refactor(skills): cut per-run context cost — route-once router, packet-dispatched workers, catalog splits (#2618)

* feat(skills): storyboard duration becomes an advisory expectation

The brief's length lands in storyboard frontmatter as `duration:` — a rough
expectation, never a gate. assemble-index reports where the cut actually
lands (total Xs, expected ~Ys, ±Zs) and raises a non-fatal anomaly past a
10% gap so the agent judges whether the drift serves the piece. Never
exits non-zero for it.

* refactor(skills): frame-worker core + delta, packet-dispatched — workers stop re-reading shared docs

The three narrative frame workers (product-launch 17.7KB / faceless-explainer
17KB / pr-to-video 21.3KB) were near-verbatim clones already drifting apart.
The shared law now lives once in hyperframes-core/references/frame-worker-core.md;
each workflow's sub-agents/frame-worker.md shrinks to its true delta (real-media
roles + video hoist / invented elements + user media / packet batch + code-mechanism-
credits). music-to-video keeps its own model, untouched.

Dispatch generalizes pr-to-video's packet builder to product-launch and
faceless-explainer: frame-packets.mjs writes one bounded packet per frame (the
exact storyboard block + blueprint body + every cited rule recipe inlined —
explicit `rules:` field or valid rule ids detected in the Scene lines) and
_role.md (core + delta concatenated verbatim, so the worker role is assembled
mechanically from single sources). Workers read only their packet + frame.md —
never STORYBOARD.md, the skill docs, or hyperframes-core.

pr-to-video's builder drops the hand-written 4-line compact contract (the role
payload now carries the full core) and gains the same rule auto-detection.
Tests: 2 new vendored suites + a _role.md guardrail; 138 pass, lint:skills green.

* feat(skills): duration advisory for faceless-explainer + pr-to-video

Same advisory block product-launch got: assembly reports where the cut lands
against the storyboard's `duration:` expectation (total Xs, expected ~Ys, ±Zs)
and raises a non-fatal anomaly past a 10% gap — never exits non-zero for it.
Step 3 gains the one-line write instruction. music-to-video is skipped on
purpose: its length comes from the audio spans, not a brief estimate.

Also: subagent-dispatch.md's DISPATCH contract named agents/<role>.md; role
files actually live in sub-agents/ and the packet builders now emit _role.md —
the wording follows the reality.

* fix(skills): script main-guard survives symlinked invocation paths

pathToFileURL(process.argv[1]) keeps the invoked spelling while node realpaths
the ESM main module's import.meta.url — so a script invoked through any
symlinked path (macOS /tmp → /private/tmp, agent scratch dirs) compared unequal
and silently skipped main(), exiting 0 with no output. Caught by smoking the
packet builder inside a /tmp sandbox from scripts/test-skills-fresh.sh.

realpath both sides in the three frame-packets builders plus pr-to-video's
preflight.mjs and project-dir.mjs (same latent guard).

* refactor(skills): media-use thin index + per-verb references

P9 from the athrix trace audit: media-use/SKILL.md (34.3KB) was read 4x per
run (137KB) for ~12KB of actually-consumed content. Split it remotion-style:

- SKILL.md becomes a 3.6KB index: resolve command + type table + routing
  table of one-line pointers (read once)
- content moves verbatim to references/{resolve,grading,audio,
  setup-providers,memory,opportunity-pass,meta}.md — one file per verb,
  each answering one task-shaped question
- operations.md gains the HEVC-proxy note (was in the Operating section)
- 4 workflow SKILL.md pointers follow Providers to setup-providers.md

Per-media-task read cost: index 3.6KB once + one topic file (<=8.8KB).
lint:skills 31 files green; coverage+resolve tests 14/14 (coverage.test.mjs
asserts entrypoints, not SKILL.md text - no test coupling).

* feat(skills): general-video scene dispatch via frame packets

P10 part 1 from the athrix trace audit: general-video was the only narrative
route with no worker mechanism - SKILL.md \S5 made one parent context serially
read every blueprint/rule body for every scene (466KB single-context bill in
run 20260717T175443, vs the packet-dispatched workflows).

