Historical lab document. Original machine, code and acceptance scope apply. Private paths are redacted; this is not a fresh public 0.6.1 benchmark. Read the latest outcome and public configuration.
Known gaps ও সমাধান — Phase 4–11
৫ নভেম্বর ২০২৬। বর্তমান bounded performance evaluation-এর Phase 11 ফল, register ও method দেখুন; Phase 10 ফল ও Phase 9 ফল retained। Active release-এর original gates Phase 8 register ও আগের Phase 4–7 register-এ। Phase 8 ফল, Phase 4 calibration, workflow/recovery/release contract ও operational guide দেখুন। Historical reports ও তাদের নিজস্ব scope retained; Phase11 overall promotion gate pass হয়নি।
| ID | সমস্যা | ব্যবস্থা | বর্তমান প্রমাণ / সীমা |
|---|---|---|---|
| G01 | Hardware-কে সর্বদা available/fast ধরা | Actual no-fallback encode ও forced hardware decode surfaces | 4K Main10 ও unsupported-decoder refusal pass; universal fastest নয় |
| G02 | অপ্রয়োজনীয় re-encode | All-stream Matroska copy | Real packet hashes ও 10-bit/colour metadata pass |
| G03 | Thread oversubscription | Independent decoder/encoder/filter/validation pools, sequential lock | Real long HD/UHD/live-action calibration; M4 Max eligible preset ও explicit override |
| G04 | ভুল headroom/admission | Compressed-memory-aware policy, explicit scoped override | Real default refusal; owner waiver current runs-এ recorded |
| G05 | চলমান job-এর resource growth | Sampled footprint/disk/total deadline, durable full JSONL trace | 30-minute real-media elapsed soak, 355 records, 54,000 decoded frames; overnight/kernel quota পৃথক |
| G06 | Worker হারালে child চলে যাওয়া | Independent guardian, pre-exec ready handshake, kernel birth/nonce ও inherited lock | Worker/supervisor/both-loss actual matrix-এ owned media stopped; OS/guardian simultaneous failure guarantee নয় |
| G07 | Source/output overwrite | Read-only source identity, staging, atomic no-clobber link | Race tests এবং real audio/video/image/remux publish |
| G08 | Shell injection/raw argv | Strict typed schema/argv ও canonical path control-character refusal | Actual symlink-alias input/output/state rejection; fresh MCP refusal ও 126-test matrix |
| G09 | Indirect/network media fetch | Local input, protocol/demuxer allowlists | Real network playlist refusal ও MCP input/root checks; broad fuzzing future |
| G10 | Duplicate/replayed encode | SQLite idempotency ও active output reservation | Concurrent test, real CLI/MCP replay |
| G11 | Unbounded concurrent batch | Host-wide sequential lock, bounded drain | Lock tests; repeated real batch workloads |
| G12 | Ambiguous agent interface | Versioned JSON/errors ও 8 typed MCP tools | Real stdio queue→worker→output pass |
| G13 | Cancel/timeout-এ ফল হারানো | Owned group stop, immutable attempt/partial retention | Controlled process tests ও real queued MCP cancel |
| G14 | Progress unit/format ambiguity | Optional native JSONL; microseconds; encode receipt retains last progress | Core JSONL/kv/invalid option ও actual JSONL encoding pass |
| G15 | Log/stdout deadlock | Selector drain, bounded tail/log/request | Large-stderr ও split-record tests |
| G16 | Exit 0-কে valid output ধরা | Strict input errors, metadata/full decode/hash/atomic publish | Truncated refusal, nonzero audio, 4K, VFR, dense/sparse pixels ও full-decode proof |
| G17 | Stale source/binary/code plan | Source/runtime ctime/OS/load paths; MCP startup workflow digest | Restored-mtime/race rejection ও stale source-process refusal; checkpoint source content SHA; non-checkpoint full source snapshot future |
| G18 | HDR/bit-depth/colour loss | Explicit static PQ/HLG video ও 16-bit sRGB PNG; unknown signalling refuse | Independent analytic pixels, tags/alpha/16-bit matte pass; untagged publisher HDR refusal; mastering-display judgement পৃথক |
| G19 | Remux-কে exact trim বলা | Remux trim নিষিদ্ধ; decoded range export, explicit VFR/CFR | 60-frame VFR PTS ও 90° rotation pass; wider camera timing corpus future |
| G20 | Frame timestamp অনুমান | Requested time বনাম decoded source PTS | Single frame ও 16-frame sheet timestamps independently match |
| G21 | ভুয়া bitstream resume | Validated CFR video segments, source/code/tool SHA ও continuous AAC pass | Actual interruption/corruption/retry, 3,600 correct frames ও identical final audio packets; VFR checkpoint refuse |
