@rig-c/create 0.0.0-stage → 1.0.1

package/docs/INTERVIEW.md

@@ -0,0 +1,886 @@
+# The interview
+
+The direction a user gives before any rig work starts, and the shape every
+answer takes in `project.json`. The interview is organised by the **waypoints**
+the rig work always passes through — the sections of a `@rig-c/parts` config —
+and the user's decisions attach at each one. New adopted records use [per-ask choices](#per-ask-choices-interview2); the sentence format below remains historical. Camera, 2.5D depth and timing are
+**post-production**, asked separately after the rig is accepted. The plan the
+work follows is drafted from these answers ([PROJECT.md § The plan](PROJECT.md#the-plan));
+a worked example is in [../templates/example-interview.md](../templates/example-interview.md),
+with its record in [../templates/example-project.json](../templates/example-project.json).
+
+Why this exists: an agent rigging a painting makes direction calls at every
+waypoint — which parts move on physics, where the torso is cut, what stays
+rigid, whether the hands move. Made alone, each of those calls is only found out
+when a person watches the result and sends it back, and every bounce re-runs a
+full build and review. The interview moves the calls to before the work. A user
+who answers "auto" has chosen to hand that call to the agent, and the record
+says so.
+
+## The user's three steps
+
+The user sees parts/mesh review, the rigging questionnaire, rig/animation correction, and direction. The historical three-step grouping below applies to legacy records; new projects use the split order in PROJECT.md. Everything else in this file — waypoints, states,
+provenance — is the machinery under them and is not something the user has to
+learn; the interviewer asks in plain questions and never shows a waypoint id or
+a form.
+
+| User step | In this file | When |
+| --- | --- | --- |
+| 1. **Base motion** | Framing (`rig_type`, `reference`), then a guide at each rig waypoint: `seethrough`, `assemble`, `bones`, `meshes`, `constraints`, `motion` | stage `interview`, before any rig work; the parts judged at `parts-review` before rigging, the rig at the first review, on a still frame |
+| 2. **Direction** | Post items: `camera_depth`, `post_timing`, `post_composition` | stage `post-interview`, after the rig is accepted; judged at the second review |
+| 3. **Polish feedback** | Verdicts on a shown version, each tied to a waypoint or post item | at either review, as often as it takes |
+
+## Every item has an owner
+
+No answer is left empty for the agent to fill later. Each question ends in
+exactly one of four states:
+
+| State | Meaning | Who writes it |
+| --- | --- | --- |
+| `null` | Not asked yet. The interview cannot be signed while a framing answer or the guide of a waypoint the plan passes is `null`. Post items are asked when `post-interview` opens, not before. | — |
+| user answer | The user's own words, verbatim, plus a normalised `value`. "None" is a user answer (`"value": "none"`), not a delegation: *there is no reference* is a decision. | the interviewer, from the user |
+| delegated | The user said "auto" (or "you decide", or pressed a control that means it). The agent decides that item and writes its decision and reasoning into `agent_decision` beside the delegation. | the interviewer records the delegation; the agent fills `agent_decision` |
+| not applicable | A prompt's `required_for` does not include the answered `rig_type`, or the pipeline for this project has no such section (a pre-layered input skips `seethrough`). Derived, not judged, so the interviewer may write it. | the interviewer |
+
+Rules that follow from the table:
+
+- **A delegated item stays delegated.** Signing the plan that shows the agent's
+  decision does not turn it into a user answer. If the user wants it to be
+  theirs, they answer it: the new answer replaces the old, and the old one moves
+  to that question's `history[]`. Reason: a bounce on a delegated item is a cost
+  the user chose, and the record must still say so after the plan is signed.
+- **The agent's decision on a delegated item is written before the plan is
+  signed**, so the user sees every agent choice the plan rests on at the moment
+  they sign. A delegated item the plan does not need yet may wait, but its
+  `agent_decision` is written before the step that uses it starts.
+- **A bounce is attributed to the items it is about.** Each verdict lists
+  `on_items[]`: the answer ids it concerns and whether each was a user answer,
+  delegated, or never asked ([§ Polish feedback](#polish-feedback)).
+- **A changed answer reopens what rests on it.** Every plan step whose
+  `rests_on` names that answer needs the plan signed again before it runs.
+
+`rig-create interview --check` is the machine check of these rules — refusing
+by name every waypoint with neither a guide nor a delegation, every step that
+rests on no answer, every stage moved without a sign-off; this document is the
+contract it checks and [CHECKS.md](CHECKS.md) lists its codes. Delegation is a
+pass for that check, never a hole: the refusal fires only on silence.
+
+## Answer shape
+
+New adopted projects use [per-ask choices](#per-ask-choices-interview2). The
+examples in this section describe the preserved legacy sentence format.
+
+Every record in `project.json` that a person made — an answer, a sign-off, a
+verdict — carries the same three fields: `by` (who), `at` (when, ISO 8601) and
+`verbatim` (the words used). One shape, so a reader never has to learn which
+record calls the date what.
+
+A user answer:
+
+```json
+{
+  "delegated": false,
+  "by": "user",
+  "at": "2026-11-10T14:05:00Z",
+  "verbatim": "Breath and blink. The scarf lifts when the wind comes.",
+  "value": [
+    { "motion": "breath" },
+    { "motion": "blink" },
+    { "motion": "scarf_lift", "note": "only with the wind" }
+  ],
+  "source": "interview, chat"
+}
+```
+
+A delegated answer, after the agent has decided:
+
+```json
+{
+  "delegated": true,
+  "by": "user",
+  "at": "2026-11-10T14:07:00Z",
+  "verbatim": "auto",
+  "source": "interview, chat",
+  "agent_decision": {
+    "value": { "passes": ["full", "head"] },
+    "by": "agent",
+    "at": "2026-11-10T15:20:00Z",
+    "reasoning": "Blink needs the eyelids as their own layers; at full-figure resolution the face is too small for the split to find them."
+  }
+}
+```
+
+Not applicable:
+
+```json
+{ "not_applicable": true, "because": "rig_type = fullscreen.minimal", "by": "interviewer", "at": "2026-11-10T14:09:00Z" }
+```
+
+- `verbatim` is kept in the language the user used. `value` is the interviewer's
+  normalisation of it; when the two disagree, `verbatim` wins and `value` is
+  corrected.