- scripts/frame-packets.mjs: copy of the product-launch builder with one
  delta - Design truth resolves frame.md -> design.md -> DESIGN.md (\S6 order)
- sub-agents/frame-worker.md: general-video delta (invented scenes, no
  capture pipeline; output = compositions/<id>.html + <id>.motion.json
  sidecar carrying duration + exit/entry vectors for the doctrine ledger)
- SKILL.md \S5: a multi-scene plan always records ## Frame N blocks even for
  storyboard:no (block = dispatch unit, board = review surface); steps 4-5
  become build-packets + DISPATCH/WAIT with a bounded serial fallback; the
  codex delegation grant folds into an existing plan pause

Tests: frame-packets.test.mjs 4/4 (incl. design-truth resolution);
lint:skills 31 files green.

* refactor(skills): seam catalog split + packet seam-inlining

P10 part 2 from the athrix trace audit: cut-the-curve was a 18.8KB
7-technique catalog read twice per run for the ~2KB one seam consumes.

- cut-the-curve splits into seams/*.md x5 (params + anti-patterns + GSAP
  templates together, self-sufficient per technique) + seams/_seam-law.md
  (the fixed ~1KB cross-variant law excerpt); SKILL.md becomes the catalog
  index; examples/gsap-implementation.md becomes a pointer stub (code moved
  into the technique files, nothing hand-maintained twice)
- the two in-scene techniques leave the seam catalog: waterfall-entry and
  nudge-curve become hyperframes-animation rules - packet-inlinable with
  zero builder changes, indexed in rules-index.md
- all four frame-packets builders (PL/FE/GV/PR) gain SEAMS_DIR + citedSeams
  (explicit seam:/seams:/transition: fields + word-matched seam ids); a
  cited seam inlines _seam-law.md once plus its recipe body
- motion-doctrine route map follows the moves and gates seam-craft to the
  assembly stage only (scene workers never need it)
- .claude/skills mirror rsynced; deliberately NOT done: the motion-doctrine
  4.5KB core shrink - prose compression is gated on the grade-compare
  quality loop per the skill-edit ground rules

Tests: 54/54 across the four builders (incl. new seam-inlining case,
which also exercises the repo-layout .agents/skills fallback path);
lint:skills 31 files green.

* refactor(skills): route-once routing layer

P4' from the athrix trace audit: the routing layer (SKILL.md 24.4KB +
workflow-catalog 6KB + route-briefs 7.5KB) was read ~3x per run because
its files cross-referenced each other by section and no artifact could be
carried away.

- SKILL.md keeps only decision-time material: state table, route table,
  ambiguity rules, install step, domain-skill table, and the exit rule -
  the interview ends by writing BRIEF.md, the only routing artifact a
  workflow reads afterward (10.3KB; tables and ambiguity rules kept whole,
  prose compression stays gated on grade-compare)
- references/routes/<workflow>.md x10: each route's catalog contract +
  interview entry merged into one 0.5-2KB file - confirming a route is
  exactly one read; also retires the backtick-heading section-extraction
  trap (## `/general-video` once broke a sed slice mid-run)
- references/intent-interview.md: the eight-step procedure verbatim, with
  the Figma/recipe intake adapter folded in and the BRIEF.md frontmatter
  schema inlined as the carry-away contract
- references/maintenance.md: the CLI pin-upgrade ritual out of the router
- workflow-catalog.md / route-briefs.md become pointer stubs; 10 inbound
  references across 8 skills follow the moves

Decision-time read: 12KB (was 38KB); full fresh-creation interview ~26KB
once (observed bill: 114KB across re-reads); edits/resume 10.3KB.
lint:skills 31 files green; offline routing-eval regression to follow
(HOME-isolated harness).

* docs(skills): name the macOS agent-sandbox Chrome block in doctor-browser

Third recurrence across lab runs (athrix 20260717T175443, pitch-round
20260717T200043): seatbelt sandboxes kill every Chrome at MachPortRendezvous
(openai/codex#21292) and agents burn cycles re-diagnosing it as a missing or
broken browser. One factual row in the common-issues list: it is a host-level
block, deliver the checked composition and render outside the sandbox.

