Repository navigation
vnext: integrate learning-system stack through R7 - #64
Merged
Merged
Conversation
…, planner (#42) Greenfield learning core per docs/vnext/capability-model-v0.md — no UI, no legacy coupling. The engine answers: what can the learner verifiably do, on what evidence, did it survive delay, did it transfer, what next. - capabilities.js: 12-capability graph across 5 modalities with an acyclic prerequisite check; Vietnamese risk probes attached per capability. - evidence.js: append-only EvidenceEvent schema — 11 event types, typed context (mission/practiced/transfer/prompt family), observed vs self-report, per-support flags. answerBearing() is the single rule deciding what counts as help. - projection.js: pure event-log → milestones → state derivation. Locks the doctrine: aided/self-reported success never earns INDEPENDENT; RETAINED needs a ≥24h unaided re-success; TRANSFERRED needs a novel promptFamily (replayed practiced prompts don't count); FLUENT = transfer across ≥2 families; modality-mismatched events earn nothing; replay is deterministic, ids dedupe on resync. - risk-priors.js: 12 Vietnamese population priors — diagnostic probes only, learnerStateEffect is a constant, not a write path. - planner.js: deterministic next action — resume → due retrieval → remediation (only for taught work; a baseline fail routes to teaching, not retry) → transfer → unaided attempt → probe/expose gated on INDEPENDENT prerequisites. - tests/vnext.test.mjs: 9 groups pin every rule above plus the full vertical slice (baseline fail → input → retrieval → supported → feedback → retry → INDEPENDENT → RETAINED → TRANSFERRED → FLUENT). Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- only attempt-bearing event types may advance state; an exposure, support_use, or feedback record carrying outcome:success is contact, not performance - canonical replay order is (occurredAt, id) — arrival order can no longer change the projection or the plan for equal timestamps - projectLearnerState/planNext require learnerId; foreign-learner events are dropped before projection so a mixed log can never cross-contaminate - a failed/partial due check routes to remediation instead of rescheduling delayed_retrieval forever - FLUENT requires ≥2 novel transfer families AND ≥2 transfer successes measurably faster than the independent baseline (FLUENCY_LATENCY_RATIO); without latency data the ceiling is TRANSFERRED — two changed-context wins alone only prove breadth - rehearsedPromptFamilies tracks every practiced-context family (aided or not); a support-rehearsed prompt can never be re-sold as a novel transfer context Regression tests pin each invariant in tests/vnext.test.mjs §9. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Two evidence-honesty fixes from re-review: - FLUENT is no longer auto-promotable. The latency-ratio rule was an uncalibrated invented threshold that rewarded answering faster — the doctrine requires hesitation + intelligibility + successful turns + repairs + stability. FLUENT stays in the enum/milestones as reserved but nothing sets it; TRANSFERRED is the v0 ceiling until a dedicated fluency-evidence contract exists. - conditions.supportAllowed was dead data: an attempt could earn INDEPENDENT while using support the capability never permits. isIndependent now requires every used support kind to be within the capability's declared conditions. 'repeat_once' is only satisfied with provenance — support.repeatCount === 1; a bare repeat flag cannot distinguish once from many and fails the condition. Support fields gain repeatCount for this. Regression tests pin both: repeat on a no-support capability caps at SUPPORTED; a repeat_once capability accepts exactly one recorded replay; even fast transfers stay TRANSFERRED. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
vnext: headless capability engine — graph, evidence, projection, planner (#42)
) Teaching item ≠ transfer item ≠ assessment item — the engine now enforces it structurally instead of trusting callers: - contracts.js: MissionContract/TaskContract + validators. Missions check capability existence, prerequisite closure, task ownership, an eliciting path per target, and freshness collisions (a practiced family can never masquerade as fresh assessment/transfer). Tasks enforce purpose semantics: transfer needs declared changed dimensions + fresh_transfer family; assessment needs fresh_assessment family, zero allowed support, and no mid-attempt answer reveal. - bind.js: bindAttempt/bindObservation are the only way learner-facing code mints EvidenceEvents. capabilityId/taskId/revision/modality/ missionId/promptFamily/practicedOrTransfer/purpose/evaluation all derive from the contract — caller-supplied copies throw. eventType is constrained by purpose (an interaction task cannot emit checkpoint). - effectiveAllowedSupport = capability ∩ task — a task may narrow, never broaden; violations are recorded honestly then demoted. - projection: sticky support per attemptId (union across canonical order — a hint revealed mid-attempt permanently taints that attempt), evaluator-authority gate (only deterministic/human can award INDEPENDENT; self_report/asr/ai_llm cap at SUPPORTED), and the support-conditions check now uses the bound effective policy. - fixtures.js: two complete headless missions — "meet a new person" (diagnostic → input → retrieval → supported → feedback → unaided → delayed → transfer → assessment) and "order a drink". - validateMissionContent enforces declared language + injected budgets; no universal thresholds hard-coded. tests/vnext-contracts.test.mjs pins all 15 required invariants. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- evaluation.authority is contract-derived; a caller reporting a