+- `source` says where the words were given (chat, a call, a note) so the
+  record can be traced; chat is not a record — the decision is copied here.
+  A **derived answer**, taken from words already on record instead of asked
+  again, says so here ([§ Derived answers](#derived-answers)).
+- A sentence may leave part of a waypoint open ("the rest is up to you"). Then
+  the agent's choice for that part is written to `agent_decision` beside the
+  sentence, with `covers` naming what it fills, and that part reads as
+  delegated.
+- A guide may grow after the interview: a sentence added from polish feedback
+  or from an `x_` prompt goes into the guide's `additions[]`, each with `by`,
+  `at`, `verbatim`, `value` and `source`. The original sentence is not edited,
+  so a provenance `quote` taken from it stays findable; an addition is quoted
+  the same way.
+- A replaced answer moves whole into the question's `history[]`; the record
+  never loses what was once decided.
+- Collection members add a `basis` field to every answer
+  ([COLLECTIONS.md § Inherited, overridden, own](COLLECTIONS.md#inherited-overridden-own)).
+
+### Derived answers
+
+Before asking, the interviewer reads what is already on record — earlier
+answers, an intake note in `inputs/`, a reviewer's past verdicts — and takes
+from it every answer those words already give. Only what is left is asked, in
+one short message. Asking a user what their own earlier sentence already
+answered is a cost the record could have saved.
+
+A derived answer is an ordinary user answer; only `source` says where it came
+from:
+
+```json
+{
+  "delegated": false,
+  "by": "user",
+  "at": "2026-11-08T10:12:00Z",
+  "verbatim": "Only the chest breathes; the hips stay where they are.",
+  "value": [{ "motion": "breath", "note": "chest only, hips still" }],
+  "source": "derived from inputs/intake.md"
+}
+```
+
+- `by`, `at` and `verbatim` are those of the words on record — who said them,
+  when, the exact fragment — not of the derivation. A reviewer's past verdict
+  keeps the reviewer as `by`.
+- `source` starts with `derived from` and names the record: a file under
+  `inputs/`, an answer (`interview.waypoints[motion].guide`) or a verdict
+  (`versions[v1].verdicts[0]`). The `verbatim` fragment is found in that
+  record, so anyone can trace it.
+- Only words are derived. An answer the record merely implies, and anything
+  read off the image, is still a question.
+- The message that asks what is left lists each derived answer in one line, so
+  the user can correct it; a correction replaces it, and the derived answer
+  moves to `history[]`.
+
+For unadopted records no new field is needed: `interview --check` reads a derived answer as the user
+answer it is, and a provenance `quote` taken from it resolves against its
+`verbatim` like any other.
+
+## Direction and exact words
+
+In unadopted records, the split and direction are received in the user's own sentences —
+"cut the bust, not the waist, and give it jiggle physics", "the hem moves in the
+wind down to the knee" — not as a drawing or a form. The prompts below draw
+those sentences out; they are not fields to fill.
+
+- The sentence is the answer, stored `verbatim`. `value` is the agent's
+  normalised reading of it and never replaces it. One sentence may answer
+  several questions; it is stored once and referenced.
+- Every part, effect, motion or extent a sentence names is a line the work is
+  checked against: a part said to be cut must exist as a part; an effect said to
+  be on a part (wind, physics, jiggle) must be on that part and on no part the
+  sentence excludes; an extent ("down to the knee") bounds what moves. A
+  sentence that says "must" is a bar, not a preference. Turning these lines into
+  `@rig-c/core` `invariants` starts with the pure translator in
+  [`src/invariants.ts`](../src/invariants.ts), documented in [INVARIANTS.md](INVARIANTS.md).
+  Its sourced entries and explicit gaps are reproduced in
+  [`example-invariants/invariants.json`](../templates/example-invariants/invariants.json).
+  Project/check/build wiring remains a follow-up; until then these lines are
+  read at review against the version's provenance.
+- Anything the agent shows back — the labelled split image, a motion sketch — is
+  its reading of the sentences for the user to sign, not the user's input.
+
+## Waypoints: where the guide attaches
+
+A waypoint is a place the agent always passes while rigging, with the user's
+guide waiting at it. The waypoints are not a hand list: **they are the top-level
+sections of the `@rig-c/parts` config the work writes**, grouped by the decision
+made there. A new config section is a new place a decision is made, so it joins
+a waypoint (or becomes one) without anyone editing a list. The template's
+waypoint list is a snapshot of the sections at the package version named in
+`interview.question_set`; project creation is expected to regenerate it from the
+installed package.
+
+Not every section is a waypoint. A section is one when writing it is a
+**direction call** — something the user could have an opinion about before
+seeing a build. Read from `@rig-c/parts` 2.4.0:
+
+| Config section | Where it goes | Why |
+| --- | --- | --- |
+| `seethrough` | waypoint `seethrough` | how the painting is decomposed into layers |
+| `assemble`, `limbs`, `scene`, `waist` | waypoint `assemble` | what becomes a part, where the cuts are, what the painting is, which outfit sets exist (`assemble.alternates` is a field of `assemble`, not a section of its own) |
+| `bones` | waypoint `bones` | where bones sit, which parts get chains |
+| `meshes`, `regions` | waypoint `meshes` | what deforms and what rides rigid |
+| `constraints`, `stops`, `cloth` | waypoint `constraints` | what moves on physics and how far |
+| `motion` | waypoint `motion` | what moves, how far, how fast |
+| `turn`, `depth` | post item `camera_depth` | head-turn and 2.5D depth are direction, asked after the rig is accepted |
+| `key`, `generation`, `proposed_rig_scale`, `requirements`, `motion_seams`, `note` | not a waypoint | a name, a record of a run, a value the tool proposes, a path to a checks file, an observation, an annotation — none is a direction call. `requirements` is where translated guide sentences will land, not where they are asked. |
+
+The agent writes the checks file named by the config's `requirements` path.
+Declare `setup_painting.required: true` there. Measure its limits on the build
+the user accepted at the split checkpoint, using parts' own `pixel_count` per
+region on that build. Record them as a "no worse than the accepted split" bar
+with its source, including the accepted build and the measurement. Create
+chooses no limit values. Parts refuses missing limits by name with
+`SETUP_PAINTING_REQUIRED_LIMIT`, and enforces the declared limits through
+`SETUP_MATCHES_PAINTING`. The file also needs an authored `fps`, which has no
+default. Create writes no config or requirements file during init.