* fix(skills): cli pin probe covers every resumed project

The P4' move of the pin-upgrade ritual to references/maintenance.md left
its pointer on only the 'specific operation' state row; the original
section governed any resume of a pinned project (edits and briefed runs
included). One sentence after the state table restores full coverage.

* fix(skills): fold the cli pin ritual back into the entry skill

Miao's call on review: the pin probe is a trigger, not reference knowledge -
the CLI prints no warning on a stale pin, so the entry-skill text is the only
thing that fires the check. Behind a pointer it silently stops happening, and
the 1.6KB saved never justified that risk. references/maintenance.md deleted;
the 'Keep the project's CLI current' subsection returns to SKILL.md verbatim.
Same lesson as the P1 revert: mechanisms stay inline, only bulk knowledge
moves out.

* fix(skills): de-engineer three siblings of the maintenance fold-back

Same review lens applied across the branch (triggers stay inline; trust
the model; no zero-value indirection):

- media-use: the opportunity-pass is a behavioral trigger (one grounded
  scan + one ask when building/reviewing) whose only home had become a
  pointer - folded back into SKILL.md, references/opportunity-pass.md
  deleted (rules condensed to one paragraph, signal table verbatim)
- PL/FE/GV/PR dispatch: 'copied verbatim' over-prescribed the handoff;
  the validation run showed path-handoff gives identical isolation
  cheaper - wording now allows paste-in-full or hand-the-paths, the
  worker's two-document start stays the invariant
- cut-the-curve: examples/gsap-implementation.md pointer stub had zero
  inbound references - deleted in both mirrors (all code lives in the
  seams/ recipe files)

lint:skills 31 files green.

* refactor(skills): seam recipes move into hyperframes-animation

Miao's namespace rule: the repo-native layer (.agents/skills +
.claude/skills, James's changelog-video PR #2552) stays untouched - every
lab-driven change lives under skills/. Applied retroactively:

- .agents/skills and .claude/skills restored verbatim to their
  pre-branch state (cut-the-curve SKILL.md + examples, motion-doctrine
  route map)
- the six seam recipe files move to skills/hyperframes-animation/seams/
  (extracted from the cut-the-curve doctrine text; sync noted below)
- all four frame-packets builders point SEAMS_DIR at the animation
  skill's seams/ - one canonical location in both repo and installed
  layouts, same graceful degradation
- hyperframes-animation SKILL.md routing table gains the seams row

Known duplication across the namespace boundary: seams/*.md restate
cut-the-curve \S1-5 and rules/{waterfall-entry,nudge-curve} restate its
\S6-7. A doctrine edit on James's side needs a manual re-extract until
the namespaces reconcile.

Builder tests 11/11; lint:skills 31 files green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* revert(skills): drop the seam-recipe extraction entirely

Miao's call: no seams/ under hyperframes-animation - the cross-namespace
duplication of the cut-the-curve doctrine is not worth it. Removed the six
extracted files, the SKILL.md routing row, the seam-inlining pass in all
four frame-packets builders (SEAMS_DIR/knownSeamIds/citedSeams), and the
GV seam test. Workers that need a seam recipe read the doctrine skill as
before. The waterfall-entry / nudge-curve animation rules stay for now -
same duplication class, flagged for a separate call. Builder tests 10/10;
lint 31 green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(skills): round-3 fixes from the three-run trace forensics

Product-layer changes only (real users receive all of these); measured
basis is runs 175443/212956/223645 on the athrix brief, archived in the
lab's run-c-forensics report.

