iter36 close: env-gated DIAG removed, 5/5 PASS retained

Backend cleanup: -131 lines (iter29/30/33/35 env-gated dump probes).
Load-bearing fixes (α-25, α-29, α-30) untouched. Framework env knobs
(LIBVA_V4L2_DUMP_OUTPUT, etc.) retained.

5-codec regression post-cleanup: all 5 PASS with identical hashes to
pre-cleanup. Backend tip 7db15a5.
This commit is contained in:
2026-05-14 18:13:39 +00:00
parent 7c06c519e7
commit 7e3eadf983
+43
View File
@@ -0,0 +1,43 @@
## Iteration 36 — Phase 8 (close): Remove env-gated DIAG probes
Closes 2026-05-14, seventh campaign-day milestone after iter35 (MPEG-2 verification). User requested "clean up the env-gated diagnostics".
### Removed (env-gated, no behavior change with env unset)
- **iter29** `LIBVA_HEVC_DUMP_SLICE_TAIL` in `src/h265.c` — refuted 40-byte slice_data inflation theory before α-29 fix
- **iter30** `LIBVA_TS_SCALE` in `src/picture.c` — refuted timestamp magnitude theory before α-29 fix
- **iter33** `LIBVA_VP8_DUMP_FRAME` in `src/vp8.c` — led to discovering α-30 fix (header strip)
- **iter35** `LIBVA_MPEG2_DUMP_FRAME` in `src/mpeg2.c` — confirmed libva MPEG-2 ctrls are semantically correct
Total: **-131 lines / +7 lines** (one α-7 comment refresh to remove the iter30 reference).
### Retained (framework-level, preexisting before this campaign)
- `LIBVA_V4L2_DUMP_OUTPUT` (picture.c α-16) — bitstream OUTPUT dump; useful for future debugging, was instrumental in α-30 discovery
- `LIBVA_V4L2_DUMP_CAPTURE` (surface.c) — sibling CAPTURE dump
- `LIBVA_V4L2_ZERO_CAPTURE` (picture.c) — debugging knob to zero CAPTURE buffers
- `LIBVA_V4L2_REQUEST_VIDEO_PATH` / `MEDIA_PATH` / `NO_AUTODETECT` (request.c) — required env config for device override
### Regression test post-cleanup (kernel 7.0-14)
| Codec | libva sha-16 | kdirect sha-16 | Status |
|---|---|---|---|
| H.264 | dd4f5f2d552c07bc | same | PASS |
| HEVC | 108f925bb6cbb6c9 | same | PASS |
| VP9 | cf35908ae0f9ab60 | same | PASS |
| VP8 | d3231e5b6c0ee10b | same | PASS |
| MPEG-2| 95c5905890c937d4 | same | PASS |
5/5 PASS. Identical hashes to pre-cleanup. No regressions.
### Substrate state
- Backend fork tip `7db15a5` (load-bearing fixes α-25 + α-29 + α-30 retained; env-gated DIAG probes removed)
- Kernel `linux-fresnel-fourier 7.0-14` clean (unchanged)
### Open items remaining (post-iter36)
1. α-26 cosmetic revert — `decode_params->short_term_ref_pic_set_size = picture->st_rps_bits` is mis-routed (rkvdec doesn't use that field). Could revert to 0. Cosmetic; no behavior change.
2. Libva multi-device probe — architectural change in libva backend (~200-400 LOC). Single-session H.264/HEVC/VP9 (rkvdec) + VP8/MPEG-2 (hantro) without `LIBVA_V4L2_REQUEST_VIDEO_PATH` override per codec. Design judgment from user welcome.
Campaign in clean ship state. No urgent work.