+
+A section that cannot be placed on this table when it appears is a question for
+the user before it is written. The table lives as data in `src/waypoints.ts`,
+and the section list is read from the installed package, so
+`interview --check` refuses such a section by name ([CHECKS.md](CHECKS.md)).
+
+**Every waypoint the plan passes carries a guide — the user's sentences or an
+explicit delegation. Nothing is decided silently.** The guide is one answer in
+the shape above, stored on the waypoint (`interview.waypoints[].guide`). A
+waypoint the plan does not pass is answered *not applicable* with the reason,
+and is not asked. "None" is a guide: *nothing moves on physics* is a decision
+about `constraints`.
+
+Each waypoint has **prompts**: the questions the interviewer asks to draw out the
+sentence. `rig_type` and `reference` frame everything and are asked first,
+because `rig_type` decides which waypoints are passed and which prompts apply,
+and `reference` decides what every review compares against.
+
+| # | Waypoint | What the agent decides there | Prompts | Guide sentence, example |
+| --- | --- | --- | --- | --- |
+| F1 | framing | which waypoints are passed; the plan's outline | `rig_type` | *"A standing character, idle; she holds a lantern."* |
+| F2 | framing | what reviews compare against | `reference` | *"The idle in our trailer, but slower."* |
+| W1 | `seethrough` | how the painting is decomposed into layers: resolution, the head crop, the seed | `x_layers` | *"Keep the hat with the head; the lantern she holds is its own layer."* |
+| W2 | `assemble` | which layers become which parts and in what draw order, where limbs are cut, what is authored new, what is kept under occluders, which outfit sets exist | `split_granularity`, `split_emphasis` | *"Cut the bust, not the waist."* · *"Keep her face exactly as painted."* |
+| W3 | `bones` | where bones sit and which parts get chains (hair, coat tails, sleeves) and how long | `x_chains` | *"Two chains in the back hair, one in each coat tail; the scabbard stays rigid on the hip."* |
+| W4 | `meshes` | which parts deform as weighted meshes and which ride one bone as rigid regions; mesh density | `x_deform` | *"The waist stays one piece that bends; the face is rigid on the head."* |
+| W5 | `constraints` | which parts move on physics, how strongly, with wind and gravity, and where a chain must stop | `split_effects` | *"The hem moves in the wind down to the knee; the bust jiggles; the sleeves barely move."* |
+| W6 | `motion` | the tracks (what moves, how far, how fast), the loop length, blinking; for a game-logic scene, the states | `motions`, `game_states` | *"A slow breath, about three seconds; blink now and then; the head does not sway."* |
+
+[CRAFT.md](CRAFT.md) holds the craft notes: per waypoint, the default the
+interviewer may propose ("Usually X — shall we?"). The interviewer may read one
+out as a proposal; the user's reply is the guide, and a default nobody was asked
+about decides nothing.
+
+### Prompts
+
+Each prompt keeps its `id`, its `text` and its `required_for`; for a `rig_type`
+it does not apply to, it is skipped and the waypoint is still asked through its
+other prompts. `source` says whether the prompt is part of the standard set
+(`rig-create`) or was added for this project (`project`, see
+[§ Questions outside this set](#questions-outside-this-set)).
+
+| id | At | Prompt | Required for |
+| --- | --- | --- | --- |
+| `rig_type` | framing | What is this: a character rig — standing idle, complex motion, or a monster with its own moves — or a full-screen scene — minimal motion (breath, blink) or motion tied to game logic? | all |
+| `reference` | framing | "Make it feel like this": a video, an existing demo, a game you have in mind? | all |
+| `x_layers` | `seethrough` | What must the layer split keep together or apart, and what resolution, head crop and seed should the passes use? | all |
+| `split_granularity` | `assemble` | Which parts are cut and how finely, in what draw order, with what authored new or kept under occluders, and which outfit sets? | all |
+| `split_emphasis` | `assemble` | Which parts matter most — more detail, or kept exactly as painted? | all |
+| `x_chains` | `bones` | Which parts get chains or stay rigid, where do the bones sit, and how long are the chains? | all |
+| `x_deform` | `meshes` | What must bend smoothly as one piece, what must stay rigid, and at what mesh density? | all |
+| `split_effects` | `constraints` | Which parts have special effects, how strongly and how far do they move, what gravity applies, and where must chains stop? | all |
+| `motions` | `motion` | Which moves? Name each one, how far and how fast it moves, the loop length, and whether it blinks. | all except `fullscreen.game_logic` |
+| `game_states` | `motion` | Which states and transitions, how far and how fast does each motion move, and how long is the loop? | `fullscreen.game_logic` |
+
+Values the agent normalises a guide into, per prompt (the sentence stays the
+answer):
+
+- `rig_type`: `character.idle`, `character.complex`, `character.monster`,
+  `fullscreen.minimal`, `fullscreen.game_logic`. Any other value is refused
+  (`RIG_TYPE_UNKNOWN`). A `fullscreen.*` project is a CG build: its rig
+  versions declare the three CG gates ([CHECKS.md § CG gates](CHECKS.md#cg-gates)).
+- `reference`: a list of `{ "ref", "take" }`, or `"none"`. A file the user hands
+  over goes to `inputs/references/` with its sha256 — it is an input, not
+  documentation.
+- `split_granularity`: `coarse` (only the parts a listed motion moves),
+  `standard` (limbs and garments at joints), `fine`, or a part list.
+- `split_emphasis`: a list of `{ "part", "want": "detail" | "untouched" | <text> }`,
+  or `"none"`.
+- `x_deform`: `{ "meshes": [<part>], "rigid": [<part>], "bust": … }`, where
+  `bust` is `"default"` (meshed with its own regions; also what no `bust` means)
+  or `{ "excluded": true, "reason": "<the user's reason>" }`. A blank reason is
+  refused (`BUST_EXCLUSION_UNREASONED`); [CRAFT.md M2](CRAFT.md#m2-the-bust-is-meshed-unless-there-is-a-reason).
+- `split_effects`: a list of `{ "part", "effect", "extent" }` — effect from
+  `physics`, `wind`, `jiggle`, `rigid`, `still`, or free text; extent as said
+  ("down to the knee").
+- `motions`: a list of `{ "motion", "note" }`.
+- `game_states`: `{ "states": [{ "name", "loop", "motions" }], "transitions": [{ "from", "to", "trigger" }] }`.