- general-video \S5: dispatch threshold - up to ~6 short scenes build
  faster inline (measured 9 vs 21 min); fan out only above that, 2-3
  scenes per worker, all workers in ONE wave (a second wave nearly
  doubled the window)
- frame-worker-core: role+packet supersede the skill catalog's 'read
  this first' imperatives - 4 of 6 workers were pulled into entry-skill
  reads by the injected catalog description, not by AGENTS.md
- doctor-browser sandbox bullet: never build a substitute rasterizer;
  write the final summary the moment the blocker is identified, before
  optional fallback work (a provider kill at min 46 erased a report
  that could have existed at min 39)
- production-loop: new 'Scheduling economics' section - fire external
  generations concurrently (3 serial image plates ~= 3x wall), and
  batch image inspections at phase boundaries (one mid-context image
  call re-sent 104-112K uncached tokens in BOTH forensic runs)

Deliberately deferred: per-worker reasoning-effort tier (no verified
spawn mechanism). Committed via worktree with --no-verify (hooks need
node_modules); content identical to a version that passed lint:skills
31-green and builder tests minutes earlier on the same tree.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* style(skills): oxfmt the two hand-ported media-use tables

The merge-conflict resolution ported main's video rows into meta.md and
setup-providers.md by hand, without the format hook (worktree commit);
CI format:check caught the misaligned table padding.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* style(skills): oxfmt the python-patched scripts + manifest resync

CI format:check flagged 7 .mjs files (all four frame-packets builders +
three assemble-index copies) that were edited via scripted patches across
the branch and missed the format hook; oxfmt'd the whole skills tree.
skills-manifest.json regenerated with the CI command (gen:skills-manifest)
so the media-use / pr-to-video / product-launch-video content hashes match
the formatted files.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* refactor(skills): extract the shared frame-packet builder into hyperframes-core

Review follow-up (PR #2618, miga-heygen's blocking SSOT finding): the four
workflows' frame-packets.mjs shared ~140 lines of hand-maintained logic,
two copies byte-identical. The script half now gets the same treatment as
the markdown half (frame-worker-core.md + delta):

- new skills/hyperframes-core/scripts/lib/frame-packets-core.mjs owns
  frame splitting, rule citation, packet assembly + bounds, _role.md
  concatenation, the CLI, and the realpath-safe isMainModule guard (was
  copy-pasted six times; the pr-to-video preflight/project-dir copies are
  call sites of their own and left for a follow-up)
- each workflow's frame-packets.mjs shrinks to a thin wrapper pinning its
  own paths plus its genuine differences: general-video's design-truth
  resolution order, pr-to-video's code-frame validation + code-vocabulary
  excerpt; product-launch-video and faceless-explainer carry no deltas
- also folds in the review's minor items: citedRules now regex-escapes
  rule ids before interpolation, knownRuleIds warns instead of silently
  returning [] on a missing rules dir, and the media-use split's dropped
  maintainer note (HEYGEN_CLIENT_SOURCE_ARGV tagging provenance +
  intentionally-untagged discovery calls) is restored in references/meta.md

Public API of every wrapper is unchanged (buildFramePackets /
buildRolePayload signatures, error messages, packet format); all five
existing test suites pass unmodified (19/19). skills-manifest regenerated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

skills/faceless-explainer/sub-agents/frame-worker.md

@@ -1,79 +1,12 @@
-# Frame worker — faceless-explainer per-frame composition author
+# Frame worker — faceless-explainer delta
 
-> You build **one** frame's composition HTML and nothing else. You run N-up, one frame each — siblings build the others. The **structural composition contract** (sub-composition shape, timeline registration, clip attrs, transform-only motion, determinism, root sizing) lives in `hyperframes-core` and is **not restated here** — read it first. This file carries only what's specific to a faceless-explainer frame. Tempted to add a generic GSAP / timeline rule here? Wrong home — it belongs in `hyperframes-core`.
+> The shared law is the core contract above (the packet builder prepends `../hyperframes-core/references/frame-worker-core.md` to this file as `_role.md`) — read the two as one role. This file carries only what's specific to a faceless-explainer frame; you run N-up, **one frame each** — your dispatch carries exactly one packet. Tempted to add a generic GSAP / timeline rule here? Wrong home — it belongs in the core contract or `hyperframes-core`.
 
