Research note 06: concept-first against floors only; two self-checks for quick

[ZLHad]

Why

#44 made the workflow concept-first because the maintainer worried it over-constrained the model. Nobody had measured whether it helps. This PR runs a small A/B test and fixes what the test found.

The test (docs/research/06-concept-first-ab.md, English in en/)

• Requests: two one-line requests, both silent:
  • a 30 s explainer on why you can't see the Milky Way from a city;
  • a 20 s opening film for a bakery that sells one bread.
• Arms: four films by the same model (Opus 5.5 subagents, HyperFrames, about 45 min each).
  • A keeps only five floors and opens no workflow docs.
  • B follows the quick path card, including the concept step.
• Review: a fresh reviewer scored all four blind on the eight scoring dimensions, compared each pair, and rated how similar each pair's ideas were.
• Stillness: measured for all four films the same way (frame differences).

	Milky Way B	Milky Way A	Bakery A	Bakery B
Concept	7	6	7	8
Phone readability	5	7	7	8
Motion	6	6	7	5
Mean of 8	5.8	6.1	7.0	6.3
Stillness ≥ 2 s, total	12.8 s	14.3 s	9.0 s	10.2 s

What it found:

• The reviewer's choices: both times it found B more original, and both times it chose A to post. The means differ by only 0.3 and 0.7, likely within run-to-run noise.
• Why B lost: labels sitting exactly at the 44 px floor (about 8 px on a phone), and a key transformation done as a dissolve. Slowness is not the reason: the two arms are about as still, in opposite directions in the two pairs. B's stillness does start at the opening, but the hook scores don't tie that to the result.
• Same idea without the workflow: in one pair (Milky Way, 8/10 similar) the model reached much the same idea on its own.
• Other: all hooks scored 3–5, and time and tokens were about equal.
• Caveats: n = 2, one run each, a model reviewer, quick only, picture only. These are listed in the note.

Changes it led to

• quick self-check (CLAUDE.md quick card, step 5, and playbook/02's effort note): add a phone-size contact sheet and a first-2 s strip. playbook/02 gains the strip command and what 44 px becomes on a phone.
• playbook/12-ideation.md §6: readouts people must read should not sit at the floor.
• Digital-silence floor: scoped to films with sound (CLAUDE.md, playbook 12, TASTE_CHECKLIST Design notes: v0.3 roadmap #18). Both workflow agents had to reason this out.
• Type 02: caption weight falls back to the local font's heaviest weight.
• Type 03: a fictional brand name gets searched for real ones. One film named its bakery 一条, a known brand.
• Indexes: docs/research/README.md and en/README.md gain a row and figure 6.1. CHANGELOG entry added.
• Open question for the maintainer: the 44 px floor is written for 1080 wide. Should it scale for landscape (1920 wide)?

Checks

• tools/ci.sh --committed, also run with VH_BASH=/bin/bash: all passed.
• bin/vh effort quick prints the scoped floor.