+
+The full label map — every part outlined on an image with one label — is the
+agent's reading of the `assemble` and `constraints` guides, shown back for the
+user to sign at the `assemble` checkpoint ([PROJECT.md § Split approval](PROJECT.md#split-approval-is-a-plan-checkpoint)).
+It is not asked.
+
+When the `assemble` guide is answered, tell the user, in plain words, that they
+will see the assembled parts before anything is rigged, and why: a part with
+nothing under its occluder, or a seam between two parts, is cheap to fix then
+and expensive after every rig build made on it
+([PROJECT.md § The parts are reviewed before rigging](PROJECT.md#the-parts-are-reviewed-before-rigging)).
+Their sign-off is the verdict they save in rig-review's parts view: it is
+recorded with the `parts-review` sign-off as evidence, and a part changed after
+it means they look again.
+
+## Post-production: asked after the rig is accepted
+
+2.5D depth and camera work are far from rigging: nothing in the rig is built or
+judged for them. They are not rig waypoints. Their guides are asked when
+`post-interview` opens, after the rig has passed `rig-review`, in the same
+answer shape and under the same rule: every post item the plan passes carries a
+sentence or a delegation. Asking them earlier would ask about a rig that may
+still change. "Accepted" means one rig version accepted at `rig-review` on its
+still frame; the post work runs on that version, which `rig-create handoff`
+names ([PROJECT.md § The hand-off to post-production](PROJECT.md#the-hand-off-to-post-production)).
+
+| id | Config sections | Prompt | What the agent decides there | Guide sentence, example |
+| --- | --- | --- | --- | --- |
+| `camera_depth` | `turn`, `depth` | Does the view move — a 2.5D head-turn with parallax, a push-in — or is it fixed? | whether depth is built, how far layers shift, the camera path | *"Fixed camera, no parallax."* · *"A slow head-turn to the left, as if she notices someone."* |
+| `post_timing` | — | How long, does it loop, and where are the beats? | clip length, loop point, timing of motions against each other | *"Ten seconds, looping; the gust comes once, in the middle."* |
+| `post_composition` | — | What is around the figure — background, framing, other figures, light? | framing, background plate, what else is on screen | *"Frame from the knees up; keep the lighthouse in shot."* |
+
+## Bars have reading commands
+
+Before putting a bar to the user for sign-off, the drafter names the command
+that produces **each** reading in it, the inputs it reads and the output that
+contains the reading. Check each command and flag against the installed CLI's
+`--help`. A missing command, or a command that only produces inputs for a
+missing measurement, is a **STOP before sign-off**. Write the command beside
+its clause in the plan's `bar` text. No new record field is needed.
+
+For example, for a compiled synthetic eight-second `idle` candidate in
+`derived/rig/`, these are two draft clauses, each with its reading command.
+The candidate and its named animation must exist before the draft is signed.
+
+- A warm idle loop has a frame series and bone geometry available for review.
+  Run `node_modules/.bin/rigc render --candidate derived/rig --animation idle --fps 24 --geometry --preroll 3 --out derived/warm`.
+  Read the recorded loop's frames and `geometry.json` in the reported set
+  directory. This provides the review inputs for one warm loop.
+- At 4 seconds the setup's named chain tips are visible for a pose review.
+  Run `node_modules/.bin/rigc inspect --candidate derived/rig --animation idle --at 4 --fps 24 --show art,bones --out derived/idle-at-4.png`.
+  Read the image and its companion legend's bone origins and tips. A pose
+  review is read by eye, with its verdict written beside the image.
+
+### Appearance belongs to a person's review
+
+An appearance item that needs the whole piece before it can be judged is not
+an intermediate step bar: perceived jiggle size, perceived motion and "looks
+right in the loop" belong in the plan as items for a later **person's review**.
+Route a rig step's item to `rig-review`, and a post step's to `direction-review`.
+Attach the step's measured readings as context; keep the step bar about those
+readings, rather than stopping mid-step for an appearance verdict.
+
+No agent can watch a loop: an image reader shows one frame. A call that needs
+a moving image ("seen at 1x in the loop") therefore goes to a person's review
+by design, never to an agent's step. Reading a frame series or measuring bone
+positions does not supply that person's appearance verdict. A numeric reading
+beside an appearance phrase does not turn the appearance call into a step bar.
+
+`interview --check` warns with `STEP_APPEARANCE_REVIEW` when a step bar contains
+one of the [documented phrases](CHECKS.md#appearance-bar-wording) without a
+numeric reading. This wording heuristic is a WARN, never a refusal: it can
+miss a paraphrase or flag a quotation, and cannot decide whether a motion
+looks right. Its absence is no approval of an appearance bar.
+
+### Warm loops and multi-period twins
+
+Today's `rigc render --preroll <n>` walks whole loops of a looping track before
+recording one loop, carrying physics across each wrap. The help says `n` is a
+whole number of 1 or more. It lists refusals for a walk past what a step list
+holds, a skeleton with no animation or no physics constraint, a set of 0
+frames, and `--environment` or `--parallax` beside it. `check` refuses that
+frame set as a reference. It supplies one warm recorded loop, with the preroll
+recorded in its metadata.
+
+To read an environment schedule or each successive boundary, make a
+**measurement twin** with one animation several source periods long, then
+render it once with `--geometry` and, if needed, `--environment schedule.json`.
+The schedule uses the twin's time range. This keeps physics running through
+internal period boundaries without asking `inspect --at` or an environment
+schedule to read beyond a source animation's duration. `inspect --at` still
+uses the animation's `--fps` frame grid.
+
+Keep the source unchanged. Record the source, twin, period, number of periods,
+frame rate, poser, grid and every track difference with the evidence. A blink
+keyed once is **not repeated** in later periods. The twin therefore cannot
+prove those periods have the source's repeated blink. Name any other one-shot
+track and its difference too. A multi-period render is measurement evidence
+only for the behaviour it actually reproduces.
+
+### Post bars on a looping idle
+
+Draft the loop-seam bar with these three clauses before seeking sign-off.
+
+1. Compare the whole-figure changed-pixel count across a boundary with physics
+   on against physics off **at that same boundary**. Breath can have its
+   fastest phase at the wrap, so its in-loop maximum excludes the very step
+   being judged.
