vnext: author missions 3–7 (R6 spec)

[Thunderkill016]

Goal

Author the remaining five A1 missions per the GPT R6-validated table
(research log in #49), on top of the authoring gate in #58.

Mission table (R6)

Mission	Targets	Carriers	Supports
M3 meet_at_a_time	reception.listen.understand_clock_time, production.speak.state_clock_time	greet, interaction.ask_name	reception.listen.identify_spoken_number
M4 complete_small_order	interaction.request_item, interaction.answer_simple_choice	greet, thank	reception.listen.understand_simple_choice
M5 buy_small_item	interaction.ask_price, reception.listen.understand_spoken_price	interaction.request_item, thank	reception.listen.identify_spoken_number
M6 find_a_place	interaction.ask_location, reception.listen.follow_short_direction	greet, thank	reception.listen.identify_basic_direction_term
M7 talk_about_self_family	production.speak.state_basic_self_detail, production.speak.describe_family_member_basic	production.speak.say_own_name	lexical only

Every target needs a full evidence package: baseline diagnostic → input →
retrieval → production → delayed → fresh_transfer (explicit
changedDimensions on signature fields) → fresh assessment. Transfer
signatures per target are specified concretely in the R6 log.

New capabilities to add

reception.listen.identify_spoken_number, reception.listen.understand_clock_time,
production.speak.state_clock_time, reception.listen.understand_simple_choice,
interaction.answer_simple_choice, interaction.ask_price,
reception.listen.understand_spoken_price, interaction.ask_location,
reception.listen.follow_short_direction, reception.listen.identify_basic_direction_term,
production.speak.state_basic_self_detail, production.speak.describe_family_member_basic,
interaction.thank (carrier). Numbers/direction-terms stay support
substrate — no hard DAG edges (R6).

Acceptance

• checkCurriculum green on all 7 missions.
• Slice/pilot coverage proves at least one new mission end-to-end.
• Evaluator matchers + VN copy for every new requiredFunction.
• npm run verify, test:browser, test:firestore green.

[Thunderkill016]

Progress — PR #60 open (b592945, stacked on #58)

Authored all 5 missions per the R6 table. Highlights beyond raw authoring — three semantics fixes the gate/traces surfaced:

Missions (each target: diagnostic → input → retrieval → interaction guided+unaided → delayed → transfer w/ changed-context signature → fresh checkpoint):

• meet_at_a_time — clock-time hear+say; greet/ask_name carriers; spoken-number support
• complete_small_order — request_item + answer_simple_choice (reception split stays a lazy support)
• buy_small_item — ask_price + understand_spoken_price
• find_a_place — ask_location + follow_short_direction
• talk_about_self_family — self detail → family detail (cap-edge)

Semantics fixes found while authoring:

1. Runner: intents on non-target caps (carriers/prereqs) that the mission can't serve are now idle + skippedIntents backlog — previously any stale prior-mission evidence going due (24h) would blocked-freeze every later mission. Target intents still fail closed.
2. Gate: prerequisiteCapabilities excluded from the ≤6 active surface (read-only DAG gates, not bandwidth consumers).
3. UI registry: mission page now resolves the whole-curriculum task registry so M1-bound evidence verifies inside M3+ — was a genuine production deadlock.
4. One honest cap-edge added (state_clock_time ← understand_clock_time); numbers/direction-terms stayed support substrate, never edges (R6).
5. validateMissionContent language closure: every task's required language now ⊆ introduced∪assumedKnown within budget (12/12/4).

Verification: curriculum gate 0 problems on all 7; npm run verify (typecheck 120 files); browser 25 groups; Firestore emulator PASS; M4–M7 smoke-traced to checkpoint; M3 cross-mission slice test added (seeded M1 evidence → prereq gates → carriers rehearse → lazy support never routed → terminal idle with visible backlog).

GPT/R7 review consult in flight — verdicts go here before merge.

[Thunderkill016]

R7 consult resolved — both blocking issues fixed in 55b2271 on PR #60.

R7 verdict (from the GPT thread):

1. Fake prerequisite edges. Hard-edge legitimacy bar: an edge is valid only when no valid assessment exists that a learner could pass downstream while failing upstream — sequencing is pedagogy, not DAG. By that bar every authored edge was lesson-order. → All capability.prerequisites emptied (incl. the understand_clock_time → state_clock_time edge added in wave 1); curriculum order moved to pedagogy.recommendedAfterMissions. DAG machinery stays dormant for real dependencies.
2. Support-on-paper. Lazy supportCapabilities with no demand path = dead declaration (REJECT CURRENT). → Dropped from all mission declarations; support caps keep providesFunctions metadata as substrate. The real mechanism (missingFunctions → providesFunctions → SUPPORT_DEMAND → probe task) is follow-up issue vnext: demand-driven support routing (missingFunctions → providesFunctions → SUPPORT_DEMAND) #61.

Adjustments applied:

• Transfer signatures normalized to exactly the 3 designed dims (cueTopology, setting, interlocutorRole); channel/register/relationship constant vs practiced → ≤3 confounds per transfer claim. Family ids recomputed via canonicalFamilyId.
• M6 gained a practiced landmark-retrieval task so follow_short_direction rehearses inside its own mission.
• vnext.test.mjs vertical slice rewritten: baseline fail routes to teaching the capability itself (no prereq walk); earliest-due ordering now proven over two independent evidence timelines.
• vnext-slice.test.mjs M3 updated: target-first ordering, terminal idle + visible backlog, lazy-support non-routing asserted.

Gates after the wave: npm run verify (typecheck 120 files + all suites + build) ✓ · browser 25 groups ✓ · Firestore emulator ✓ · all 8 vNext suites ✓ · curriculum gate 0 problems over 7 missions.