fix(k6b): remediacion de integridad del verifier y Assurance Graph (v2.51.0)

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "ospec-workflow",
   "description": "Spec-Driven Development workflow for VS Code Agent Customization with OpenSpec, strict TDD, phase agents, skills, hooks, and verification contracts.",
-  "version": "2.50.0",
+  "version": "2.51.0",
   "author": {
     "name": "Manuel Michael Retamozo García"
   },

.plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "ospec-workflow",
   "description": "Spec-Driven Development workflow for VS Code Agent Customization with OpenSpec, strict TDD, phase agents, skills, hooks, and verification contracts.",
-  "version": "2.50.0",
+  "version": "2.51.0",
   "author": {
     "name": "Manuel Michael Retamozo García"
   },

CHANGELOG.md

@@ -5,6 +5,17 @@ All notable changes to this project are documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.51.0] - 2026-08-27
+
+### Added
+- **Remediación de integridad K6b (`k6b-verification-integrity-remediation`)**:
+  - PASS exige evidencia admisible por cada obligación MUST del Execution Graph; `strategy satisfied ≠ Execution Graph satisfied`.
+  - Contrato aditivo `assessment/v1` (`ospec://schemas/kernel/assessment/v1`): binding persistible EvidenceId ↔ role ↔ obligation_id ↔ node_id ↔ policy, sin mutar `evidence/v2` ni K1 v1.
+  - Provenance fuerte derivada del canal del harness (`input.collector` / `input.collectors[]`); `collector` en el sobre del worker falla cerrado (`UNTRUSTED_COLLECTOR`). Desacuerdo claim↔canal simétrico; weak+allowlist no escala a clase fuerte.
+  - `graph_id` fingerprinta inputs canónicos; proyección fail-closed (`GRAPH_PROJECTION_FAILED` / `GRAPH_DIVERGENCE`); `rejectForbidden` por kind/namespace.
+  - ADRs `docs/adr/adr-20260827-007` a `010`. Specs `independent-verification`, `assurance-graph` y `kernel-contract-schemas`.
+  - K6b queda `done`; K6c pasa a `next-eligible`. Ciclo SDD completo (ruta standard, high-risk, size:exception, 4R successor approved). Verify: 2754 pass, 0 fail. Archivado en `openspec/changes/archive/2026-08-27-k6b-verification-integrity-remediation/`.
+
 ## [2.50.0] - 2026-08-27
 
 ### Added

docs/adr/adr-20260827-007-additive-assessment-family-id.md

@@ -0,0 +1,23 @@
+# ADR-001: Additive assessment family `$id`
+
+- Status: proposed
+- Change: k6b-verification-integrity-remediation
+- Date: 2026-08-27
+
+## Context
+
+Assumption `sdd-propose-001` left the persistable binding `$id` to design. Intent forbids in-place mutation of `evidence/v2` and K1 v1. Assessment identity must include `role` and `obligation_id` without collapsing physical `EvidenceId`.
+
+## Decision
+
+Publish family `assessment` at `$id` `ospec://schemas/kernel/assessment/v1`, `kind: "assessment/v1"`, path `schemas/kernel/assessment/v1.schema.json`, `schema_version: 1`. Register additively in `manifest.json` and `contract-claims.json`. Keep `evidence/v2`, `verification/v2`, and K1 v1 bytes/pins frozen.
+
+## Alternatives
+
+- Mutate `evidence/v2` to carry role/obligation: rejected; observation identity would absorb evaluation identity.
+- Evolve `verification/v3` with embedded bindings: rejected; verdict unique-sort of `evidence_ids` still hides distinct roles.
+- `$id` `ospec://schemas/kernel/assessment-binding/v1`: rejected; breaks the `{noun}/vN` kernel `$id` pattern.
+
+## Consequences
+
+Consumers pin a third K6b-era family without migrating evidence. Rollback deletes the additive family. Four roles over one observation become four `assessment_id` values by construction.

docs/adr/adr-20260827-008-must-walk-after-strategy-assessment-identity.md

@@ -0,0 +1,23 @@
+# ADR-002: MUST walk after strategy; assessment identity ≠ EvidenceId
+
+- Status: proposed
+- Change: k6b-verification-integrity-remediation
+- Date: 2026-08-27
+
+## Context
+
+`evaluateStrategy()` only checks role minimums. PASS was therefore possible without Obligation Manifest coverage. `obligation_ids` on classified evidence are ephemeral, so `satisfies` edges cannot be replayed from `evidence/v2`.
+
+## Decision
+
+Keep `evaluateStrategy` as the role-shape gate. After it succeeds, walk every non-deferred `must` obligation, emit `assessment/v1` records, and require those assessments for PASS. `computeEvidenceId` stays observation-only. `assessment_id` fingerprints `evidence_id`, `role`, `obligation_id`, `node_id`, `candidate_id`, and `policy_snapshot_id`.
+
+## Alternatives
+
+- Fold MUST coverage into `evaluateStrategy`: rejected; mixes strategy policy with graph obligations and reason codes.
+- Walk obligations before strategy: rejected; delays cheap role failures.
+- Treat unique-sort of `verification.evidence_ids` as binding identity: rejected; one EvidenceId used as four roles would collapse.
+
+## Consequences
+
+Reason codes split cleanly (`MISSING_STRATEGY_MINIMUM` vs `UNFULFILLED_MUST` / `UNKNOWN_OBLIGATION_ID` / `WRONG_IMPLEMENTING_NODE`). Projector `satisfies` edges consume assessments only. Existing tests that omit obligation bindings will fail closed until updated.