+2. Name every background used for that count, for example white `#ffffff` and
+   sentinel magenta `#ff00ff`, with the same pixel grid, frame rate, boundary
+   times and counting definition for the on/off pair.
+3. Keep the physics reading for each named chain tip. Its boundary jump must
+   be no larger than that tip's own largest in-loop frame-to-frame step.
+   Compare distances in the same world units and at the same frame rate.
+
+Today's `rigc render --help` offers neither a physics-off pose command nor a
+whole-figure boundary pixel counter. `--environment` controls wind and gravity;
+its help provides no physics-off setting. The whole-figure comparison therefore
+has no reading command today. A bar that requires it is a **STOP before
+sign-off** until a tool supplies that reading. Rendering frames is only an
+input to counting them. Do not give the user an executable-looking command for
+a measurement tool that does not exist.
+
+The filled example adds only the chain-tip clause that can be read from
+`render --geometry`, with its reduction stated beside the command. It omits
+the unavailable whole-figure comparison. A project requiring all three clauses
+must resolve the missing reading before it can sign its post bar.
+
+## Provenance: every config section cites its guide
+
+Each section of the config the agent writes cites the guide sentence or the
+delegation it rests on. A section with no basis is a value the agent invented,
+which is what the interview exists to stop. The citation lives in `project.json`
+beside the version built from that config — `versions[].provenance` — not inside
+the config, because the config lives with the build in `derived/` and is not
+kept:
+
+```json
+"provenance": {
+  "config": { "path": "derived/v1/config.json", "sha256": "…" },
+  "sections": {
+    "assemble": [
+      { "entry": "plan[bust]", "guide": "assemble", "quote": "Cut the bust, not the waist." },
+      { "entry": "plan[*] order", "guide": "assemble", "delegated": true }
+    ],
+    "constraints": [
+      { "entry": "bust_physics", "guide": "constraints", "quote": "the bust jiggles" },
+      { "entry": "hem_*", "guide": "constraints", "quote": "The hem moves in the wind down to the knee" }
+    ]
+  }
+}
+```
+
+- `sections` is keyed by config section name; `entry` names the item in the
+  section (a part, bone, constraint or track, with `*` for a rule that covers
+  several); `guide` is the answer id it rests on (a waypoint, framing or post
+  item id); `quote` is the exact fragment of the stored sentence it follows;
+  `delegated: true` points at the delegation or `agent_decision` instead.
+- An entry with no citation is a decision nobody gave. It is either cited (by
+  finding the sentence, or asking for one) or removed before the version is
+  shown. `interview --check` refuses, by section name, a section with no
+  citation and a citation that does not resolve — a `guide` that is no answer,
+  a `quote` found in none of that answer's stored sentences, a `delegated: true`
+  on an answer the user did not hand over
+  ([CHECKS.md § Provenance](CHECKS.md#provenance)).
+- A bounce's `on_items` names the waypoint, and the provenance says which
+  sentence or delegation the bounced entry came from.
+
+## Polish feedback
+
+What the user says after seeing a version is stored verbatim in that version's
+`verdicts[]`, with `review` (`rig-review` or `direction-review`) and
+`on_items[]`. Each item names the answer it touches (a waypoint id, a framing id
+or a post item id) and `answer_was`:
+
+- `user` — the version did not follow a sentence the user gave: an agent error,
+  fixed in the next version.
+- `delegated` — the agent's own decision, on an item the user handed over: a
+  delegated cost, not an agent error. The user may now answer it themselves
+  (the new sentence replaces the delegation, which moves to `history[]`).
+- `not-asked` — the feedback names something no guide covered. It becomes a new
+  sentence in that waypoint's guide (`guide.additions[]`), quoted from the
+  feedback with its date and marked `"source": "polish feedback, v<N>"`, before
+  the next version is built.
+  It is never fixed silently, so the same feedback is not needed twice: the next
+  version's provenance must be able to cite it.
+
+The next version's `answers_feedback` lists the verdicts it answers (version
+number and verdict index), and its `changed` says what changed. Feedback on the
+rig given at the direction review goes back to the rig waypoints and makes a
+new rig version, which needs a new first review.
+
+### Recording a verdict
+
+```
+rig-create review <version> --verdict accepted|bounced|provisional --by <who>
+  --review rig-review|direction-review --verbatim "<words>" --source <where>
+  [--item <id>[:user|delegated|not-asked]]... [--quote <id>=<fragment>]
+  [--value <id>=<json>] [<project dir>] [--json]
+```
+
+Appends one verdict to `versions[v<N>].verdicts[]`, dated now, with the words
+stored exactly as given, and prints what it recorded. The review must be the
+one that judges that kind of version (`rig-review` a rig version,
+`direction-review` a post version); rig feedback given at the direction review
+is recorded on the post version shown and names the rig waypoints in its items.
+
+`answer_was` comes from the record wherever the record decides it, and a value
+the record contradicts is refused by name (`FEEDBACK_ITEM_CONTRADICTS`):
+
+| The item's current answer | `answer_was` when not named | Also allowed |
+| --- | --- | --- |
+| a delegation | `delegated` | — |
+| a user sentence | `user` | `not-asked` — the feedback names something the sentence did not cover, which only the reader can tell |
+| a user sentence that left part open (`agent_decision` beside it) | none: name it (`FEEDBACK_ITEM_AMBIGUOUS`) | `user`, `delegated`, `not-asked` |
+| not applicable, or no answer | refused (`FEEDBACK_ITEM_NOT_APPLICABLE`, `FEEDBACK_ITEM_SILENT`): ask the question first | — |
+
+A `not-asked` item is written in the same turn as a new sentence in that
+answer's `additions[]`: `by` and `at` of the verdict, `verbatim` = the
+`--quote` fragment (default: the whole feedback; it must be a fragment of it),
+`value` = `--value` (default `null`, to be normalised before the next version)
+and `source: "polish feedback, v<N>"`. The original sentence is not edited.
+Because the guide changed, the steps resting on it need the plan re-signed
+before they run again (`GUIDE_CHANGED_UNSIGNED`).
+
+An item id that is no framing, waypoint or post item is refused
+(`FEEDBACK_ITEM_UNKNOWN`); `review` does not create an `x_` prompt. A prompt
+is a question with its own text and `required_for`, asked of the user like the
+rest (§ Questions outside this set) — a verdict cannot write it — and an `x_`
+prompt's answer lives in its waypoint's guide, so the item the verdict names is
+that waypoint either way. Add the prompt, then name the waypoint.
