iter6 PARTIAL close: Bug 6 narrowed to H-E (kernel-side hantro VP8 partial-write)

Phase 3 Candidate K executed: H-D (slot rotation) ELIMINATED via
instrumented bind+read site logging. Slot v4l2_index matches at
BeginPicture and at vaGetImage for every surface; destination_data[0]
matches slot->map[0]. No rotation mismatch.

H-A/B/C/D all eliminated. H-E (kernel-side hantro VP8 partial-write)
confirmed by elimination. The libva backend submits correct controls,
correct slice bytes, correct slices_size, correct slot indices.
Kernel writes erratic partial content (per-frame Y plane transitions
at row 536, 24, ... — not a clean buffer-size truncation, not slot
rotation).

iter6 close PARTIAL: 5 of 6 Phase 1 criteria PASS; criterion 1
(libva_vp8 == kdirect) PARTIAL — kernel-side fix needed, out of
iter6's locked backend-only scope.

No patches landed. Fresnel substrate unchanged: fork tip 70196f8,
backend SHA 2c6ff82c... (identical to iter5b-β close).

Net deliverable: Phase 3 narrowing reduces Bug-6 hypothesis space
from 5 to 1. Future iter7+ (or kernel-agent campaign) picks up the
kernel-side investigation.

Pattern recognized: iter2 HEVC transitive PASS masked Bug 5;
iter3 VP8 transitive PASS masked Bug 6. Both surfaced under direct
verification post-iter5b-β. Transitive proofs against ONE artifact
(control payload) don't catch bugs in OTHER artifacts.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

phase3_iter6_findings.md

@@ -39,11 +39,26 @@ VP8 keyframe control payload byte-identical between libva and kdirect on the cur
 
 ## Remaining hypotheses
 
-### H-D — CAPTURE slot rotation mismatch: open
+### H-D — CAPTURE slot rotation mismatch: ELIMINATED (user pick Candidate K, executed 2026-05-12)
 
-Not directly tested in this session. Would need: log `slot->v4l2_index` at cap_pool_acquire time and at copy_surface_to_image read time; verify they match. If they diverge, libva reads from a stale slot while kernel wrote to a different one.
+Instrumented `surface_bind_slot` in surface.c and `copy_surface_to_image` in image.c to log slot indices and `destination_data[]` pointers. Re-ran VP8 sweep.
 
-### H-E — kernel-side hantro VP8 quirk: open and increasingly likely
+Empirical result (excerpt):
+
+```
+H-D bind: surface=0xaaab0111d630 slot=v4l2_index=0 dst_index=0 map[0]=0xffffa465e000
+H-D bind: surface=0xaaab01122110 slot=v4l2_index=1 dst_index=1 map[0]=0xffffa450c000
+H-D bind: surface=0xaaab01126bf0 slot=v4l2_index=2 dst_index=2 map[0]=0xffffa43ba000
+H-D read: surface=0xaaab0111d630 dst_index=0 current_slot=… destination_data[0]=0xffffa465e000 destination_data[1]=0xffffa473f000
+H-D read: surface=0xaaab01122110 dst_index=1 current_slot=… destination_data[0]=0xffffa450c000 destination_data[1]=0xffffa45ed000
+H-D read: surface=0xaaab01126bf0 dst_index=2 current_slot=… destination_data[0]=0xffffa43ba000 destination_data[1]=0xffffa449b000
+```
+
+For each surface: the slot v4l2_index at `surface_bind_slot` (BeginPicture time) **matches** the dst_index at `copy_surface_to_image` (vaGetImage time). The `destination_data[0]` pointer matches `slot->map[0]` returned by `cap_pool_acquire`. **No slot rotation mismatch.** H-D eliminated.
+
+(Bonus observation: cap_pool acquires slots in increasing index order — slot 0, 1, 2, 3, … through 12+ over a 3-frame decode. LRU semantics working as designed.)
+
+### H-E — kernel-side hantro VP8 quirk: CONFIRMED by elimination of H-A/B/C/D
 
 The output bytes show an **erratic partial-write pattern**:
 
@@ -108,12 +123,9 @@ iter6 hands off Bugs 4/5/6 to iter7+. Memory updates for the iter6 lesson on tra
 
 ## Decision point
 
-This is a user-decision point. Phase 3 has done its narrowing job. Bug 6's actual fix is either:
+User picked Candidate K. Phase 3 executed. **H-D eliminated**, **H-E confirmed**. iter6's Phase 1 locked scope was backend-only ("backend-side fix expected"). Bug 6 is kernel-side. **iter6 closes PARTIAL** — Phase 3 narrowing delivered as the iter6 contribution; Bug 6 fix deferred to iter7+ (would target kernel-side work, similar to original iter5 Candidate B scope).
 
