docs(plan-integration): T5.1 — design §5 reviewed against the code; three amendments

Item 5 (D-F) was designed on 2026-09-16 before its RED. Read against
learning/decision.py at 7208eb6, three of its sentences do not fit the
tree; each is amended under §5 with the source it rests on, so T5.2's RED
is written against a design the code can carry.

1. "recovered / gentle review -> low" names a row _struggle_candidates never
   emits: it selects only confidence in ("struggling", "learning") or
   last_teachback_score < 14. The low-demand class is the `learning` row;
   fresh `struggling` (last_seen <= 14 days) is high, old `struggling` or a
   weak-teach-back-only row is medium. Demand is derived once in the
   collector and carried in candidate metadata.

2. "listed in energy_deferred" cannot hold a repair: DeferredMilestone has a
   mandatory milestone_index and all three renderers (cli/_now.py,
   learning/recap.py, today-panel.js deferredNotes) print
   `milestone {index + 1} "{title}"`. A deferred repair gets its own frozen
   DeferredRepair in a new additive key, energy_deferred_repairs, and each
   renderer gains one line for it.

3. "opens the existing body-double session route (web/routes/body_double.py)"
   names the read-only focus reader. The session door is `studyloop study
   "<topic>" --mode co-study` on the CLI and the Body Double view's session
   start on the Web; _evidence_command has no conversation branch and would
   fall through to `studyloop progress … -c learning` (a write), so the
   body-double candidate sets its evidence_command explicitly.

tasks.md T5.1 ticked with the receipt. Full suite on this branch: 30 failed
/ 5120 passed / 14 errors, the 44 failed+errored ids byte-identical to the
item-4 control's committed environmental set (run - control = empty).

Author: NetDevAutomate

openspec/changes/plan-integration-followons/design.md

@@ -238,6 +238,36 @@ among what MCP revises and the row names every schema property.
 - **Rubric row 3b** (owner scores): scenario 3's fixture at low energy now yields the deferred repair named in
   `energy_deferred` and a body-double primary (or the due recall if one exists).
 
+**T5.1 review against the code (2026-09-18, tree `7208eb67`) — three amendments, each from reading
+`learning/decision.py`, not the text above:**
+
+1. **Demand classes are the struggle collector's classes.** `_struggle_candidates` emits a row only when
+   `confidence in ("struggling", "learning")` or `last_teachback_score < 14`; "recovered / gentle review" is not a
+   row it produces. So: `struggling` with `last_seen` ≤ 14 days → `high`; `struggling` older than 14 days, or any
+   row whose only signal is a weak teach-back → `medium`; `learning` → `low`. Demand is derived once, in the
+   collector, and carried in the candidate's `metadata` beside `confidence` so the scorer and the renderers read
+   one value. Required capability `high → 6`, `medium → 4`, `low → 0` stands (the `low` class is what "repair is
+   cheaper than encoding" was always about).
+2. **`energy_deferred` is milestone-shaped and cannot carry a repair as it is.** `DeferredMilestone` has a
+   mandatory `milestone_index`, and all three renderers (`cli/_now.py`, `learning/recap.py`,
+   `today-panel.js::deferredNotes`) print `milestone {index + 1} "{title}" needs energy {floor}/10`. A deferred
+   repair gets its own frozen `DeferredRepair` (`plan_id`/`plan_title` when plan-related, else `None`, `concept`,
+   `topic`, `confidence`, `energy_demand`, `required_capability`, `energy_capability`, `reason` naming the
+   struggle), carried in a **new additive key `energy_deferred_repairs`** — not folded into `energy_deferred`,
+   whose consumers would print "milestone None". Same "readable off the top" rule as the closing review's
+   evidence lines: each renderer gains one line per deferred repair.
+3. **The body-double door is a session start, not `web/routes/body_double.py`.** That route is the read-only focus
+   reader (`GET /api/body-double/focus`). The session door is `studyloop study "<topic>" --mode co-study` on the
+   CLI and a session start from the Body Double view (origin `body-double`) on the Web. `_evidence_command` has
+   no branch for a `conversation` candidate and would fall through to `studyloop progress … -c learning`, which is
+   a write, not a door — so the body-double candidate carries `evidence_command = 'studyloop study "<plan title>"
+   --mode co-study'` set explicitly, and `_evidence_command` is not asked to guess. `source="body_double"`,
+   `action_type="conversation"`, base score below `MILESTONE_BASE_SCORE` (48) so any real candidate outranks it.
+
+Rule 3's *deferral* of repair is the change; rule 3's *eligibility* of plan-related due recall is untouched. The
+no-plan golden stays byte-identical because a body-double candidate requires an active plan and the golden world
+has none; `INTERLEAVE_RATIOS["low"]` unchanged.
+
 ## 6. Verification
 
 `scripts/verify/plan_integration.py` gains registered checks for: the two architect grants (the ten names in

openspec/changes/plan-integration-followons/tasks.md

@@ -192,8 +192,13 @@ writer, through the existing gate, closes that. Kept out of item 3 so item 3's f
 
 ## Item 5 — energy demand + body-doubling floor (D-F) · own round
 
-- [ ] **T5.1** Design §5 reviewed against the code (`_struggle_candidates`, `_score_candidates`, rule 3) — amend if
-      the code contradicts it.
+- [x] **T5.1** Design §5 reviewed against the code (`_struggle_candidates`, `_score_candidates`, rule 3) — amend if
+      the code contradicts it. (2026-09-18: three amendments recorded under §5 — the demand classes are the struggle
+      collector's own (`struggling` fresh/old, weak teach-back, `learning`; no "recovered" row exists); a deferred
+      repair needs its own `DeferredRepair` in a new additive `energy_deferred_repairs` key because `DeferredMilestone`
+      and its three renderers are milestone-shaped; the body-double door is `studyloop study … --mode co-study` /
+      the Body Double view's session start, not the read-only `body_double.py` focus route, so the candidate sets
+      its `evidence_command` explicitly. T5.2's RED names hold; a sixth test pins the new key's rendering.)
 - [ ] **T5.2** RED `tests/test_now_plan_guidance.py`: `test_live_struggle_repair_defers_at_low_energy_like_new_work`,
       `test_recovered_repair_stays_eligible_at_low_energy`, `test_body_double_candidate_is_synthesised_when_nothing_plan_related_fits`,
       `test_body_double_is_a_proposal_not_a_filter`, `test_body_double_never_appears_without_an_active_plan`,