+
+`review` never moves `stage` and writes no sign-off: a verdict says what a
+person thought of a version, and a stage moves only on a sign-off. An
+acceptance is recorded with this command and signed with its sign-off in the
+same turn ([PROJECT.md § Stages](PROJECT.md#stages)).
+
+`interview --check` then refuses a bounced version that a later version does
+not answer, an `answers_feedback` ref that points at no earlier verdict, and a
+`not-asked` item whose sentence is not in the guide; and it counts every bounce
+by cause, so a bounce on a delegated item reads as the user's chosen cost
+([CHECKS.md § Polish feedback](CHECKS.md#polish-feedback)).
+
+## Checking the interview
+
+```
+rig-create interview --check [<project dir>] [--json]
+```
+
+Before any rigging, and after any change to the record, the check reads
+`project.json` and refuses by name everything still open: a waypoint the plan
+passes with neither a guide nor a delegation, a delegated item the agent has not
+decided while the plan is signed, a plan step resting on no answer, a stage
+moved without its sign-off, a `@rig-c/parts` section the waypoint table does not
+place. It reports every refusal, not the first; exit `0` means the record is
+complete for its stage. The codes, what fires each and how to clear it are in
+[CHECKS.md](CHECKS.md).
+
+## Questions outside this set
+
+When a plan step needs a decision no prompt reaches, it is asked like the rest:
+add a prompt with an `x_` id (`x_wind_direction`) at the waypoint it belongs to,
+with `"source": "project"`, and record the user's sentence or delegation in that
+waypoint's guide. A step never rests on nothing. An `x_` prompt that comes up on
+a second project is a candidate for the standard set.
+
+## A user check names its size floor or stage
+
+When a user asks for a check such as "specks in layers mean re-run", ask for
+the smallest island area that counts (in pixels, at the named resolution), or
+which stage the check applies at: raw layers, merged layers, or assembled
+parts. Record the answer verbatim beside its normalized bar. A raw-layer
+speck rule without a floor can reject harmless small islands before merging;
+a check intended for assembled parts must say so. Do not invent the floor or
+silently apply the check at another stage. This is an interview question,
+not an automated check.
+
+## Recording an agent-found bounce
+
+A route rejected by the agent before a review belongs in the waypoint guide's
+`history[]`: `{ "at": "<date>", "bounce": { "cause": "agent", "failed":
+"<what failed>", "bar": "<the bar it failed>", "replacement": "<what replaced
+it>" } }`. Keep all three descriptions non-empty. This does not attribute a
+human verdict or alter the user's guide; a replacement requiring a changed
+user choice still goes back to that user. See
+[PROJECT.md](PROJECT.md#prepared-inputs-and-agent-found-bounces) for the
+prepared-input record and the per-entry compatibility date.
+
+## Name every skipped ask
+
+The canonical rig asks are `WAYPOINT_ASKS` in `src/waypoints.ts`, derived from
+W1–W6 above. Before finishing a waypoint, compare its guide with every applicable
+ask. Say each unanswered question in plain words. If the user delegates it,
+name that skipped question back as delegated, record its id in the guide's
+`asks` list with `state: "delegated"`, and write the decision beside the guide.
+Record answered asks with `state: "user"`. Never treat an answer to the crux as
+an answer to the remainder, or silently turn silence into delegation. The user
+still sees plain questions, not internal ids. `agent_decision.covers` remains
+free text; per-ask coverage is a separate list (PROJECT.md § Per-ask guide coverage).
+
+### Projection and ground at assemble
+
+Ask projection on every new assemble interview. Core keeps the part mask as
+its boundary. Visible follows the painting rim and can take backdrop pixels.
+Record the `projection` ask in the guide's `asks`, with `value.project` set to
+`core` or `visible` for a user answer. A delegated choice goes in
+`agent_decision.value.project`, with the existing reasoning and `covers` rules.
+
+Only visible asks `ground`. Plain preserves the existing rim projection.
+Busy limits backdrop pickup around thin parts. The author answers `plain` or
+`busy`, recorded as `value.ground` for a user answer or
+`agent_decision.value.ground` for a decided delegation. Keep the user's words
+and time with the guide. For example, a user guide can contain
+`"value": {"project": "visible", "ground": "busy"}` and
+`"asks": [{"id": "projection", "state": "user"}, {"id": "ground", "state": "user"}]`
+beside its other asks.
+
+The agent writes these answers as `assemble.project` and, for visible,
+`assemble.ground` in the project config. Create has no config-writing command.
+Parts' `ASSEMBLE_GROUND_PLAIN` WARN reports a plain-declared border that is not
+mostly one colour. A busy ground and a figure crossing the border can both
+cause it, so keep the warning for the author to judge. A measured interview
+proposal awaits parts#1023's public CLI reading.
+
+Guides predating `PROJECTION_ASKS_SINCE` (2026-11-06T00:00:00Z) retain the old
+contract and receive `GUIDE_ASKS_LEGACY` when a new applicable ask is absent.
+Core has no ground ask or ground refusal.
+
+
+## Rigging questions follow the approved parts
+
+In the explicit split order, the initial interview asks framing plus source
+and parts preparation only (`seethrough`, `assemble`). Bones, deformation,
+constraints and motion guides wait for `rigging-interview`, after the person
+has accepted the parts and mesh/chest structure. Null later guides are not
+missing early answers. The rigging plan is signed separately, against the
+hash-bound parts approval and structure snapshot. The person's earlier mesh
+and chest declarations are carried into that interview as approved context.
+
+The context is outside the answer union. This card keeps the existing
+user/delegated/not-applicable/inherited states without prescribing the later
+multiple-choice option set or correction-path schema. Existing sentence answers
+remain history; adoption does not re-date them or copy them as new answers.
+The source and rigging approval resolver is `approvalStageOfStep`; all delegation
+and plan checks read it. Source planning cannot require an unanswered later
+rigging guide. Hand pass means the correction loop of rig-review, not another
+signature whose meaning nobody defined.
+
+## Per-ask choices (interview@2)
+
+New projects explicitly adopt `interview.order.spec: "rig-create/stage-order@2"`,
+with `adopted_at` for the split order and `choices_adopted_at` for choices.
