Review rounds get caps; reading time follows the film's Pace instead of fixed seconds

[ZLHad]

Why

A studio film (two-constellations) ran 11 full independent review rounds: every fresh reviewer found new taste nits, the scores hovered at 7–9, and small fixes sometimes broke something else. The maintainer, verbatim:

我觉得可以了 其实我觉得也不必要你每次这么十几轮审阅 太多了 给项目加一个限制，每一种问题不超过5次，总共不超过10次，不要为了一个小问题就全部来一遍Review。另外，项目中关于字的节奏也不要写死间隔时间，有的时候太慢。

1. Caps on review rounds

Where the loop is defined: templates/TASTE_CHECKLIST.md (scoring layer), CLAUDE.md / AGENTS.md (effort table, a new 评审有上限 paragraph, hard rule 4), playbook/01 (stage 7, unattended mode), playbook/02 (layer 5), both READMEs, templates/NOTES.md, templates/REVIEW.md, video-types/09.

• Total cap: at most 10 full review rounds per film (a fresh-context reviewer scoring all 8 dimensions over the whole film). At the cap: ship, or take the scores and remaining items to gate ③.
• Per-issue cap: the same kind of issue (one checklist item on one shot or segment) gets at most 5 fix-and-check cycles, then it goes into NOTES as a known limitation with its last evidence, for the human. A floor item at the cap goes to the human too, never shipped silently.
• Small fixes don't trigger a full round: a local fix is checked on its own span (frames and strips, the automated checks covering it, the author's look). A new round is for structural changes or the final pass.
• Minimums unchanged: standard 1 round; studio ≥ 3, all 8 ≥ 8. The draft-count cell stays "不设上限，直到过线" and points to the review cap.
• Log: NOTES gets 评审记录 (one line per round), a 问题计数 table (issue | rounds so far | status) and 已知局限.
• Regressions: when a later round contradicts an earlier one on a taste item, the author may keep the earlier decision with a DECISIONS line, which goes to the next reviewer with the taste overrides. The reviewer also gets the known limitations and the BRIEF's Pace (so Document auto-merge; match the ruleset file to the live ruleset #5 only FAILs under the floor).
• bin/vh effort standard|studio prints the caps paragraph.

2. Reading time: no fixed seconds

	formula	16 CJK chars	4 CJK chars
floor (【底线】, FAIL)	max(1.5 s, CJK/7 + other/20 + 0.8 s), #53's short-label rule	3.1 s	1.5 s
relaxed target	CJK/4.5 + other/15 + 1.0 s	4.6 s	1.9 s
normal target (default)	CJK/6 + other/18 + 0.8 s	3.5 s	1.5 s
brisk target	CJK/7 + other/20 + 0.8 s (= the floor)	3.1 s	1.5 s

(old rule: max(2.5 s, CJK/4.5 + other/15 + 1.5 s): 5.1 s and 2.5 s)

• New BRIEF Spec line - Pace: relaxed | normal | brisk, default normal.
• Tuning: brisk equals the floor. The 8 CJK/s + 0.3 s first proposed for brisk is under the floor at every length, so it could never warn. Relaxed keeps lemo's reading speeds with a 1.0 s pad instead of 1.5 s. Normal sits between them, with the Latin rate scaled to match. 20 other chars/s is about 240 wpm, close to Brysbaert's (2019) meta-analytic silent reading mean of 238 wpm for non-fiction. None of these numbers has been tested on viewers; playbook/03 §2 says so.
• Text the viewer already knows and short labels may run at brisk; text on a musical grid ends on a bar line instead of being padded; a WARN is "take a look", and a sheet of WARNs means the Pace or the word count is wrong.
• tools/readcheck.py: --pace, else the nearest BRIEF.md's Pace line, else normal. FAIL (exit 1) only under the floor, WARN (exit 0) under the target; output OK / WARN / FAIL. data-pace on a clip/scene (nearest wins, sub-compositions included) and "pace" in texts.json; --export writes it. --mode label / data-read="label" kept and folded in (target = brisk). --cjk-cps/--latin-cps/--pad/--min kept; they retune the target only. --budget shows the pace's target and the floor; bin/vh readcheck -h prints the rules. tools/rhythm.py follows (red under the floor, amber under the target, the project's Pace).
• Places that restated 2.5 s: TASTE_CHECKLIST Document auto-merge; match the ruleset file to the live ruleset #5, playbook/03 §2, CLAUDE.md floor list, BRIEF Text rules, STORYBOARD checklist, video-types/01's prompt block, five style presets (four others keep their 2.5 s holds as style beats), the recipes (black-card / breath-title-card / brand-imprint-open recomputed at normal: holds 46 / 62 / 45 frames, with their durations, QA frames and index rows; sequence budgets say 读时 instead of ≥ 75 frames), cases/promo-video-shotcraft.md, a current-state line in docs/research/04 (zh, en).
• Subtitles are unchanged (Netflix ceilings + 1.8 s floor): the voice sets their pace.

Verification

• tools/ci.sh --committed and VH_BASH=/bin/bash tools/ci.sh --committed: all checks passed.
• With linters, as on Linux CI: a shellcheck shim (uvx --from shellcheck-py shellcheck) on PATH and uv run --no-project --with pyflakes -- bash tools/ci.sh --committed: shellcheck -S warning and pyflakes pass, all checks passed.
• New CI tests: the --budget lines; a WARN band at normal with exit 0, while a label and a "pace": "brisk" text pass at the floor; FAILs under the floor and under 1 s; the BRIEF's Pace (relaxed) and --pace brisk overriding it; a bad pace refused (exit 2); a project made by bin/vh new carries Pace: normal, and readcheck reads it.
• By hand: data-pace inheritance into a sub-composition and a bad data-pace note; bin/vh rhythm with a WARN and two FAILs; bin/vh recipes check 32/32.

After independent review

A fresh-context review asked for changes; these went in before merging:

• The floor is Videos move to a GitHub release; the intro film is re-paced to 103 s; the READMEs are rewritten #53's short-label rule again, max(1.5 s, CJK/7 + other/20 + 0.8 s). The first version had lowered it to 1 s and +0.5 s while saying it kept it, and that rule is the one the approved 103 s intro film was cut to ("not a flash either"). normal's pad is 0.8 s too, so the paces never cross. Recipes and CI recomputed.
• Style holds that are part of a style (editorial-data, monumental-scifi, symmetry-pastel, dark-math) stay 2.5 s, marked as style beats; type 01's question card stays a 2–3 s pause to think.
• Caps: studio's three rounds run even between small fixes; a gate ③ return restarts both counts; a floor FAIL at its cap goes to the human and never ships as final (unattended output is labelled). Changing Pace needs a DECISIONS line and, when the human owns rhythm, their say.
• bin/vh new writes Pace relaxed for math and paper, brisk for meme.
• 7 CJK chars/s is marked as an assumption; product-film-60s's breath beats are rhythm, not reading time.
• A targeted re-review approved; its two one-line nits are fixed.