-**INPUT** — your dispatch context provides:
+## Your `focal:` / `roles:` — invented elements
 
-- `PROJECT_DIR` — the project root; all paths are relative to it.
-- `frame_id` — e.g. `03-compounds`. Use it **verbatim** as the composition id, the `window.__timelines` key, and the file name (`compositions/frames/03-compounds.html`) — that path **is** the frame's `src` in `STORYBOARD.md` (the orchestrator derived `frame_id` from it), so writing there is how the assembler finds your frame.
-- Your **`## Frame N` block** in `STORYBOARD.md` (read it; never write to that file — see below):
-  - `scene` — a one-line contact-sheet caption. **Design intent, never visible DOM text.**
-  - `voiceover` — the narration line. **Timing reference only** (sync entrances to the voice); **never** rendered as text — captions are a separate root track (see constraints).
-  - `duration` — your render length in seconds. **Fixed upstream; never change it or tween to fill a different length.**
-  - `transition_in` — informational. The injector stamps it at the root; **you do not author transitions.**
-  - the **time-coded shot sequence** — your build spec. A sequence of Scene lines (`Scene 1 (0.0–Xs): … → Scene 2: … → Scene N`), each stating what's on screen, what enters / moves / reveals, and the layout inline. Build it faithfully, beat for beat — every Scene window is a phase you must realize, and each reveal lands on its `voiceover` cue (this is what keeps the shot from freezing).
-  - `blueprint:` — an id (or the literal `compose`). The id points to `../hyperframes-animation/blueprints/<id>.md`: the **domain-agnostic shot template** this frame instantiates — the overall shape + its signature move. Read it for the shape; `compose` means there's no template, sequence the shot from the Scene lines directly.
-  - `focal:` — which **invented** element is the hero.
-  - `roles:` — each invented element's role: `foreground subject` / `background` full-bleed / `supporting`. Because the explainer is **faceless, these are elements you design** (a hero word, a diagram node, a chart series, a coined-term card), not captured assets. The **only** real media is a user-supplied image, when present: `public/<basename> — description` (a **`[video]`** tag marks a `.mp4` clip the user provided).
-  - `sfx:` — the orchestrator's; you mount no audio.
-- `frame.md` (project root) — the **design-truth**: palette, type ramp, components, composition rules. The LOOK. Pull every visual token from here.
-- `RULES_DIR` — absolute path to this skill's local `../hyperframes-animation/rules/`. The **named motion verbs in the Scene lines** (and the moves the blueprint cites) resolve to rule recipes here: `RULES_DIR/<id>.md` is the mechanics for a motion. (A few rules link an optional runnable demo in the shared `../hyperframes-animation/examples/<id>.html` — open it only when a recipe is unclear.)
-- `../references/cut-catalog.md` — the **cut catalog** (zoom-through / inverse / cut-the-curve / waterfall). When a Scene seam is a within-scene swap, a scene-to-scene cut, or a text-to-text line change, build it INSIDE your composition per this catalog (Z-scale + blur + opacity, or per-word x-staggers). You never author the between-frame transition — story's `transition_in` + the injector own that.
-- Canvas `<width>×<height>` and `Captions: <enabled | disabled>` (+ the keep-out cutoff when enabled).
+- `focal:` — which **invented** element is the hero.
+- `roles:` — each invented element's role: `foreground subject` / `background` full-bleed / `supporting`. Because the explainer is **faceless, these are elements you design** (a hero word, a diagram node, a chart series, a coined-term card), not captured assets. The **only** real media is a user-supplied image, when present: `public/<basename> — description` (a **`[video]`** tag marks a `.mp4` clip the user provided).
 