+`stage-order@1` keeps the split order and sentence answers; no date opts a
+project in. `interview-options --adopt [dir]` upgrades an existing @1 project,
+preserving its original adoption date and all answers. Earlier current answers
+warn `CHOICE_ANSWERS_LEGACY` until actually revised. `CHOICE_ANSWERS_SINCE`
+labels legacy warnings only. New answers from the explicit adoption require
+choices. Signoffs, verdicts, checkpoints and mechanical n/a remain unchanged.
+
+On a legacy-order project at `post-interview`, the same `interview-options --adopt`
+command adopts only post choices, recording `interview.post.choices_adopted_at`.
+It preserves the stage order, earlier answers and signatures. Only `camera_depth`,
+`post_timing` and `post_composition` can use this adoption; framing and waypoint
+choices still require explicit split-order adoption. Repeating adoption preserves
+its original date. Post picks and approval use the Review job and `--from-review`.
+
+Every newly asked framing question, waypoint and post item uses a retained
+option set. **Each applicable ask gets its own pick.** A waypoint is never a
+single bundle hiding several decisions. A framing/post item is a one-ask set
+whose ask id is its question id. A waypoint uses exactly its applicable
+`WAYPOINT_ASKS` in table order. Motion's game-state asks depend on rig_type;
+assemble includes ground if any offered projection is visible. Sets offering
+both core and visible must make the ground combinations explicit. Other and
+auto are always available, never preselected. Each ask has at least two value
+options and exactly one recommended value option with a nonempty reason.
+
+The agent writes a retained file, for example `evidence/interview/motion-001.json`:
+
+```json
+{
+  "schema": "rig-create/interview-options@1",
+  "id": "motion-001",
+  "question": "motion",
+  "stage": "rigging-interview",
+  "drafted_by": "agent",
+  "drafted_at": "2026-11-11T11:01:00Z",
+  "context": { "kind": "parts", "accepted_parts": {}, "structure": {}, "draft_id": "example", "parts": {} },
+  "asks": [
+    { "id": "tracks", "prompt": "Which moves?", "options": [
+      { "id": "breath", "kind": "value", "label": "Breath", "effect": "Only the torso breathes.", "fragment": [{"motion":"breath"}], "recommended": true, "reason": "Keeps the approved rigid accessories still." },
+      { "id": "still", "kind": "value", "label": "Still", "effect": "No track is added.", "fragment": [], "recommended": false },
+      { "id": "other", "kind": "other", "label": "Other", "effect": "Describe this decision.", "recommended": false },
+      { "id": "auto", "kind": "auto", "label": "Let the agent decide", "effect": "The agent records its decision before approval.", "recommended": false }
+    ] }
+  ],
+  "assembly": { "rule": "concat" },
+  "constraints": [{"asks":["tracks","blink"],"allowed":[["breath","yes"],["still","no"]]}]
+}
+```
+
+The sketch omits the remaining motion asks and context bindings; it is not a
+valid filled record. Use the complete writer-generated
+[consumer fixture](../templates/example-interview-choice.json) and raw
+[option snapshots](../templates/example-interview-options/) for executable data.
+The full file is the snapshot as shown. Identity is `(raw file sha, ask id,
+option id)`; ids are unique within each ask. A redraft uses a new set id and
+file; never overwrite an already bound file. Hash raw UTF-8 bytes, not a
+canonical reserialization. Label/effect/note/Other strings retain all newlines,
+Unicode and trailing whitespace without normalizing or translating them.
+
+Context is a projection of existing recorded identities, never a new producer
+measurement:
+
+- `inputs`: `{kind:"inputs", files:[{path,sha256}]}` binds the declared source
+  inputs available before parts. It never claims future parts were shown.
+- `parts`: `{kind:"parts", accepted_parts:{signoff_at,evidence}, structure,
+  draft_id, parts:{path,files:[{path,sha256}]}}` binds the live parts-review,
+  retained structure snapshot and full recorded draft input inventory,
+  including parts.json and every image. Inventory file paths are project-relative.
+- `rig`: `{kind:"rig", rig_version, signoff_at, config:{path,sha256}, parts}`
+  binds the accepted rig, its rig-review signoff, retained provenance config
+  and the parts inventory from its build-time approval. Post asks wait for it.
+- `scene`: `{kind:"scene", scene:{n,path,sha256}}` replaces the rig kind for a
+  post set while the latest scene is current on the latest live accepted rig
+  ([PROJECT.md, "Scene option context"](PROJECT.md#scene-option-context)).
+
+Draft time and selection time must both see that context live. New selection
+or decision after changed files/approval/structure or a relevant reopen refuses
+`INTERVIEW_OPTIONS_STALE`; redraft first. An old answer is checked against the
+context live when it was given, so later approvals do not rewrite history.
+Later appended source inputs do not change which immutable inputs an earlier
+person saw; a fresh writer checks the current complete source identity.
+Ignored draft bytes may be unavailable on a historical clone, retaining the
+existing `DRAFT_RIG_UNAVAILABLE` warning. Fresh registration/selection/decision
+requires the actual viewed bytes present. Retained option, structure, source
+and config bytes remain mandatory. Paths stay within the project, with literal
+segments, no symlink traversal, regular files and no ignored option storage.
+
+`assembly.rule` is a closed enum:
+
+- `single` is allowed only for a one-ask set; its contribution is the value as
+  is, including scalar `rig_type` and reference strings.
+- `concat` concatenates array fragments in asks order.
+- `merge` shallow-merges object fragments in asks order, refusing a key collision.
+
+No semantic inference, coercion or silent adjustment occurs. Optional
+`constraints` name distinct asks and allowed tuples of their value-option ids.
+An Other or auto fragment on a constrained ask resolves to the **unique** value
+option with a deep-equal fragment. Zero or several matches, or a tuple not in
+allowed, refuses `INTERVIEW_CHOICE_CONSTRAINT`: redraft the set to offer that
+option. Thus a novel Other value on a constrained ask requires a redraft.
+Unresolved auto/Other constraints wait until their contributions land.
+
+The writer extends the existing answer union with:
+
+```json
+{
+  "choice": {
+    "spec": "rig-create/interview-choice@1",
+    "set": {"path":"evidence/interview/motion-001.json","sha256":"<raw sha>"},
+    "picks": [{"ask":"tracks","option_id":"breath","shown":{"label":"Breath","effect":"Only the torso breathes."}}],
+    "note": "Keep it subtle."