-- 1-2 hours more empirical work (Candidate K — likely productive)
+Remaining user pick: which target for iter7.
-- Multi-session kernel-side work (Candidate L)
-- Pivot to a different bug (Candidate M)
-- Close iter6 partial and document (Candidate N)
 
 ## Substrate state at iter6 Phase 3 close
 

phase8_iteration6_close.md

@@ -0,0 +1,105 @@
+# Iteration 6 — Phase 8 (close)
+
+Closes 2026-05-12. iter6 = Bug 6 (VP8 partial output post-iter5b-β) targeted; root-cause narrowed but **not fixed**. iter6 closes PARTIAL with the narrowing as the iter6 deliverable. Bug 6's actual fix lives in kernel territory, outside iter6's Phase 1 locked scope.
+
+## Summary
+
+| Metric | Value |
+|---|---|
+| Iteration target | Bug 6 — VP8 libva output diverges from kdirect (74.8% zero with real keyframe content) |
+| Hardware | RK3399 hantro-vpu-dec for VP8 |
+| Fork tip start (iter5b-β close) | `70196f8` |
+| Fork tip end (iter6 close) | `70196f8` (unchanged — no patches landed) |
+| Phase 1 criteria | 5/6 PASS, 1/6 PARTIAL |
+| Bug 6 status | **narrowed to kernel-side (H-E)**; backend-side fix not available |
+| Campaign scoreboard | unchanged from iter5b-β close: 5/5 with 2 direct (VP9, MPEG-2) + 3 mixed (H.264 keyframe-partial, VP8 partial-with-Bug-6, HEVC transitive-only with Bug 5) |
+
+## Deliverable: Bug 6 root-cause narrowing
+
+Phase 3 systematically eliminated 4 of 5 hypotheses via empirical investigation:
+
+| Hypothesis | Status | Evidence |
+|---|---|---|
+| H-A — slice data corruption in libva path | **ELIMINATED** | libva-dumped slice 0 (300614 bytes) byte-identical to raw VP8 frame 0 from .webm post-header (SHA `9e74956c…`). |
+| H-B — `slices_size` wrong on OUTPUT QBUF | **ELIMINATED** | slices_size = fp_size + sum(dct_part_sizes) = 22742 + 277872 = 300614 exactly. |
+| H-C — CAPTURE-side cache coherency | **ELIMINATED** | VP9 uses same `image.c::copy_surface_to_image` path and works fine. msync attempt empirically no-op (output hash unchanged). |
+| H-D — CAPTURE slot rotation mismatch | **ELIMINATED** (Phase 3 Candidate K) | Instrumented bind + read sites; slot v4l2_index matches between BeginPicture and vaGetImage for every surface tested. destination_data[0] pointer matches slot->map[0]. |
+| H-E — kernel-side hantro VP8 partial-write quirk | **CONFIRMED by elimination** | The control bytes are right, the slice bytes are right, the buffer pointers are right, the slot indices are right. Yet the kernel writes only a partial subset of the frame to the CAPTURE buffer (per-frame transition rows: frame 0 Y row 536, frame 0 UV row 134, frame 1 Y row 24 — erratic). |
+
+The libva backend is doing the right thing on the V4L2 protocol surface. The kernel-side hantro VP8 driver writes partial content to the CAPTURE buffer.
+
+This narrowing is iter6's deliverable. Future investigation (iter7 or kernel-agent work) needs ftrace, hantro source-read, possibly a kernel patch.
+
+## Per-criterion verdict
+
+| # | Criterion | Verdict | Evidence |
+|---|---|---|---|
+| 1 | VP8 libva == kdirect | **PARTIAL** — kernel-side bug confirmed via systematic elimination, not fixed | post-β VP8 hash `bcc57ed5…`, kdirect `136ce5cb…` |
+| 2 | VP9 unchanged | **PASS** | `4f1565e8…` == kdirect (iter5b-β state preserved) |
+| 3 | MPEG-2 unchanged | **PASS** | `19eefbf4…` == kdirect (iter5b-β state preserved) |
+| 4 | H.264 keyframe-partial unchanged | **PASS** | `71ac099b…` (Bug 4 still deferred) |
+| 5 | HEVC unchanged | **PASS** | `06b2c5a0…` all-zero (Bug 5 still deferred) |
+| 6 | Control-payload anchors hold | **PASS** | No control-submission code changes in iter6 |
+
+5 of 6 PASS. Criterion 1 PARTIAL: the locked goal of `libva_vp8 == kdirect_vp8` not reached, but Phase 3 narrowed it from "unknown" to "kernel-side H-E confirmed."
+
+## Cost of iter6
+
+- Phase 0+1+2+3 docs.
+- 0 fork commits (no patches landed).
+- ~3 hours of diagnostic instrumentation (slice-data dump, msync test, slot-index logging) all reverted at end of Phase 3.
