diff --git a/SKILL.md b/SKILL.md index 01cc380..b0abe6e 100644 --- a/SKILL.md +++ b/SKILL.md @@ -278,6 +278,7 @@ re-encoded to `-b:a 256k`: that added ~128 kbps for no audible gain, ~+10% file - **Kill by PID only**, one lock per tier, one NVENC job at a time, and pause GPU tiers before a driver update (an update wedges GPU work silently — the output-growth watchdog is the net). - **A multi-day run must not be tied to the agent session.** Two tiers launched as Hermes background jobs with `persist_on_release` were both killed mid-encode the moment the user closed their session (measured 2026-10-08 10:37: `*_out.mp4` and `.err` frozen in the stage dir, no `PASS-DONE` in either log, ~40 min of work lost). Launch them **orphaned** instead — `powershell -NoProfile -Command "Start-Process -FilePath -ArgumentList -WindowStyle Hidden"`, the script itself looping `library_run.py` with a retry — so the parent exits immediately and the agent's process cleanup cannot reach the tree. Verify the chain: the job's parent must be a bash whose own parent is already gone. Cost: no exit notification (watch the log) and **stopping is by PID only**. - **The `copy` tier is not idempotent — never re-run a whole `--tier copy` pass.** Its only skip test is "already HEVC hvc1", and a copy output is H.264 by definition, so a second pass redoes every file it ever produced (193 of them here, ~40 GB of share I/O). For a handful of newly-arrived files use the single-file path per file: `library_run.py --tier copy --path --keep-source`. +- **A re-do of an already-playable file earns a 0.975 mean bar; a first-time conversion does not.** Measured over one pass of this library: 11 attempts were rejected purely on the window mean (0.9655–0.9795) while accepted encodes had *worse* frame minima (0.33–0.90) — the bar was splitting a continuum, not separating defects (`0718 马斯克原理` failed with min 0.887 and **0.00 %** of frames below 0.85). A re-do exists only to reclaim an old setting's bloat from a file that already plays (HEVC hvc1 + AAC), and reaching a 0.98 mean costs more bits than the source's own bitrate — which the size ceiling forbids — so a strict bar strands the file with its bloat for ever while burning an encode every pass. `SSIM_MEAN_MIN_REDO = 0.975` therefore applies **only** when the live file is genuinely playable (HEVC hvc1 **and** AAC) and this run is replacing it from the snapshot original. The frame-share test (>1 % of frames below 0.85) is unchanged in both cases and remains the test that catches a real defect — under the relaxed bar `失控` (0.9655) is still rejected, and `0808 陌生人效应` (0.9739) keeps the full 0.98 bar because its live file is H.264 — not a re-do at all. - **Full measured detail, every failure mode and its detector: `references/pitfalls.md`.** ## The owner's standing requirements — read this first diff --git a/STATE.md b/STATE.md index 126a8f9..64bb09b 100644 --- a/STATE.md +++ b/STATE.md @@ -54,6 +54,20 @@ bash /d/tmp/xcode/convert_new4.sh # 逐个 --tier copy --path --keep- - `1004 反内卷` 10:48:53 完成 ✓（535 MB → 535 MB，视频 100% 逐位不变，音频转 AAC，149 s，rc=0） +## 图片闸门：re-do 用 0.975（业主决定 2026-10-08） + +第一轮实测：一趟里 11 次编码**只差在 mean**（0.9655–0.9795）就被拒，而被接受的那些帧最低分反而更差 +（0.33–0.90）—— 阈值在切一条连续的分布，不是在分开缺陷（`0718 马斯克原理`：min 0.887、**0.00%** 帧 +<0.85，照样被拒）。被拒的文件本来就已能播（HEVC hvc1 + AAC），re-do 的唯一目的是清掉旧设定的膨胀； +而要够到 mean 0.98 得花掉超过源片自身的码率，会被 size ceiling 拦住 —— 结果是文件连同膨胀被永久搁置， +而且每一趟还白烧一次编码。 + +新规则（`library_run.SSIM_MEAN_MIN_REDO = 0.975`）：**只有**"重做一个已经在播的文件"才用 0.975，条件为 +live = HEVC hvc1 **且**音轨是 AAC（真能播）**且**本次是从快照原档重做；首次转换仍用 0.98。 +帧占比判据（>1% 帧 <0.85）两档都不变 —— 它才是抓真缺陷的那条：`失控`（0.9655）仍会被拒， +`0808 陌生人效应`（0.9739）因为 live 是 H.264、根本不是 re-do，也仍走 0.98。 +先在一个文件上验证（`0321 随他们去`，上一轮 mean 0.9789 被拒）→ 见 `/d/tmp/xcode/run_pilot.log`。 + ## 目录与日志 | 用途 | 路径 | diff --git a/scripts/library_run.py b/scripts/library_run.py index 42d5a30..75afcce 100644 --- a/scripts/library_run.py +++ b/scripts/library_run.py @@ -88,6 +88,15 @@ SIZE_CEIL = 1.0 # output video bitrate must not exceed the source's QUALITY_STEPS = 2 # on a quality-gate failure, retry with the quality raised this many # steps (cq -2 / crf -2) before giving up; the retry only fires on a # file the cheaper setting could not carry +SSIM_MEAN_MIN = 0.98 # the window-mean bar for a normal encode +SSIM_MEAN_MIN_REDO = 0.975 # ... and for a RE-DO of a file that already plays everywhere + # (HEVC hvc1 + AAC). Owner decision 2026-10-08: measured on this library, + # 11 of a pass's attempts died on a mean of 0.9655-0.9795 while files with + # WORSE frame minima (0.33-0.90) passed - the bar was splitting a + # continuum, not separating defects (0718 马斯克原理 failed with 0.00% of + # frames below 0.85 and min 0.887). A re-do's only job is to reclaim the + # old settings' bloat, so it must not be stranded with the bigger file for + # ever; 0.975 still rejects the two genuine stragglers (0.9655, 0.9739). def log(*a): @@ -599,6 +608,7 @@ def process(src, args, dest=None, origin=None): if info["dur"] < 1.0: return "FAIL stub (%.1fs) - needs a re-download, not a transcode" % info["dur"], None live = dest or dst_path_for(src) + redo_playable = False # True once we know this run is re-doing a file that ALREADY plays if info["vcodec"] == "hevc" and info["tag"] == "hvc1" and info["pix"] in ("yuv420p",) \ and not T.unusable_timeline(info, src)[0]: @@ -614,6 +624,14 @@ def process(src, args, dest=None, origin=None): info["size"] / 1e6, o_size / 1e6), None log(" re-doing from the original: live %.0f MB > original %.0f MB" % ( info["size"] / 1e6, o_size / 1e6)) + # A re-do replaces a file that already plays on iPad + Android; its only purpose is to + # reclaim the old settings' bloat, so it earns the relaxed mean bar (owner decision, + # SSIM_MEAN_MIN_REDO) - but only when the live file really is playable end to end + # (HEVC hvc1 video, which the branch above guarantees, plus AAC audio). + redo_playable = info["acodec"] in ("aac", "mp4a") + if not redo_playable: + log(" re-do of a file whose audio is %s (not playable) - the normal bar applies" + % (info["acodec"] or "none")) src, info = origin, probe(origin) if not info: return "FAIL original unreadable", None @@ -687,17 +705,24 @@ def process(src, args, dest=None, origin=None): # Measured on this library: good encodes 0 of 1500 frames below 0.95; the file that # triggered this rule 5 of 1500 below 0.85 (0.33%) at 0.98 mean; the genuine # decoder-defect files 44 of 1499 (2.9%) with the mean falling too. + # The MEAN bar is 0.975 instead of 0.98 when this is a re-do of a file that already + # plays (SSIM_MEAN_MIN_REDO) - the frame-share test below is unchanged and is what + # actually catches a defect. bad_share = (sstat.get("below", 0) / float(nfr)) if nfr else 0.0 - systematic = (mean is not None and mean < 0.98) or bad_share > MAX_BAD_SHARE + mean_min = SSIM_MEAN_MIN_REDO if redo_playable else SSIM_MEAN_MIN + systematic = (mean is not None and mean < mean_min) or bad_share > MAX_BAD_SHARE if systematic: if q_off < QUALITY_STEPS and (attempt == 1): q_off = QUALITY_STEPS - log(" quality gate: mean %.4f, %d of %d frames below %.2f (%.2f%%) - " - "retrying at quality+%d" % (mean, sstat.get("below", 0), nfr, - BAD_FRAME_SSIM, 100.0 * bad_share, q_off)) + log(" quality gate: mean %.4f < bar %.3f, %d of %d frames below %.2f " + "(%.2f%%) - retrying at quality+%d" % ( + mean, mean_min, sstat.get("below", 0), nfr, + BAD_FRAME_SSIM, 100.0 * bad_share, q_off)) continue - return ("FAIL quality: %s - systematic: mean %.4f, %.2f%% of frames below %.2f" - % (ssim_note, mean, 100.0 * bad_share, BAD_FRAME_SSIM)), None + return ("FAIL quality: %s - systematic: mean %.4f < bar %.3f, %.2f%% of frames " + "below %.2f%s" + % (ssim_note, mean, mean_min, 100.0 * bad_share, BAD_FRAME_SSIM, + " (re-do of an already playable file)" if redo_playable else "")), None if lo is not None and lo < SSIM_TARGET: log(" %d scattered frame(s) below %.2f (%.2f%% of the window) with a healthy " "mean - treated as fades/black, not a defect"