different authority is a binding error, never an upgrade path
- fresh_assessment binds its own 'assessment' context kind; only
'transfer' context can feed TRANSFERRED, and assessed families are
still rehearsed (cannot be re-sold as novel transfer)
- projection only credits contract-bound attempts: binding.purpose must
exist and agree with eventType, so raw makeEvent() successes cap at
SUPPORTED
- validateMission scopes evidence-path/freshness checks to taskIds and
flags tasks that claim the mission without being declared
- validateMissionContent drops the capability-language whitelist:
mission language must be declared in introduced/assumedKnown, and
assumedKnown is structured {chunks,vocabulary,constructions}
- attempt-boundary support history keys on taskId::attemptId, so a
reused attemptId cannot leak support across tasks
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
…ontracts - projectLearnerState now takes the registered task list and re-derives every semantic field from it: taskId+revision, capabilityId, modality, purpose↔eventType, promptFamily, context kind, evaluation authority+contractId, and the recomputed effective support policy. A stamped binding is data, not proof — a raw makeEvent() carrying a forged binding can never verify and caps at SUPPORTED. - Eliciting purposes require a real evaluation.contractId at construction (makeTask) AND at verification, so authority: deterministic + contractId: null can never mint independent evidence. - planner passes the task registry through to the projection. - tests: contract-test registry + core suite registers synthetic contracts mirroring binder output; new §13 pins forged-binding and missing-contractId cases. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- The registry is keyed taskId@taskRevision so a v2 contract cannot overwrite v1 — historical evidence keeps verifying under the contract that produced it and replay stays stable. Duplicate id@revision registrations are rejected outright. - verifyEventTask now requires the resolved task to pass validateTask() first — a hand-rolled invalid contract (e.g. purpose 'transfer' with zero changedDimensions) mints nothing, however closely an event's fields happen to match it. - tests: revision coexistence, duplicate rejection, invalid-transfer regressions; the core suite's synthetic registry now meets the same purpose-specific requirements validateTask enforces. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
vnext: Mission/Task/Transfer/Assessment contracts + evidence binder (#45)
nextMissionTask() maps planner intents to concrete registered tasks inside mission.taskIds — declared baseline diagnostics run first, then expose/resume/retry/independent_attempt/delayed_retrieval/transfer map to their exact purposes; a missing task yields blocked, never a silent substitute. Assessment is gated on TRANSFERRED so it samples achieved ability instead of standing in for it. runMissionTrace() drives a scripted learner through selector → binder → append-only evidence → projection, emitting an auditable per-step trace (taskId, purpose, reason, before/after state). Mission A gains the missing interact.ask_name phases (baseline diagnostic, input, retrieval, remediation) so the full path diagnostic fail → input → retrieval → supported → feedback → retry → INDEPENDENT → 24h delayed → RETAINED → transfer → TRANSFERRED → fresh assessment is selectable end-to-end and FLUENT stays unreachable.
Review round 2 orchestration fixes: - Task identity is taskId@taskRevision end-to-end: selection returns taskRevision, mission ids resolve to the highest registered revision, trace lookup matches exact revision, duplicate id@revision throws. - Consumption is verified-only: an event marks its task run only if it passes verifyEventTask against the registry — forged or mis-revisioned events can no longer hide an unrun baseline or input task. - Mission integrity fails closed: a declared taskId absent from the registry or resolving to an invalid contract blocks selection instead of being dropped by filter(Boolean). - expose/resume fallback is bounded to input/notice → retrieval / production / interaction — it can never drift into remediation, delayed_retrieval, transfer or assessment. - EVENT_TYPES_FOR_PURPOSE gains input/notice → observation-only types so exposure tasks verify under the same purpose↔eventType contract.
