Open Point Detect Runner

---
name: open-point-detect-runner
description: Performs full-lens, read-only detection of unresolved questions for the current slice.
---

# open-point-detect-runner

## Goal

Expose unresolved questions that matter to the current slice as one
candidate batch.
`$CTX.guide.cognitive_frame` (D1) and `$CTX.guide.intent_anchor` (D2)
jointly guide Detect as a whole.

## Script Macros

Entry points. Contract in `--help`.

| Macro | Command |
|---|---|
| `$OPEN_POINT_CTL` | `python3 "$SKILL_ROOT/compose/scripts/inductive/open-point/open_point_control.py" --out-dir "$INDUCTIVE_OUT_DIR"` |

## Inputs

Parent invoke arguments only.

- `--out-dir` is required. Host root is process cwd.

## Detection context

Slice materials from the two control commands.

1. `$CTX` is `$OPEN_POINT_CTL detect-context` stdout, the roster:
   `opens_snapshot`, `pending_lenses`, and `guide`.
2. `$CTX.guide` is this stage's Domain `cognitive_frame` (D1) and
   `intent_anchor` (D2).
3. `pending_lenses` names the lenses on this pass, not a scan order.
4. `$OPEN_POINT_CTL detect-lens-context --lens <key>` is one lens
   packet: `kw_criteria`, `lens_registry`, and `facts_snapshot`.

## Cognition

Packet terms carry the meanings defined in these units.

| Term | Unit |
|---|---|
| `facts_snapshot`, fact | `../../references/cognition/fact.md` |
| `lens_registry`, lens | `../../references/cognition/lens.md` |
| `kw_criteria` | `../../references/cognition/kw-ruler.md` |
| `$CTX.guide` (D1, D2) | `../../references/cognition/profile/domain.md` |
| unknown, candidate Open | `../../references/cognition/producer/induce.md` |

One Detect pass composes them per lens:

```text
candidates(lens) = Detect(lens; kw rows, facts)
                   − SettledBy(facts)
                   − RepresentedBy(opens_snapshot)
```

## Detection dimensions

Rulers for judgment. Question type lives in `references/detect-means.md`.

1. Completeness obligation is the `kw_criteria` rows in the lens packet.
2. Probe every row in the packet, not only the coarsest silent one.
   List every point the facts cannot state, each tagged with its row
   in `kw`. A row yielding several points yields several candidates.
3. `gap_kw` names the coarsest row among those candidates.

## Boundaries

Invariants for this pass.

- Read-only. Control calls: `detect-context` once; `detect-lens-context`
  once per `pending_lenses` key. A failed fetch fails the pass. Run no
  control command beyond these two.
- Evidence closure: the two control stdout payloads, this SKILL,
  `references/detect-means.md`, and the five Cognition units are the
  whole evidence for one Detect pass. Read nothing outside the closure.
  An unclear term is judged from the closure, not looked up.
- Do not produce options, leanings, or solutions.

## Return

Hand-back shape. `means` is stamped per `references/detect-means.md`.

- One object: one verdict per `pending_lenses` key.
- `gap_kw` is null exactly when `candidates` is empty.
- `candidates[].kw` is the row the point belongs to; `gap_kw` equals
  the smallest `kw`. `kw` is Detect metadata, not an Open field.

```json
{
  "verdicts": [
    {"lens": "CTX", "gap_kw": 1, "candidates": [
      {
        "question": "What is the current rollback path?",
        "basis": "KW1 is silent on failure recovery",
        "blocking": true,
        "kw": 1,
        "means": "probe"
      }
    ]},
    {"lens": "GO", "gap_kw": null, "candidates": []}
  ]
}
```