feat: imported-course mode for needs-analysis, assessment-design, course-builder

[savvides]

Summary

Three skills (needs-analysis, assessment-design, course-builder) assumed net-new course design and broke down for courses imported from a cartridge. The TMC-430 test report flagged this as Cluster #C — three issues with the same root cause. This PR adds an early branch on `import_metadata.source` to each affected skill.

Skill	Without this PR (TMC-430 case)	With this PR
needs-analysis	Asks "is training the right intervention?" — meaningless for a credit-bearing course already in the registrar's catalog.	Skips the gate, records `justified=true, confidence=10, rationale="design audit only"`, runs a design-fit check instead.
assessment-design	Designs new rubrics from scratch even though the cartridge has 6 existing ones.	New Mode 3 — Audit Existing: reads existing rubrics, classifies criteria on Bloom's, flags alignment gaps. Doesn't propose new assessments unless asked.
course-builder	Would regenerate syllabus, modules, rubrics that already exist. Half-measure at line 211 acknowledged `import_metadata` but didn't act on it.	New gap-fill mode: generates ONLY the artifacts upstream skills flagged as missing. Skips Steps 2–7 entirely.

All three announce the chosen mode at the start of the conversation, so the user can override the auto-detection if it's wrong.

Mode-detection contract

Both conditions must be true to enter the audit-existing/gap-fill branch:

1. `import_metadata.source` is one of `cartridge`, `scorm`, `canvas-api` (whitelisted formats only).
2. The skill-relevant section is non-empty:
   • needs-analysis: `import_metadata.items_imported.modules > 0`
   • assessment-design: `assessments.items` non-empty OR `course_content.assessments` non-empty
   • course-builder: `course_content.modules` non-empty

Default is design-new mode — the conservative choice. Borderline states (e.g., `import_metadata.source` set but zero modules imported) ask the user via `AskUserQuestion` rather than guessing.

Schema additions (additive, no version bump)

New optional fields in `templates/manifest-schema.md`:

• `needs_analysis.mode` — `"design-new"` or `"audit-existing"`
• `assessments.mode` — `"Mode 1"`, `"Mode 2"`, or `"Mode 3"`
• `assessments.audit_notes[]` — Mode 3 outputs (`{target_id, action, description, reason}`)
• `course_content.mode` — `"build-new"` or `"gap-fill"`
• `course_content.recommended_generation_targets[]` — gap-fill outputs (`{description, source, status, output_path}`)

Existing manifests stay readable; downstream skills that don't read the new fields ignore them.

Other changes

• assessment-design adopts `bin/idstack-manifest-merge` for its manifest writes (landed in v2.1.0.0). Replaces the inline-write pattern's misleading "include the COMPLETE schema structure" instruction with two scoped merge calls (one per top-level section the skill writes).

Test plan

• Smoke tests: 153/153 passing (was 150). 3 new assertions: each mode-aware skill template references `import_metadata.source` (drift guard against accidental removal).
• Generated freshness: `bin/idstack-gen-skills --dry-run` — 0 stale.
• Manifest schema review: new optional fields are additive; `bin/idstack-migrate` LATEST_VERSION stays at 1.4 (no migration needed).
• Out-of-band on TMC-430 (after merge): the regression target. Expected:
  • `needs-analysis` reports "Mode: audit-existing (cartridge import from Canvas)" instead of "Mode: design-new"; skips the training-justification gate.
  • `assessment-design` reports "Mode 3: 6 existing rubrics audited; 2 alignment gaps found" instead of designing new rubrics.
  • `course-builder` reports "Mode: gap-fill; 4 generation targets queued" (matching the targets the v2.1.0.0 audit identified) instead of attempting to generate a syllabus.

Risks (covered)

• R1 — Mode-detection edge cases: explicit decision tree, both whitelist + non-empty check; ambiguous states ask the user.
• R2 — Design-new branch silently regresses: audit-existing branch is strictly additive before existing workflow; no edits to existing Step bodies.
• R3 — Item shape inconsistency between modes: both modes produce items in the canonical shape (same `id, type, title, weight, ilos_measured, rubric_present, alignment_status` fields).
• R4 — Test coverage: smoke-test grep is a weak signal but acceptable; behavioral verification is the TMC-430 re-run.
• R5 — User confusion when manual `import_metadata` is partial: mode is announced as the first sentence, user can override.