…dence (#50) Feasibility pilot for the evidence engine, not an efficacy study: five scripted learners through the meet-a-new-person mission across five deterministic sessions (day0, +25h, +50h, +75h, +100h). - runPilotLearner: bounded runMissionTrace per session sharing one append-only log; harness stamps id/learnerId/occurredAt (learner isolation is the harness's job) and counts per-run service calls (ctx.call) so scripted outcomes are replay-deterministic. - evaluateClaim: recomputes learnedByFlashday from PRIMITIVE verified evidence — milestones are outputs, never claim inputs. Baseline-pass → acquisitionSource PREEXISTING (never claimable); unresolved- contradiction rule replaces the brittle "no fail after independent"; retained24h/retained72h horizons reported separately; milestone-vs- primitive mismatches surface as integrityFailures. - mission-runner: a failed assessment is re-probed after remediation (latest checkpoint outcome ≠ success re-serves while TRANSFERRED holds) instead of dead-ending at first non-pass. - tests/vnext-pilot: 10 checks — fast/dependent/forgetful/able/ oscillator archetypes, claimRate over eligible claims, zero invalid promotions, forged-revision immunity, replay + isolation. Research basis: issue #49 round 2 (feasibility criteria, unresolved- contradiction claims, separate retention horizons, no vanity metrics). Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Per R3 research (#49): engine defines what happened; policy defines how much evidence is enough. Pedagogical thresholds are data. - src/vnext/policy.js: LEARNING_POLICY_V1 frozen + registry, resolvePolicy fails closed on unknown versions/malformed objects, makePolicy for declared test variants. Thresholds: retention minLagMs, independent successfulUnaidedRetrievals/minDistinctSessions/ minSpacingGapMs, remediation minConsecutiveFailures, claim require* flags + blockOnUnresolvedContradiction. - projection: {policy} opt feeds retention lag; new engine fact consecutiveFailures (trailing fail/partial count); return carries policyVersion. - planner: remediation routes on consecutiveFailures >= policy.remediation.minConsecutiveFailures (v1=1, same behavior). - mission-runner/pilot-harness: policy plumbs through selection, trace and claim; evaluateClaim reads every threshold from policy and stamps policyVersion on the claim. - tests/vnext-policy: registry fail-closed, frozen objects, same- events-different-lag retention, 2-strike remediation, version stamping, lenient-vs-strict claim divergence. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Per R3 research (#49): persistence lands before any UI — if a lost-ack retry can double an attempt, the UI could mint INDEPENDENT by accident. - src/vnext/persist.js: users/{uid}/vnext_events/{eventId} adapter on the cloud-compat fs surface. appendVnextEvents is per-event transactional — absent → create, identical content → dedupe, different content under the same id → hard conflict (evidence identity is never last-write-wins). loadVnextEvents returns the canonical (occurredAt, id) replay order. toEventDoc keeps both occurred_at (client-observed, drives lag) and recorded_at (server arrival) plus schema_version/policy_version/mission_run_id. - firestore.rules: users/{uid}/vnext_events — create+read only; update AND delete denied even for the owner (immutable evidence, stricter than lesson_events); learner_id must equal the path uid — evidence for another learner is a forgery. - tests/vnext-persist: fake-fs round-trip provenance, append → reload → replay parity (projection AND claim identical), exactly-once retry, id-conflict throw, out-of-order canonicalization, isolation. - tests/firestore-vnext-emulator: real rules — owner CRUD, update/ delete denied for owner, cross-owner + anon denied, learner_id forgery denied, malformed docs denied, e2e persist → reload → replay → identical claim. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
nextMissionTask → bindAttempt → append → projection, surfaced through four screen types (intro / learn-input / reusable task / checkpoint). - ui-session.js: headless session controller — selector-driven screens, deterministic attemptIds, support_use events + immutable snapshot on the attempt, prompt→committed→feedback phases, run persistence across reloads without a new missionRunId. - evaluators.js: reusable deterministic contracts (eval.required_functions.v1, eval.choice.correct.v1) scoring real response shapes via mission-checks; fixtures now bind these contracts instead of hollow per-task ids. - planner/selector: uncovered intents exclude per (capability, kind) via skipIntentFor — the cap stays on the surface for prerequisites and its other intents still route; only verified evidence consumes a task; novelty-free purposes may re-elicit, diagnostics stay single-sample. - vnext/ page: Vietnamese-first copy describing observed evidence — no engine labels, no mastery claims; pre-commit support is recorded and stamps the attempt. - tests: 13 ui-session checks + 2 browser groups covering baseline first, support before commit, stamped attempt, reload resume. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