| G22 | Unmeasured speed claim | Frozen baseline, excluded warmup, paired real-media trials, all-frame quality ও size gates | Four long software profiles calibrated; final comparison আলাদা epoch; universal Stock advantage দাবি নেই |
| G23 | Unrelated app config loss | Private backup, new local MCP entry, original tables preserved | Configured entry verified; current chat reload আলাদা |
| G24 | Fork/update/build দুর্বলতা | Pinned source, C patch, checksum replay, isolated build/cache/resume | Compiled M4 fork, local FATE, native activation; security refresh manual |
| G25 | Universal format support ধরে নেওয়া | Typed contract, full native CLI, broader actual import/export matrix | Real HD/UHD/live-action corpus; 15 format tests ও 15 selected external FATE targets; সব camera/codec নয় |
| G26 | বড় job ছোট job আটকানো | Priority/aging; checkpoint job durable segment boundary-তে yield | Actual urgent job completed, lower job resumed ও validated output; arbitrary non-checkpoint mid-bitstream preemption নয় |
| G27 | Platform end-to-end অজানা | Fresh real stdio MCP এবং existing Studio timeline worker-এর pinned CLI | Actual cut/retime/fade/source-audio export, retained-part identical recovery ও tampered manifest refusal; creative approval/GUI deployment নয় |
| G28 | সব hard-kill case guarantee | Guardian-এর pre-exec registration/ready handshake ও inherited host lock | Actual worker/supervisor/both SIGKILL; guardian/OS loss বা instantaneous kernel quota-এর universal guarantee নেই |
| G29 | RSS বনাম actual footprint ও sampling overhead | Native macOS VM/libproc API, bounded child tree ও CLI fallback | No-helper/live-child/ABI tests; sampled footprint policy বহাল, instantaneous GPU/kernel quota নয় |
| G30 | Shared library provenance | Cold SHA-256, ctime/OS/device/inode/load paths; pinned release engine snapshot | Active 38 files; device-ID change-এ guard refusal ও fresh native/extended reactivation; reproducible binary future |
| G31 | Repeated frame startup ও sparse full decode | Dense single-pass ও count≤8 bounded labelled seeks | 3-trial 600s single-pass→seek 5.983→0.361s, same PTS/pixels; GOP-aware calibration future |
| G32 | Quality metric frame misalignment | Decoded-frame-number alignment at 30fps | Exactly 180 samples; superseded 181-sample pass retained |
| G33 | Explicit quality request silently ignored | Software quality mode অথবা typed hardware refusal | Controlled request-contract regression pass |
| G34 | Valid cache/result পুনর্নির্মাণ | Source/tool/config/binary matching cache; diagnostic stage resume | Actual already_built & patch already_applied; successful media reused |
| G35 | Short-child group-signal PermissionError race | Owned Popen direct-signal fallback ও pipe cleanup | Deterministic regression ও Phase 1-এর 20 real repetitions pass |
| G36 | Repeated capability/help child processes | Selected/custom dependency cache; digest/OS/ctime/alias invalidation | Native warm lookup ≈1.7–1.9ms; corrupt/blocked/changed dependency tests; unresolvable closure uncached |
| G37 | প্রতি phase-এ Python supervisor startup | One reusable independent supervisor per worker pass | Matched workload gains; real worker-loss/cancel/timeout pass; full validation বহাল |
| G38 | Offline Metal compiler অনুপস্থিত | Optional runtime shader source compilation | NV12/P010 GPU smoke ও matched 60-frame hashes; first-use compile latency থাকে |
| G39 | Metal command over-release cleanup crash | Scoped per-dispatch autorelease pool | Stock exit −11 reproduced; native exit 0, identical frame hashes ও ~90s GPU soak |
| G40 | ffplay/zscale/subtitle feature gaps | Existing SDL2, pinned private static zimg ও libass | Headless playback, analytic tone map, Bengali/ASS visible timing; all-font/GUI/perceptual review পৃথক |
| G41 | Metadata measurement failure অদৃশ্য হওয়া | Live unmeasurable child fail-closed; already-exited short child confirm/reap | Deterministic short-child race regression ও fresh native matrix; no hidden live-child measurement bypass |
| G42 | CPU download/scale ও closed-pipe wait | Explicit hardware-frame scale; guarded closed-pipe wait | 60s real 4K→1080 five-pair calibration-এ VMAF −6.45; hardware scaler default-এ promote হয়নি |
| G43 | HD encoder/validation concurrency সর্বোচ্চ ৪-এ আটকে থাকা | Scoped M4 Max HD 8/2/2/8 profile; software-only explicit encoder/validation≤8; nominal encode budget≤14 | Nine-pair HD/ToS confirmation-এ পুরোনো Native থেকে সময় 22.31%/30.77% কম; quality/size pass; explicit/4K/resize/checkpoint behaviour retained |
