docs(state): de-duplicate the closed T-SPEED-GPU-REGISTRY-ORPHAN row

[lusoris]

Summary

Ledger-only correction. T-SPEED-GPU-REGISTRY-ORPHAN-2026-06-19 was fixed on master long ago by commit a0bf83c214 (PR #1004), but docs/state.md carried the bug twice in "Recently closed" and kept two identical tombstone comments in "Open bugs". A session reading the ledger cold could take that as unsettled and re-investigate. This PR collapses it to exactly one row per bug id and records the fresh re-verification against origin/master. No source, build, or test behaviour changes.

Verified at origin/master (not taken from the ledger prose):

• git ls-tree origin/master core/src/feature/ | grep -i feature_extractor → only feature_extractor.cpp and feature_extractor.h; the dead .c twin is gone.
• Externs present at lines 105-106 (vmaf_fex_speed_{chroma,temporal}_cuda), 134-135 (_sycl), 216-217 (_hip).
• Registry entries at line 310 (inside #if HAVE_SYCL), 327 (#if HAVE_CUDA), 405 (#if HAVE_HIP) — each verified to sit before the matching #endif. CPU twins registered unconditionally at line 276.
• Later touches 300423b9e and fcae339b5 did not remove them.

So vmaf_get_feature_extractor_by_name("speed_chroma_cuda" | "speed_temporal_cuda" | …_sycl | …_hip")` resolves on a build with the matching backend enabled.

Type

• docs — documentation only

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• pre-commit run --files green on every touched file (markdownlint, dispatch-registry check, secret scan).
• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.

Bug-status hygiene (ADR-0165)

• docs/state.md updated in this PR — the duplicate "Recently closed" row and the duplicate tombstone comment are removed; the surviving row is rewritten in past tense citing commit a0bf83c214 (PR fix(registry): restore orphaned GPU SpEED registrations + delete dead feature_extractor.c #1004) plus the 2026-09-06 re-verification.

Netflix golden-data gate (ADR-0024)

• No golden assertion touched — this PR changes only docs/state.md, CHANGELOG.md (rendered) and one changelog.d/ fragment.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: trivial ledger-only correction; the verification is inlined above.
• Decision matrix — no alternatives: only-one-way fix; a duplicated bug row can only be de-duplicated.
• AGENTS.md invariant note — no rebase-sensitive invariants.
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG fragment — changelog.d/fixed/t-speed-gpu-registry-orphan-ledger-dedup.md; CHANGELOG.md regenerated with bash scripts/release/concat-changelog-fragments.sh --write.
• Rebase note — no rebase impact: ledger-only change, touches no upstream-mirrored source.

Reproducer

# exactly one ledger row for the bug id
grep -c '^| \*\*T-SPEED-GPU-REGISTRY-ORPHAN-2026-06-19\*\*' docs/state.md   # -> 1

# the fix is really on master
git show origin/master:core/src/feature/feature_extractor.cpp | grep -n -i speed
git ls-tree origin/master core/src/feature/ | grep -i feature_extractor      # no .c twin

# rendered changelog matches the fragment tree
bash scripts/release/concat-changelog-fragments.sh --check

Known follow-ups

None. T-CUDA-INIT-SUBMIT-LEAKS-2026-06-19 also has both an open row and a tombstone comment in docs/state.md; that one is a separate bug id and is deliberately left untouched here.

[lusoris]

Superseded by #1322, which merged first.

T-SPEED-GPU-REGISTRY-ORPHAN-2026-06-19 was one of the duplicate-row pairs #1322 removed wholesale. On current master the id appears exactly once, under ## Recently closed (line 391), with no stale Open row left to de-duplicate:

git show origin/master:docs/state.md | grep -c '^| \*\*T-SPEED-GPU-REGISTRY-ORPHAN-2026-06-19\*\*'   # 1

Rebasing this would either be a no-op or re-add a second row, which the gate #1322 introduced (scripts/ci/check-state-md-rows.sh) now rejects. Closing rather than rebasing.