) R5 (issue #49) called for shipping the curriculum schema before scaling content: a capability DAG, mission roles that separate claim-bearing targets from carriers, and auditable prompt families. - Capability ids namespaced by CEFR activity type: reception.listen.*, reception.read.*, production.speak.*, production.write.*, interaction.* — one migration, before any new missions depend on the names. - Missions gain carrierCapabilities: rehearsed for retention/context but never owed transfer credit. Mission A now claims only the two R5-shaped targets (say_own_name + ask_name); say_own_name gets its missing delayed + held-out transfer tasks and checkpoint sample. - contextSignature on every task: the family's auditable identity (cueTopology, setting, register, channel, interlocutorRole, ...). Canonical family ids pf.<cap>.<cue>.<setting>.<register>.<channel>.vN re-derive it, so an id that contradicts its signature fails. - curriculum-checks.js is the authoring gate: DAG validity, per-target evidenceability (practiced + delayed + fresh_transfer + assessment sample, <=3 targets), family coherence (same id -> same signature), fake-novelty detection, and transfer deltas verified as real signature deltas against EVERY practiced family. - Prerequisite closure now covers the whole declared surface (targets + carriers + support): an undeclared prereq deadlocks its dependents. - Delayed re-checks deliberately reuse the rehearsed family — a delayed task on a novel family would measure transfer, not retention. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
…roductions (#57) Second authoring-gate pass from the GPT R6 design review: - Canonical prompt-family ids gain a signature hash: pf.<cap>.<cue>.<setting>.<register>.<channel>.<sigHash8>.vN. The readable prefix stays human-inspectable; the fnv1a fingerprint covers the WHOLE contextSignature, so two families differing only in a non-id field (interlocutorRole, relationship, responseTopology, lexicalDomain) can no longer alias behind a .vN bump. Fixtures now derive ids via canonicalFamilyId(capId, sig) — id≡sig by construction. - Missions are role-aware at introduction: targets owe a mandatory baseline probe; carriers rehearse opportunistically (input + eliciting practice, never a diagnostic, never fresh_transfer, never an assessment of their own); supports are demand-driven — undeclared rather than dead surface. Mission A drops the four carrier diagnostics and ask_repeat entirely; mission B's offer check becomes post-input retrieval practice. - Surface cap: a mission declares at most 6 capabilities across all roles — beyond that the diagnostic/exposure phase explodes. - interaction.order_drink broadens to interaction.request_item (same communicative function; drink vs food/goods is a context difference, not a capability difference). Verify green: unit suites (curriculum, contracts, slice, ui-session, pilot, policy, persist), browser 25 groups, Firestore emulator. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Five new A1 missions with full evidence packages (baseline diagnostic → input → retrieval → interaction → delayed → transfer → checkpoint): meet_at_a_time, complete_small_order, buy_small_item, find_a_place, talk_about_self_family — 12 new capabilities in the CEFR namespaces. Runner: uncovered intents on non-target caps (carriers, prerequisite gates) are cross-mission backlog, not mission integrity failures — stale prior-mission evidence no longer freezes later missions. Curriculum gate: prerequisiteCapabilities are read-only DAG gates and no longer count against the 6-capability active surface. UI: all seven missions registered; mission page resolves the whole curriculum task registry so carried prerequisite evidence verifies (routing stays scoped to mission.taskIds). Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
…ms (#59) Fake prerequisite edges removed: GPT R7's bar is that a hard edge is only legitimate when no valid assessment exists that a learner could pass downstream while failing upstream — sequencing belongs to pedagogy, not the DAG. All capability prerequisites are now empty; curriculum order moves to pedagogy.recommendedAfterMissions on the mission object. - capabilities.js: all prereqs emptied; providesFunctions added on support caps (identify_spoken_number, understand_simple_choice, identify_basic_direction_term) as the substrate the future support-demand mechanism consumes. - fixtures.js: mission declarations use targets/carriers only — paper-support declarations dropped per R7 (lazy support had no demand path); transfer signatures normalized to exactly the 3 designed dimensions (cueTopology, setting, interlocutorRole) so novelty claims carry ≤3 confounds; M6 gains a practiced landmark-retrieval task so follow_short_direction gets rehearsal evidence inside its own mission. - tests: vertical-slice rewrite codifies the new semantic (baseline fail routes to teaching the capability itself, never a prereq walk); M3 slice asserts target-first ordering, visible-but-non-fatal backlog, and lazy-support non-routing.
Repository-local orchestration for long autonomous missions: QUEUED →
RUNNING → VERIFYING → DONE/FAILED/BLOCKED, one non-terminal mission at
a time, state in plain artifacts under missions/<id>/.
- scripts/swe.mjs: status/start/checkpoint/verify/finish/resume.
swe:start refuses on dirty tree or a second active mission.
Verification commands are allowlisted shapes (npm run <script>,
node tests|scripts/<file>) — a mission file can never smuggle shell.
swe:finish DONE requires the last verify green on current HEAD.
Checkpoint ingest stamps git state and enforces all 10 sections.
swe:resume rebuilds context from mission + latest checkpoint + git.