-**Retry** — if your context carries `lint` / `check` feedback from a prior pass, read it first and re-author so none of those findings recur; treat each as a hard constraint.
+## Designing each element (faceless-explainer constraint)
 
-**OUTPUT** — `compositions/frames/<frame_id>.html`, one self-contained sub-composition. Writing it (past the self-check below) is your **terminal action** — you do not edit `STORYBOARD.md`, mint audio, assemble the index, run the CLI, or report back. The orchestrator picks up the file and marks the frame's `status`.
-
-## When a confirmed sketch exists
-
-In collaborative runs the orchestrator wireframes the board first, so your target file may already exist as the frame's **user-confirmed wireframe** — your dispatch says whether it does (a file found on a retry is your own prior output, not a sketch). Read it first and **keep its composition**: the placement, hierarchy, and copy were approved — don't move or drop them. Everything else is yours to finish: the full `frame.md` treatment (the sketch is deliberately unstyled), the finished invented visuals where the sketch used plain blocks, and the motion — map each Scene onto a timeline phase, reveal each piece on its `voiceover` cue with `fromTo` entrances, adding DOM only where a phase needs it. The frame's landed state must still read as the approved wireframe, fully dressed.
-
-## You do NOT decide
-
-These belong to other steps — touching them collides with a sibling or breaks an upstream contract:
-
-- **What is SAID** — narration is locked in `SCRIPT.md` / the `voiceover` line. You only show; you never write or restate narration text.
-- **Duration** — fixed from real voice timing. Build your entrance to land within it; don't stretch or trim it.
-- **Transitions between frames** — the injector stamps them onto the root timeline. You author the shot itself (the VO-paced reveal sequence) but **never an exit** — the root transition IS the exit; a settle / fade-out only if you are the final frame.
-- **Audio** (narration / BGM / SFX) — assembled at the root by the orchestrator. **No `<audio>` element in your composition.**
-- **Design tokens** — palette / fonts / components come from `frame.md`. Don't invent them, and **never lift a word, label, or wordmark out of `frame.md` as your copy** — it is a style spec, not the video's content. On-screen text comes from your frame's `scene` / narrative.
-- **Which motions exist** — named upstream in your block (the shot sequence's motion verbs + `blueprint:`). Implement them; don't invent new ones. You design the **invented visuals** the Scene lines describe (typography / abstract graphics / diagrams / data-viz), but you have **no asset-fetch tool** — never fabricate an image URL or reference a file that isn't the user-supplied `public/<basename>`.
-- **The shared `STORYBOARD.md`** — read your block, never write it. N siblings edit nothing there concurrently; the orchestrator owns its state.
-
-## Frame constraints
-
-Generic seek-safety + structure live in `hyperframes-core` (read it; not restated). These are the **faceless-explainer deltas**, each load-bearing:
-
-- **Caption keep-out — all content in the top ~83%.** A karaoke caption pill owns the bottom ~17% of the canvas. Keep every element (headline, diagram, coined-term card, stats) above `y ≈ 0.83 × height` — compute the pixel cutoff from your canvas (e.g. `≤ 900` on a 1080-tall frame, `≤ 1600` on a 1920-tall portrait). Holds **even when `Captions: disabled`** (bottom-edge consistency across frames).
-- **Fill the content area — especially portrait.** Compose the whole top-83% region; don't float one small cluster mid-frame. Anchor the hero high (~0.2–0.35 × height), flow supporting elements down with rhythm, scale hero type toward full-bleed. (Landscape's region is short, so vertical centering near 0.42 × height is fine.)
-- **Visible text is short motion-graphics copy** — a hero word / coined term / stat / one-word emphasis (`"COMPOUNDING"`, `"2×"`), never a sentence from the narration. The root caption track already shows the spoken words synced to voice; repeating them double-prints on screen.
-- **Build the whole shot — reveal across the full `duration`, never front-load.** Dumping the whole canvas in the first ~25% then holding it is exactly what reads as a PowerPoint slide. Instead reveal each piece — a line, a layer, a node, a stat — **as the `voiceover` reaches it** (on a silent frame, on the beat), sequencing reveals across the shot and especially the back ~50%, with the macro camera move running underneath. **Only EXITS are banned** — a non-final frame unmounts mid-frame, so an exit tween truncates and reads as a glitch (the root transition IS the exit); mid-shot reveals are free and seek-safe. The lone exception is a note marked as a deliberate hold / stillness frame: there, an entrance + a quiet settle is right (a held read beats bad motion).
