[Carr1005]

Fixes the flaw the 2026-10-19 adoption run found in both agnostic specs: an option that removes a part (conversation-only memory, no grader) left other rows naming that part. The gate asked questions that contradicted the earlier answer, and the build agent had to guess. The process failure that let #58 and #59 merge with this flaw is in #60.

What changes

• §5.5 "No cross-row restating", rewritten in place, from values to parts:
  • Rows after the removing row never list or count the parts an earlier row's options remove. They name them through the owning row ("each store row D1's answer keeps").
  • Rows before it (asked first, so they cannot point ahead) list the parts as the Defaults build them, and the removing option's §7 line names each such cell.
  • Only the Trade-off cell and the Default's branch note may name the course's parts.
  • A row that the option skips by its condition is exempt.
• §7 "Tests by option": the change column becomes "Rules, criteria and Ledger cells that change (and how)".
• §6.0 gate step 5 (copied word for word into every spec), two additions:
  • leave out every part of the spec that exists only for a rule a chosen option retires (commands, flags, settings, schema fields, fixtures, permission items);
  • adapt a changed criterion whose new form is not fully stated, and list it under Adapted.
• §14 line and audit (k) match. (k) now checks moot rows below the option's row only, which fixes a Guide: option ids, "Applies only when" conditions, §7 "Tests by option" table; eval reads resolved-decisions.md #58 wording bug.
• eval-spec-vs-build: a new Adapted section, judged against the chosen options.
• Also on this branch: the skip-example placeholder fix (cherry-picked from Guide §6.0 step 3: the skip example uses placeholders, not real-looking row ids #61; merge Guide §6.0 step 3: the skip example uses placeholders, not real-looking row ids #61 first). The overlay is re-pinned; its 4 quoted passages are unchanged (grep -F).

Tested (Opus 5.5), with the #60 purpose check

For each part-removing option, a probe runs the gate on paper, then classes every place in the spec that contradicts the choice.

Guide	Spec	Option	Gate	Covered	Left out	Adapted	Guess
1632117 (before)	both	conversation-only, no grader	store row "all seven stores"; rubric and grader rows	—	—	—	(adoption audits)
c2e1546 (cross-row rule only)	A	conversation-only	0	most	—	—	7
c2e1546	B	no grader / command line / no learning loop	0	most	—	—	6 / 3 / 3
5caba53 (this PR)	A	conversation-only	8*	14	25	5	4
5caba53 (this PR)	B	no grader	1	17	39	4	2

* 5 are §1 "Not Included" items shown in the scope question, which is asked before D1 and so is true at the time it is asked. 3 are a generator slip in A's store row (the rule forbids it); the adoption repair handles it.

Residuals (accepted; logged on #48)

• §1 items shown in a row above are not Ledger cells, so the §7 line does not name them. This only matters if a person restores one and then picks the removing option.
• Low-risk guesses: whether to keep an always-null field or unused schema definitions; a flag's default once its only nonzero use is gone; a permission item tied to a skipped row.
• A known generator slip class: a row after the removing row still describing removed parts (A's store row). The audit catches it.

🤖 Generated with Claude Code