fresnel-fourier/phase6_iter5b_implementation_v2.md

Iteration 5b — Phase 6 v2 (implementation, option β)

Commits A + B + C landed on fork master 2026-05-12 after the Phase 5 v2 review cleared the β plan with two CRIT amendments. Build + install on fresnel pending — fresnel offline at time of commit push.

Commits

SHA	Files	LOC	Summary
1c548b1 (A)	src/codec.h (new) · src/codec.c (new) · src/meson.build	+102 / -2	NEW codec.h/codec.c — pixelformat_for_profile() helper. Identical to reverted iter5b commit A. Registered in meson.build sources + headers.
cc077a0 (B)	src/config.c	+11	Wire object_config->pixelformat = pixelformat_for_profile(profile) at CreateConfig. Populates the previously-dead field at config.h:46.
7055b14 (C)	src/surface.c · src/surface.h · src/context.c · src/request.h	+238 / -280	β refactor + CRIT-1 + CRIT-2 + IMP-1 + IMP-2 all in one. Strips ~200 LOC from surface.c::CreateSurfaces2; absorbs OUTPUT-side lifecycle into context.c::RequestCreateContext; adds request_pool_destroy to DestroyContext; deletes last_output_{width,height} fields + surface_reset_format_cache function.

Fork tip post-iter5b-β: 7055b14.

Phase 5 v2 amendments incorporated

ID	Amendment	Where landed
CRIT-1	Remove video_format == NULL early-return at context.c:65-66	Commit C — restructured CreateContext entry; CAPTURE probe now runs first (within CreateContext), then output_type/capture_type derived from video_format->v4l2_mplane.
CRIT-2	Add request_pool_destroy(&driver_data->output_pool) to DestroyContext before REQBUFS(0)	Commit C — inserted with explanatory comment citing the v2 review.
IMP-1	Probe CAPTURE first to derive output_type, not the hardcoded v4l2_type_video_output(true)	Commit C — IMP-1 ordering applied directly.
IMP-2	Delete dead surface_reset_format_cache function + caller + last_output_* fields	Commit C — deleted from surface.c (function definition + comment block), surface.h (declaration), context.c (call in DestroyContext), request.h (fields).
IMP-3	Document error-recovery state in commit comment	Commit C body explains the error-path state.

All 5 Phase 5 v2 amendments in commit C; no follow-up edits needed.

File-level changes

File	Action	Notes
src/codec.h	NEW (commit A)	Helper declaration.
src/codec.c	NEW (commit A)	Helper definition; switch covers all 5 codec profiles.
src/meson.build	UPDATE (commit A)	Add codec.c/codec.h to sources/headers.
src/config.c	UPDATE (commit B)	Add #include "codec.h"; wire config_object->pixelformat.
src/surface.c	STRIP (commit C)	CreateSurfaces2 reduced from ~250 LOC to ~50 LOC. Delete surface_reset_format_cache function. Delete OUTPUT-side block + CAPTURE probe + cap_pool_init + per-surface destination_* fill (all moved to context.c).
src/surface.h	UPDATE (commit C)	Remove surface_reset_format_cache declaration.
src/context.c	EXPAND (commit C)	CreateContext now owns the entire OUTPUT-side lifecycle. DestroyContext adds request_pool_destroy (CRIT-2 fix) and drops the surface_reset_format_cache call.
src/request.h	UPDATE (commit C)	Delete last_output_width/last_output_height fields; replace block with explanatory comment.

Net delta: +238/-280 LOC across 5 files + 2 new = effectively unchanged total size, but code is now in the architecturally correct place.

Build + install status

• Fork tip pushed to git.reauktion.de/marfrit/libva-v4l2-request-fourier master at 7055b14.
• Fresnel build/install: pending. Fresnel offline (ssh: No route to host) at push time. When fresnel returns:

ssh fresnel 'cd ~/src/libva-v4l2-request-fourier && \
  git pull --ff-only origin master && \
  ninja -C build && \
  sudo install -m 644 build/src/v4l2_request_drv_video.so \
       /usr/lib/dri/v4l2_request_drv_video.so && \
  sha256sum /usr/lib/dri/v4l2_request_drv_video.so'

Predicted new SHA256: TBD. Will differ from iter4-tip 6e90b7a9b2c33480… and from rejected iter5b-α' d7722da742bfcb86….

Phase 7 readiness

The Phase 4 v2 plan's C8 verification matrix is ready to run once the new backend is installed. Re-run /tmp/iter5b_p7/sweep.sh (preserved from earlier verification — needs device-path remapping per current boot).

Expected post-β hash matrix (same as Phase 4 v2 C8):

Codec	Pre-β libva	Expected post-β libva	kdirect (anchor)
H.264 1080p30	71ac099b… keyframe partial	unchanged (Bug 4 deferred)	1e7a0bc9…
HEVC 720p	06b2c5a0… all-zero	9340b832…	9340b832…
VP9 720p	06b2c5a0…	4f1565e8…	4f1565e8…
MPEG-2 720p	19eefbf4… worked	19eefbf4… (unchanged)	19eefbf4…
VP8 720p	06b2c5a0…	136ce5cb…	136ce5cb…

PASS criteria: 3 codecs unblocked (HEVC + VP9 + VP8 libva == kdirect == sw) + MPEG-2 unchanged + H.264 unchanged.

Phase 7 robustness check (specifically for the α' failure mode): run the sweep TWICE back-to-back in the same shell (or one ffmpeg-vaapi session followed by another with a different codec) to confirm the lifecycle is robust across context cycles. The α' failure mode was multi-CreateSurfaces2 mid-stream; β should handle it because each session is a self-contained CreateContext setup-and-teardown cycle.

Risk register at Phase 6 close

Risk	Status
Build fails on fresnel	Unknown until fresnel returns. Code review + grep verified all symbols, includes, and types resolve.
HEVC SIGSEGV (α' failure mode)	Mitigated — β has no in-CreateSurfaces2 destructive teardown branch.
CRIT-1 (NULL guard)	Fixed in commit C.
CRIT-2 (missing request_pool_destroy)	Fixed in commit C.
mpv probe-then-real	Expected to work — DestroyContext between cycles.
ohm backwards compat	Expected to work — MPEG-2 path gets explicit MPEG2_SLICE (which hantro previously substituted anyway).

Substrate state at Phase 6 close

• Fork tip 7055b14 on noether + gitea.
• Backend on fresnel: still iter4-tip SHA 6e90b7a9b2c33480… (revert state) until fresnel returns + pull + build + install runs.
• Kernel: linux-fresnel-fourier 7.0-1 (unchanged, no kernel work in iter5b).
• Test fixtures: unchanged.
• Phase 7 sweep: ready to execute once new backend is installed.