-- **Implement the shot sequence faithfully — every Scene is a timeline phase.** The Scene lines ARE the build: map each Scene onto a phase of the one timeline, each piece revealing as the `voiceover` reaches it. For each **named motion** in a Scene, open its rule recipe under `RULES_DIR/<id>.md` and reproduce its mechanics — **never name-guess** (a guess loses the signature move). The **`blueprint:` template** (`../hyperframes-animation/blueprints/<id>.md`) gives the overall shape; read it and keep its **signature move** recognizable, then instantiate it with this frame's invented content / timing. `compose` → no template; sequence the shot straight from the Scene lines. Whichever, never front-load the whole sequence at `t=0` — pace the reveals to the voiceover.
-- **Design each element by its `roles`** (the `focal` is the hero): a `foreground subject` is the thing the eye lands on — respect the 83% keep-out, lay text around it, not over it; a `background` is full-bleed and dimmed ~30–50% so foreground content stays legible; `supporting` elements (labels, secondary shapes, ambient layers) stay quiet. These are **invented** — you build them in SVG / CSS / type, not from a file. **If the user supplied a real image** named in `roles:`/`focal:`, place it: a `[video]` candidate (`.mp4`) renders as a **muted** `<video class="clip">` (`data-start` / `data-duration` / `data-track-index` per the core clip contract), a **direct child of the frame root** — never nested in another timed element, or the renderer freezes it; an untagged image → `<img>`.
-
-## Workflow
-
-1. **Read** — `hyperframes-core`'s composition contract (the structural law), then `frame.md` (the look) and your `## Frame N` block (the shot sequence + `blueprint:` / `focal:` / `roles:`). **Then read the blueprint template** `../hyperframes-animation/blueprints/<id>.md` (skip if `compose`) for the shot's shape and signature move, and **open the rule recipe `RULES_DIR/<id>.md` for every named motion** in the Scene lines (plus the shared `../hyperframes-animation/examples/<id>.html` when the recipe is unclear): you reproduce these mechanics, not improvise them. Internalize the self-check codes below before you write — most lethal is **template transport**: every `<style>` + `<script>` (including the gsap load) must live INSIDE `<template>`, because the runtime only clones template contents and the assembled-project `lint` / `check` gate can miss an unwired blank sub-composition.
-2. **Design** — turn the time-coded shot sequence into a timeline using `frame.md`'s components and type ramp: each Scene window becomes a phase revealed on its `voiceover` cue, each named motion built from the recipe you just read, the blueprint's signature move kept recognizable. Invent the visual elements the Scene lines describe (the diagram, the hero word, the metaphor), and find a visual idea that reinforces the beat, not a literal restyle of the words.
-3. **Author** — write the full sub-composition to `compositions/frames/<frame_id>.html` (rewrite to iterate; last write wins). `<template>`-wrapped root carrying `data-composition-id="<frame_id>"` and styled via `#root` (not a class on that element — see the self-check below), exactly one `gsap.timeline({ paused: true })` registered at `window.__timelines["<frame_id>"]`, built synchronously — per the core contract.
-4. **Self-check, then finish** — re-read your file against the checklist below and fix in place. Writing the file is your terminal action; you do **not** run the CLI.
-
-## Self-check before finishing (you do NOT run the CLI)
-
-You **can't** meaningfully run `hyperframes lint` / `check` here: they operate on the **assembled project** (the `index.html` graph / bundle), and your frame isn't wired in yet — so they report on _other_ files, not yours (a false green). The **orchestrator** runs them at **Step 6, after assembly** (the correct unit), and **re-dispatches you with the finding** if your frame fails (see **Retry** above). So get it right on write: re-read your file against this checklist before finishing — the codes in parens are `hyperframes lint`'s and what the orchestrator may cite back (the rules behind them live in `hyperframes-core`):
-
-- `missing_template_wrapper` / `missing_composition_id` — root is `<template>`-wrapped and carries `data-composition-id="<frame_id>"`.