- missions/TEMPLATE.md + README.md: authoring contract + lifecycle doc.
- tests/swe-factory.test.mjs: 21 checks in a throwaway git repo —
dirty-tree refusal, one-active-mission, checkpoint validation,
verify gating incl. HEAD-moved re-verify, red-verify cannot DONE,
blocked→resume, allowlist rejection, terminal immutability.
- npm scripts swe:* wired; test wired into npm test.
Fixed en route: git status --porcelain output must not be trimmed —
the first line's leading space is the unstaged-worktree status column;
trimming corrupted path extraction ('tests/x' → 'ests/x').
…t first line-end (#62)
001-factory-smoke: start → checkpoint → verify (PASS) → finish DONE, REPORT.md generated. Zero product changes; files-changed-vs-start = 0. 002-resume-sim: start → checkpoint --blocked → swe:resume printed the full context packet and unblocked → verify (PASS) → DONE. Proves a fresh session reconstructs state from mission + checkpoint + git alone.
…ontract (#62) Two false-success paths closed (external review findings): - Finding 1 (confirmed): DONE could land while tracked product changes sat uncommitted — verify ran on the dirty tree and finish only compared SHAs, so the report's startSha..HEAD diff hid the work. swe:finish DONE now requires a clean tree per the same policy as swe:start; --result failed stays unconditional (abandon never traps work). - Finding 2 (confirmed): mission.md was re-read per command, so a running mission could rewrite objective/criteria/verification — committed or not — and still finish DONE. swe:start now pins the mission.md sha256 into state.json; verify/finish/resume re-hash and fail closed on drift. Restoring original bytes restores operation. There is deliberately no V1 amendment workflow. - README: allowlist wording corrected — it constrains mission-specified verification commands only, not the agent's own tool access (reliability guard, not a sandbox). Tests: 10 new checks — dirty-tree refusal (tracked + unsafe untracked), commit-then-reverify-then-done, uncommitted/committed mission.md drift detection at all three guarded commands, restore-definitions-recovers, hash persisted at start.
4 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
vNext stack integration — single candidate to main
This PR carries current main + the complete vNext development stack + the SWE Work Factory on one verified branch. Integration mission ran through the Work Factory itself —
missions/003-vnext-stack-integration/holds the mission definition, checkpoint, verification log, and REPORT.md (mission DONE).Ancestry proven (git, not PR descriptions)
git merge-base --is-ancestoron every stack tip → all ancestors of55b2271(PR vnext: author missions 3–7 + cross-mission evidence registry (#59) #60 head):devin/vnext-slice-001(vNext: headless meet-a-person slice + mission runner (#47) #48)f3f5d89✓devin/vnext-pilot-001(vnext: deterministic pilot harness — cohort evidence claims (#50) #51)8c5a5c5✓devin/vnext-policy-001(vnext: versioned learning policy + append-only persistence (#52, #54) #53)ede0d9c✓devin/vnext-ui-001(vnext: minimal honest mission UI over the real contract stack (#55) #56)3f550d3✓devin/vnext-curric-001(vnext: capability DAG + canonical prompt families + curriculum gate (#57) #58)27f98da✓git log 55b2271..origin/rebuild/vnext→ empty — pre-stack core (vnext: headless capability engine — graph, evidence, projection, planner (#42) #43, vnext: Mission/Task/Transfer/Assessment contracts + evidence binder (#45) #46) fully contained.origin/main(0e44ad2) is the stack's merge-base — main ⊆ stack; main has not moved since the stack diverged.One inaccuracy found & handled: PR #63 (SWE factory) was still open, not merged on main as the mission brief assumed. Merging
devin/swe-factory-001integrates it — its ancestry already contains current main. #63 can still merge cleanly to main whenever convenient (or be superseded by this PR).Conflict resolution (1 conflict total)
package.jsonscripts — semantic union, not ours/theirs:test: full vNext suite chain +swe-factory.test.mjsappended.test:firestore: kept the vNext two-file form (firestore-emulator+firestore-vnext-emulator— superset of the factory side).swe:*commands kept.AGENTS.mdauto-merged (disjoint sections).No R7 resurrection, no new features
supportCapabilitiesstay out of missions;providesFunctionsretained as dormant substrate for vnext: demand-driven support routing (missingFunctions → providesFunctions → SUPPORT_DEMAND) #61 (deferred).prerequisitesreintroduced;pedagogy.recommendedAfterMissionskeeps curriculum order.Verification on final HEAD (9399726 + mission-artifact commit)
npm run verify:full— all green, one recorded run:Stacked PRs #48/#51/#53/#56/#58/#60 remain open as review provenance until this lands.
Generated with Devin