Out of scope (deferred — see TMC-430 test report)

• Per-skill scope improvements (fix: address Gemini code review findings on PRs #6, #7, #8 #10, fix(install): clean up pre-v2.0.1.0 dispatcher, regression-test it, add doctor #16, docs(landing): mark beta + refresh timeline through v2.4.0 #18, v2.4.0.2 — future-major version-pattern fix (Gemini, PR #21) #22, chore(docs): remove personal gstack skill references from CLAUDE.md #25): coverage cross-walk, scoring methodology, parallel-dispatch payload size, accessibility tooling, evidence currency check. Each substantial enough for its own PR.
• Minor docs/UX polish (fix: remove silent auto-update, replace with notify-only #4, refactor: fix systemic schema drift surfaced by TMC-430 test report #8). Will batch into a future cleanup PR.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces 'Imported-course mode' for the needs-analysis, assessment-design, and course-builder skills, enabling specialized workflows for imported content such as auditing existing assessments and gap-fill generation. The manifest schema is updated with mode-tracking fields and audit note arrays, and the assessment-design skill is updated to use the idstack-manifest-merge tool for atomic writes. Feedback identifies misleading section descriptions in the assessment-design skill and corrects field name inconsistencies in the course-builder skill to ensure alignment with the canonical schema.

[gemini-code-assist]

The description of the sections written by this skill is misleading. It identifies learning_objectives.alignment_matrix.ilo_to_assessment as a "section", but the idstack-manifest-merge tool operates on top-level sections (like learning_objectives). Furthermore, in Mode 3, the skill also writes to learning_objectives.alignment_matrix.gaps. It is clearer and more accurate to just name the top-level sections.

Suggested change

	and atomically updates the top-level `updated` timestamp. **This skill writes two
	sections** — `assessments` and `learning_objectives.alignment_matrix.ilo_to_assessment` —
	so call the merge tool twice (once per top-level section).
	and atomically updates the top-level `updated` timestamp. **This skill writes two
	sections** — `assessments` and `learning_objectives` —
	so call the merge tool twice (once per top-level section).

[gemini-code-assist]

The description of the sections written by this skill is misleading. It identifies learning_objectives.alignment_matrix.ilo_to_assessment as a "section", but the idstack-manifest-merge tool operates on top-level sections (like learning_objectives). Furthermore, in Mode 3, the skill also writes to learning_objectives.alignment_matrix.gaps. It is clearer and more accurate to just name the top-level sections.

and atomically updates the top-level `updated` timestamp. **This skill writes two
sections** — `assessments` and `learning_objectives` —
so call the merge tool twice (once per top-level section).

[gemini-code-assist]

The instructions refer to course_builder_notes for the mode and recommended_generation_targets fields, but the manifest schema (and the generated_files field in the same sentence) uses course_content. This inconsistency will lead to the LLM writing data to a non-existent section.

Suggested change

	4. **Record outputs.** Update `course_content.generated_files` (additive) and `course_builder_notes.recommended_generation_targets` (the list, with status `generated | deferred | declined`).
	When done, write the manifest via `bin/idstack-manifest-merge` (see Write Manifest below) and skip directly to the final summary.
	Save the chosen mode under `course_builder_notes.mode` (`"build-new"` or `"gap-fill"`).
	4. **Record outputs.** Update `course_content.generated_files` (additive) and `course_content.recommended_generation_targets` (the list, with status `generated | deferred | declined`).
	When done, write the manifest via `bin/idstack-manifest-merge` (see Write Manifest below) and skip directly to the final summary.
	Save the chosen mode under `course_content.mode` (`"build-new"` or `"gap-fill"`).

[gemini-code-assist]

The instructions refer to course_builder_notes for the mode and recommended_generation_targets fields, but the manifest schema (and the generated_files field in the same sentence) uses course_content. This inconsistency will lead to the LLM writing data to a non-existent section.

4. **Record outputs.** Update `course_content.generated_files` (additive) and `course_content.recommended_generation_targets` (the list, with status `generated | deferred | declined`).

When done, write the manifest via `bin/idstack-manifest-merge` (see Write Manifest below) and skip directly to the final summary.

Save the chosen mode under `course_content.mode` (`"build-new"` or `"gap-fill"`).