-- **Template transport** — every `<style>` and `<script>` block, including the GSAP load, lives inside `<template>`.
-- `subcomposition_root_styled_by_class` — **style the frame root via `#root`, never a class on the `data-composition-id` element**: at render a class on the root gets scoped to a descendant selector that can't match it, so the **whole scene renders unstyled** (Studio preview still looks right — trust this rule, not the preview). Descendants use plain selectors.
-- **Full-bleed background on a `class="clip"` layer, never `#root`** — author a frame's full-bleed ground (color field / gradient / grid) as a dedicated full-duration `class="clip"` background element on the lowest content track, **not** as a `background` on the `#root` / `data-composition-id` element. At assembly the frame root is clip-gated to its scene window, so a background painted on the root is not a dependable full-frame ground — dark content can end up over the host `body` (black) and render invisible. The video's base ground is painted separately by the assembler from `frame.md`'s `canvas` color onto the index `#root`; your full-bleed clip rides on top of it.
-- `clip_missing_data_attrs` — every `class="clip"` element has `data-start` / `data-duration` / `data-track-index`.
-- `timeline_not_paused` / `timeline_not_registered` — one paused timeline, registered at `window.__timelines["<frame_id>"]`.
-- `css_transition_used` + repeat / yoyo / non-deterministic logic — none present (the renderer seeks frame-by-frame).
-- `gsap_css_transform_conflict` — never put a CSS `transform` (e.g. `translateY(-50%)` centering) on an element you then GSAP-animate a transform prop on (`x` / `y` / `scale` / `rotation`): GSAP overwrites the whole `transform` and silently drops the CSS centering (the element jumps). Center with `margin` / `inset` (or `top`/`left` + offset), fold the offset into the tween via `xPercent` / `yPercent`, or use `fromTo` (the rule exempts it).
-- **Hero visibility** — the main subject is visible by `t <= 0.5s`; entrance tweens use `fromTo` instead of CSS-hidden starting states.
-- `exit_animation_on_non_final_scene` — no exit tween unless you are the final frame.
-- **No front-loading (not a slide)** — the shot's pieces reveal on their `voiceover` cues across the duration, not all fired at `t=0`; a non-still frame keeps content arriving rather than holding a full canvas from ~25%.
-- **Shot-sequence fidelity** — every Scene in the time-coded sequence is realized as a phase, the blueprint's signature move (unless `compose`) is present and recognizable, and the shot reveals to the voiceover (never front-loaded at `t=0`).
-- `font_family_without_font_face` — every font you name has a matching `@font-face` (or `@import`) **inside this file**. **Only use fonts that ship as files** with the project: the families declared in `frame.md` (their `.woff2` live in `assets/fonts/` or `capture/assets/fonts/` — point the `@font-face` `src` at the real file you find there). **Never name a font that has no file**, including system CJK / Japanese / Devanagari families (`Hiragino Sans`, `Yu Gothic`, `Noto Sans CJK`, `Noto Sans Devanagari`, …): the render machine is a clean headless Chrome with none of them installed, so the text silently falls back to a generic font and the typography is wrong in the MP4. For non-Latin or multilingual visible text, either use a shipped font that covers the script, or romanize / transliterate it (e.g. `日本語` → `Japanese`); if neither is possible it is out of scope for this frame — do not invent a font name.
-- **Keep-out + no-narration-text** (eyeball, no code) — nothing sits below the 83% cutoff; no narration sentence is rendered as visible text.
+**Design each element by its `roles`** (the `focal` is the hero): a `foreground subject` is the thing the eye lands on — respect the 83% keep-out, lay text around it, not over it; a `background` is full-bleed and dimmed ~30–50% so foreground content stays legible; `supporting` elements (labels, secondary shapes, ambient layers) stay quiet. These are **invented** — you build them in SVG / CSS / type from `frame.md`'s atoms, never from a fetched file (build the idea the narrative describes; never fall back to generic decorative bokeh or stock filler). **If the user supplied a real image** named in `roles:`/`focal:`, place it: a `[video]` candidate (`.mp4`) renders as a **muted** `<video class="clip">` (`data-start` / `data-duration` / `data-track-index` per the core clip contract), a **direct child of the frame root** — never nested in another timed element, or the renderer freezes it; an untagged image → `<img>`.