+  }
+}
+```
+
+Picks are exactly one per ask, in set order. `shown` is copied exactly from the
+retained option. Other adds `other_text` and optionally `other_fragment`.
+`verbatim` is deterministic: each pick contributes label LF effect (LF
+other_text for Other); join picks with a blank line, then LF note if present,
+even for an explicitly empty note. Existing provenance quotes can match those
+exact words. Selecting a recommendation is a human answer, not delegation.
+`value` is the assembly of user contributions; `asks[].state` is `user`,
+`delegated`, or `other-pending`. If any Other lacks a fragment, **value is null**,
+not an inferred value. Approval of that question refuses
+`INTERVIEW_CHOICE_REQUIRED` until a new set offers a tailored value option
+quoting the pending words with source/date and the person explicitly picks it.
+The old answer moves intact to history. The CLI can accept a person's explicit
+Other fragment directly.
+
+All-auto is `delegated:true` with no user value; mixed picks are
+`delegated:false` and retain their user value. Before a decision exists,
+`agent_decision:{covers:"<auto ids joined with comma>"}` names the pending
+remainder. The agent then writes `by`, `at`, `reasoning`, exact `covers`,
+`fragments:[{ask,fragment}]` for exactly the auto asks in order, and `value`
+as the assembly of those fragments. User fragments and agent fragments are
+also checked together for merge collisions and constraints. Existing approval
+and execution deadlines for undecided delegation remain in force.
+
+In adopted projects prior words become a **tailored value option**, whose label
+quotes the earlier answer verbatim and whose effect names its original source
+and date. The human still explicitly picks. It is recommended only with an
+agent reason. Never turn old words into a fabricated click or silently select
+a recommendation. Unadopted projects keep the earlier derived-answer rules.
+
+| Answer field | Writer and meaning |
+| --- | --- |
+| `refusal_baseline` | The answer writer records the pre-pending refusal multiset as `{code, path, count}[]` when auto or unresolved Other is selected. Entries have nonempty strings and positive integer counts. The record is retained when decided and in history. Each new pending answer gets a fresh baseline. Older pending answers without it keep the strict current-state comparison. |
+
+### Public writers and review ingestion
+
+```
+rig-create interview-options <question> [dir] --file evidence/interview/<id>.json
+rig-create interview-answer <question> [dir] --set <path> --sha256 <shown sha> --by <human> --pick <ask>=<option>...
+rig-create interview-answer <question> [dir] --from-review <actId>
+rig-create interview-decision <question> [dir] --by <agent> --reasoning-file <file> --fragment <ask>=<JSON file>...
+```
+
+Registration reads an already retained authored file and records its binding
+as `pending_options`; it never rewrites the option file. Answer flags include
+repeatable `--other <ask>=<text file>` and optional
+`--other-fragment <ask>=<JSON file>`, plus `--note-file` and `--source`.
+The decision writer supplies exactly the auto fragments. Changing an approved
+answer first requires the relevant stage to be reopened. Collection members
+write local overrides; inherited source answers remain historical evidence,
+never fabricated member selections.
+
+Review exposes finalized `interview-picks` acts through the standing job API.
+The writer fetches the act using its private sibling capability, binds the
+question and displayed option-set ID and SHA-256 to the pending set, and retains
+the semantic act bytes and SHA-256 in `review_acts`. Selections carry item and
+option IDs, optional Other text and the exact note. The writer derives shown
+text, assembly, coverage and verbatim from the retained options and checks
+constraints. The act's timestamp and reviewer role determine the answer's
+`at` and `by`. Identical replay returns the retained answer, including from
+history. Conflicting bytes under the same act identity refuse.
+
+`--from-review` is exclusive with pick, Other, fragment and note flags.
+Optional set/hash/by flags must agree with the act. A stale displayed set
+refuses `INTERVIEW_OPTIONS_STALE`. Malformed acts refuse `REVIEW_ACT_INVALID`.
+The removed `--picks-file` flag refuses `INTERVIEW_PICKS_FILE_REMOVED` and points
+to `--from-review`. Earlier file-route answers remain valid records. Synthetic
+generators retain historical wire fixtures while constructing answers through
+the public flag writer. Flags use the writer's clock.
+
+Writers use the check's choice validators before writing. A partial interview
+may keep unrelated existing refusals, but a writer cannot introduce a new
+one. When ending the same pending auto transition, the decision writer permits each code/path up to the larger count in `refusal_baseline` and the current record, while refusing findings absent from both and refusing any excess count. A refused project write leaves its bytes identical; retained authored
+files are never rewritten. Regenerate all consumer examples with
+`scripts/generate-example-choices.ts <fresh scratch> [--check]`; the active full
+example also reproduces through `generate-example-order.ts` and stays 0/0.
+
+## Optional intake choice in parts
+
+Recording a preflight with `intake-report` offers a new optional ask,
+`intake_suitability`, in **parts**, bound to the recorded report's raw sha256.
+It is not asked in the source interview: it needs a full run and cannot save
+the first full run. The record lives at `intake.choice`, outside required
+framing/waypoints and every completeness or approval deadline. A HIT with
+no resolved choice for that report raises `INTAKE_CHOICE_PENDING`; a no-risk
+report raises no choice warning. Neither is an approval refusal.
+
+Use `interview-options intake_suitability <dir> --file <retained-set>` to
+retain a `rig-create/interview-options@1` set with `stage: "parts"`,
+`question: "intake_suitability"`, one ask of that id and `assembly.rule:
+"single"`. Its context is `{ kind: "intake", report_id: <id>, report:
+<the entry's complete report binding> }`. The ordinary per-ask rules apply:
+two value options (for example scalar `proceed` and `choose-another`), one
+explicitly explained recommendation that is never preselected, Other and
+auto. Choose with `interview-answer`; only explicit auto may be completed
+with `interview-decision`. Other text and note are retained verbatim through
+the existing writer, together with the exact labels/effects shown. Do not
+substitute an agent summary for the person's words.
+
+Pending auto, decided auto, Other with a fragment and Other pending stay
+explicit and optional. A delegated decision is not a person's proceed pick.
+Choosing another records intent; it never changes inputs automatically or
+screens a corpus reproduction away. A new report needs a fresh shown set
+and explicit choice; the old choice and set remain history. Choices are
+local, never inherited from a collection. They may be recorded only in
+parts, but later signatures do not make a pending optional choice invalid.