| G44 | Tuned native/Stock gap বা raw-feed gap অদৃশ্য করা | Same-tuning Stock control, all retained slow trials, adjacent decoded-pixel/stage isolation | HD matched median +9.53%, ToS −0.84%; intervals equality পার করে। Adjacent compressed AV +1.19% uncertain; raw feed +4.44% ধীর, সীমিত unresolved control; universal core advantage নয় |
| G45 | Host-state variation ও calibration winner-কে স্থায়ী fix ধরা | Balanced whole-block resampling, kernel CPU/instruction/cycle counters, read-only all-frame encoder observer ও independent complete workflow | Queue 2/4/8/16 গৃহীত নয়; HD lookahead calibration −8.85% থেকে independent +39.94%। নতুন full-frame quality pass, speed fail; 0.6.0 retained। Independent raw +6.08%, interval 0.9986–1.2776; specific stable core defect ও দুই Stock gap-এর fix অমীমাংসিত |
| G46 | Macroblock progress query-তে অপ্রয়োজনীয় encoder locking | Private Apple-arm64 acquire/release progress API ও positive row-wait lower-bound cache; required-line safety/fallback retained | Captured x264 mutex traffic প্রায় 98.4% কম, bounded overflow accounted; target instructions 2.9–3.3% কম। Exact decoded HD/raw/audio; 8/10-bit TSAN ও codec regression pass। Experimental; elapsed speed promotion gate অসম্পূর্ণ |
| G47 | CPU queue/cache variation-কে encoder defect বা speedup বলা | Owned load replay, calibrated Mach counters, per-engine-bound worker observer ও pthread-hook ছাড়া process observer; actual disk reads retained | 180 encode + 45 workflow runs; raw physical reads 2.80GB বনাম cached MB-level reads observed। সব confidence gates/slow trials retained; Stock gap closure ও default update অপ্রমাণিত |
| G48 | Parent wakeup delay ও cached host statistics-কে short-run CPU/elapsed ধরা | Kernel child start/exit clock, final zombie rusage_info_v6 before reap; calibrated per-core host_processor_info |
Direct/descendant/short-child counters ও14-core capacity pass; stale/plateau counter ও accounting mismatch fail closed |
| G49 | Raw input বা executable cache residency অজানা | Same source/image mappings bounded temporary lock, pre/post residency ও direct physical-read budget | Task-local control; no global cache purge/lock-limit change; warmup cost retained outside child timer। Warm-cache result, production cold-disk throughput gain নয় |
| G50 | Generic compiler code layout-এ remaining CPU work | Same pinned row-progress source-এ disjoint-training M4/vector PGO, no LTO/instrumentation in final | HD/raw/audio exact;14 native/6 codec checks ও176 tests/runtime pass। Direct688 ও full-workflow144 cases retained। Complete HD workflow Stock থেকে1.743%/1.791% কম elapsed,9 valid blocks/load; direct HD1/9 ও raw3/9,7/9 হওয়ায় overall gate fail; candidate experimental |
| G51 | Per-case input unpin ও file-backed wire footprint অদৃশ্য | Parent-held common cache lease per paired block; process-tree footprint+held mapped bytes accounting | Read-only full residency ও partial-failure cleanup tests;12GiB own watchdog retained। Whole-run hold boot400GB guard-এ stopped; lower-demand block hold ও automatic disk reserve admission active |
| G52 | Steady saturation-কে available CPU এবং ভুল mincore bit ধরা | Global preflight busy≤6; correct SDK paged-out mask0x20 বনাম reference-other0x08;30s passive window watcher | Watcher নিজে quiet window পেয়ে independent run launch করেছে; old warmups imported নয়। Complete workflow9/9 দুই load-এ pass; strict direct within-run/block/P-share stability-তে required sample অসম্পূর্ণ, broader proof remaining |
| G53 | Volume device-ID drift-এ current/rollback binding stale | Full SHA/alias/inode/size/mtime audit; fresh original-native verification; exact frozen original package metadata reactivation | Original0.6.0-e07af57a8214502f active,0.5.0-018d61f631128c48 healthy fallback; actual rollback এবং installed CLI/MCP pass। Old snapshots retained; guard bypass/PGO promotion নয় |
Local release-এর শেষ গ্রহণযোগ্যতা VALIDATION.md ও Phase 8 register-এ। পরের bounded scope হতে পারে raw I/O/feed ও busy-host tail latency, wider/private camera corpus, overnight, reference-display HDR judgement, calibrated অন্যান্য hardware/codec, full external FATE, reproducible distribution ও upstream security refresh। এগুলো পরীক্ষিত presets-এর বাইরে। Install/update/rollback-এর পৃথক actual gate আছে।