mpeg2.c against new V4L2 stateless API
3aab187 iter1 Phase 6 commit C: delete staging-era include/mpeg2-ctrls.h
229d6d1 iter1 Phase 6 commit D: drop missed mpeg2-ctrls.h include from context.c (fix-forward)
All four authored as Claude (noether) per feedback_gitea_as_claude_noether.md.
Campaign docs (fresnel-fourier):
f720c77 iter1 Phase 0 + Phase 1 lock
cc55a6e iter1 Phase 2 situation analysis (3 bugs)
b9625af iter1 Phase 3 baseline measurements (4 baselines)
3e996d0 iter1 Phase 4 plan (6 contract clauses)
0e2e1c2 iter1 Phase 5 sonnet review (6 findings, 4 amendments)
ec9133a iter1 Phase 7 verification (5/5 GREEN)
[this commit] iter1 Phase 8 close
Three lessons distilled to durable memory:
L1 — feedback_header_deletion_check.md
Phase 2 grep audit found 2 of 3 include sites for the
staging-era mpeg2-ctrls.h header; build broke on Commit C
delete because of context.c:42. Rule: when removing a header,
let the compiler enumerate includes authoritatively (clean
rebuild after include-removal patches, before git rm). Grep
is a hint; the compiler is the authority.
L2 — feedback_review_empirical_over_theoretical.md
Phase 5 reviewer S2 flagged a numerical mismatch in
vbv_buffer_size between Baseline C (1.31 MB) and predicted
post-fix (1 MB). I rejected with confident source-read
reasoning (slot->size = sizeimage = matches). Phase 7
byte-compare empirically showed the reviewer was right —
post-fix produces 1 MB, not 1.31 MB. Operational impact nil
(kernel ignores the field), but my Phase 5 rebuttal had a
source-read gap. Rule: when a reviewer cites a concrete
numerical discrepancy, defer to Phase 7 byte-compare; don't
reject on source-read alone.
L3 — feedback_rockchip_pixel_verify_path.md
Iter1 + T4 (H.264) both empirically confirm: libva backend's
vaDeriveImage / cached-mmap readback returns all-zero NV12
on RK3399 — same iter1 patch-0011 cache-coherency bug class
observed on RK3568. Pixel verification must use DMA-BUF GL
import (mpv --vo=image, ffmpeg-v4l2request DRM_PRIME +
hwdownload). NOT ffmpeg-vaapi+hwdownload (cache-stale on
Rockchip). Codec-agnostic; applies to all 5 codecs in
campaign scope.
Backlog items deferred (campaign-internal, not durable memory):
B1: V4L2 /dev/videoN numbering shuffles across reboots
(rkvdec moved video3+media1 → video1+media0 between
Phase 0/3 and Phase 7). Backend should probe /dev/media*
for driver match. Iter2+ Phase 4 cross-cutting candidate.
B2: mpv --hwdec=vaapi-copy silently filters MPEG-2 out before
libva is loaded. mpv --hwdec=vaapi (DMA-BUF) DOES engage.
Phase 1 criterion 3 ended up anchored on ffmpeg-direct.
Mpv-side investigation as separate follow-up.
B3: Latent surface-reuse bug — picture.c:287 h264.matrix_set
reset writes byte 240 of params union, lands inside
mpeg2.iqmatrix.chroma_intra_quantiser_matrix[20] (verified
via offsetof on fresnel via gcc + libva). Per-frame
RenderPicture overwrites this byte for ffmpeg-vaapi flows
that send VAIQMatrixBufferType every frame. Latent for
VAAPI clients that reuse a surface without re-sending
IQMatrix. Iter2+ candidate.
B4: src/context.c:142-155 H.264 device-init runs unconditionally
on every CreateContext, EINVALs on hantro. Intentional
best-effort but request_log fires "Unable to set
control(s)" cosmetically. Suppress-log candidate, low
priority.
B5: vbv_buffer_size = SOURCE_SIZE_MAX (1 MB) rather than
negotiated sizeimage. Kernel ignores. Polish candidate.
Phase 4 cross-cutting work items collected:
- Add VIDIOC_EXPBUF + DMA_BUF_IOCTL_SYNC to libva backend
image-export (fixes L3's vaDeriveImage cache-stale bug for
all codecs).
- V4L2 device-discovery probe (fixes B1).
- Picture.c BeginPicture profile-aware reset (fixes B3).
- Context.c H.264 device-init log suppression (fixes B4).
Campaign roadmap (codec iterations remaining):
iter2: HEVC on rkvdec — re-enable h265.c in build, audit against
rkvdec kernel HEVC contract.
iter3: VP8 on hantro — implement vp8.c.
iter4: VP9 on rkvdec — implement vp9.c (largest control surface).
Phase 5 review S2 historical-record correction: Phase 5 reviewer
was numerically right about vbv_buffer_size. My Phase 5 rebuttal
in phase5_iter1_review.md was empirically wrong. Acknowledged in
phase7_iter1_verification.md and phase8_iteration1_close.md;
Phase 5 doc preserved as-is for the historical record.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
claude-noether