+- Substrate unchanged: fork tip `70196f8`, backend SHA `2c6ff82c…`. Net delta on fresnel: nil.
+- Net deliverable: Phase 3 narrowing doc with crisp H-A/B/C/D elimination + H-E confirmation. The hypothesis space for future Bug-6 work shrinks from 5 → 1.
+
+## Lessons distilled
+
+### Transitive-proof blind spots (re-confirmed)
+
+iter3 closed VP8 via "transitive PASS" verifying specific control fields (`first_part_size`, `first_part_header_bits`) matched kdirect. iter5b-β unblocked VP8 from the OUTPUT-format mismatch. iter6 Phase 3 confirmed: control payload is byte-identical to kdirect, slice data is byte-identical to the raw .webm, slices_size is correct — and the kernel STILL produces wrong output.
+
+iter3's transitive proof, iter5b-β's β refactor, and iter6's Phase 3 narrowing are all consistent with each other. The kernel-side partial-write (H-E) was masked by the OUTPUT format mismatch pre-β (all-zero output regardless of what else was right). Once β fixed the format, H-E became visible.
+
+Same shape as iter2-HEVC + Bug 5: transitive proofs against ONE artifact (control payload) don't catch bugs in OTHER artifacts (the kernel's actual decode behavior).
+
+### Phase 3 instrumentation hygiene
+
+iter6 Phase 3 added/reverted instrumentation three times (slice dump, msync, slot indices). Each rebuild + run + analyze cycle was ~5 minutes. Total instrumentation overhead: maybe 30 minutes across 3 cycles. The discipline of "add diagnostic, run, revert" worked — fresnel ends at clean SHA `2c6ff82c…` identical to iter5b-β close.
+
+Worth a memory note: **instrument-revert workflow is fast and safe for empirical narrowing**. Future Phase 3 work should embrace the same pattern instead of trying to deduce from source alone.
+
+### iter6 didn't need Phase 4-7 because the Phase 3 verdict was "kernel-side, out of scope"
+
+The 8(+1)-phase loop assumes Phase 3 narrows the fix site and Phase 4 plans the fix. When Phase 3 narrows OUT OF SCOPE (e.g., kernel-side when the iteration locked backend-only), Phase 4 onwards is moot. iter6 closes from Phase 3 directly to Phase 8 with the narrowing as the deliverable.
+
+This is a legitimate close pattern, not a process violation. The dev process allows for "Phase 3 → Phase 0 loopback" or "Phase 3 → iteration close PARTIAL" when the empirical evidence demands it. iter6 demonstrates the latter.
+
+## Phase 4 cross-cutting backlog status (iter6 increment)
+
+Unchanged from iter5b-β close:
+- iter4-B1 device-discrimination: still open.
+- iter4-B2 mpv-vaapi create-device: still open.
+- iter4-Q6, COLOR_RANGE, B3-B6, L3: all still open.
+
+iter6 NEW backlog items:
+- **Bug 6 narrowed to H-E**: kernel-side hantro VP8 partial-write. Per-frame Y plane transitions at varying rows (536, 24, ...). Not slot rotation, not cache, not slice corruption, not bytesused. Investigation surface: hantro driver source (hantro_g1_vp8_dec.c likely), ftrace at kernel for VP8 decode path, possibly a hantro-VP8 kernel patch.
+
+## iter6 → iter7 handoff
+
+Substrate at close:
+- Fork tip `70196f8` (iter5b-β + Commit D). Identical to iter6 open.
+- Backend SHA `2c6ff82c…` on fresnel. Identical to iter6 open.
+- Kernel unchanged.
+- Test fixtures unchanged.
+
+Candidates for iter7 Phase 0:
+- **Bug 6 H-E kernel-side fix** — heavy, kernel work. Aligns with original iter5 Candidate B scope (vb2_dma_resv era kernel patches, plus new VP8 partial-write fix).
+- **Bug 4 H.264 inter race-loss** — backend or kernel; iter4 partial work history.
+- **Bug 5 HEVC DQBUF FLAG_ERROR** — iter2's transitive PASS masked it; need to find the V4L2 protocol gap between libva-HEVC and kdirect-HEVC.
+- **iter4-B1 auto-detect harden** — backend, ~100 LOC, quality-of-life.
+- **Re-anchor regression hashes on β substrate** — establish iter7+ regression invariants.
+
+User picks at iter7 Phase 0 lock.
+
+## Memory rule note (deferred)
+
+iter6's lesson about Phase-3-empirical-revert workflow is worth a memory entry; defer the write to iter7 close or to a dedicated memory-curation session.