docs/adr/adr-20260827-009-collector-transport-provenance-allowlist.md

@@ -0,0 +1,23 @@
+# ADR-003: Strong provenance from collector/transport allowlist
+
+- Status: proposed
+- Change: k6b-verification-integrity-remediation
+- Date: 2026-08-27
+
+## Context
+
+`normalizeEvidence` currently copies `raw.provenance` into `evidence/v2`. A worker can claim `runtime-observed`. PKI is out of scope. Payload digest identifies bytes, not origin.
+
+## Decision
+
+Derive strong classes (`runtime-observed`, `host-attested`, `tool-produced`) from harness-supplied `collector.id` + `transport` via a fail-closed allowlist (`node-test`/`npm-test`/`node:test` → runtime-observed; `tool-execution` → tool-produced; `host-adapter` → host-attested). Worker/absent/unknown collectors cannot produce a strong class. Store the derived class on `evidence/v2`; never copy collector metadata onto that record.
+
+## Alternatives
+
+- Trust the payload provenance string: rejected; this is the verified defect.
+- Require signatures or PKI: rejected; out of scope for this change.
+- Treat payload digest as origin: rejected; digest is content identity only.
+
+## Consequences
+
+Callers that claim a strong class must pass allowlisted collector metadata (tests included). Mis-mapped collector ids fail closed rather than silently weakening provenance.

docs/adr/adr-20260827-010-canonical-graph-id-fail-closed-projection.md

@@ -0,0 +1,23 @@
+# ADR-004: Canonical graph_id fingerprint and fail-closed projection
+
+- Status: proposed
+- Change: k6b-verification-integrity-remediation
+- Date: 2026-08-27
+
+## Context
+
+`graph_id` currently hashes only `candidate_id` plus canonical nodes/edges. Contract or policy changes do not diverge. `verifyCandidate` returns `ok: true` when projection fails, omitting the graph. `rejectForbidden` substring-matches `id`, so `REQ-add-authorization-header` is rejected.
+
+## Decision
+
+Fingerprint `contract_digest`, `policy_snapshot_id`, Execution Graph digest, canonical OpenSpec input, `candidate_id`, and canonical nodes/edges in `graph_id`. Persist those digests as optional `canonical_inputs` on `assurance-graph/v1`. `verifyCandidate` requires a successful projection before `ok: true`. Match forbidden subjects by `kind`/`namespace`, never by `id` substring.
+
+## Alternatives
+
+- Keep the current `graph_id` preimage: rejected; replay cannot detect contract/policy drift.
+- Leave projection optional on the facade: rejected; PASS without a graph is the integrity hole.
+- Continue substring markers on `id`: rejected; legitimate requirement ids containing `authorization` fail.
+
+## Consequences
+
+Existing runtime `graph_id` values change. Replay becomes possible from persistable outputs. Facade tests that ignored projector failure must assert `GRAPH_PROJECTION_FAILED` / `GRAPH_DIVERGENCE`.

docs/architecture/harness-evolution.md

@@ -1,8 +1,8 @@
 # Arquitectura objetivo — harness gobernado por kernel, grafo y evidencia
 
 > **Autoridad:** fuente conceptual y estratégica del harness (responsabilidades y límites).
-> **Corte documental:** v2.50.0, 2026-08-27 (estado alineado al roadmap; la dirección conceptual no cambia).
-> **Estado verificado:** O3, O4+O5/O4.1, O4.2, O6A, O2B, **K1**, **K2**, **K2.1**, **K2a**, **K3**, **`k3-readiness-remediation`**, **K4a**, **K5**, **K6a**, **K4b** y **K6b** están entregados y archivados. OpenSpec/Git/Candidate siguen siendo la única autoridad semántica; el Assurance Graph es proyección. **K6c** es la siguiente iniciativa elegible.
+> **Corte documental:** v2.51.0, 2026-08-27 (estado alineado al roadmap; la dirección conceptual no cambia).
+> **Estado verificado:** O3, O4+O5/O4.1, O4.2, O6A, O2B, **K1**, **K2**, **K2.1**, **K2a**, **K3**, **`k3-readiness-remediation`**, **K4a**, **K5**, **K6a**, **K4b** y **K6b** (archivado v2.50.0; integridad cerrada en v2.51.0). OpenSpec/Git/Candidate siguen siendo la única autoridad semántica; el Assurance Graph es proyección. **K6c** es `next-eligible`.
 > **Roadmap:** orden, estado operativo y done criteria viven en [`../roadmaps/harness-evolution.md`](../roadmaps/harness-evolution.md).
 > **Precedencia documental:** ante diferencias de **orden o estado**, prevalece el roadmap; ante diferencias **conceptuales**, reconciliar antes de iniciar el slice.
 > **Investigación no normativa:** la trazabilidad completa P0–P27 vive en [`research/harness-kernel-graph-evidence-roadmap-fusion.md`](research/harness-kernel-graph-evidence-roadmap-fusion.md). La proporcionalidad de proceso y el programa de changes viven en [`research/proportional-process-and-change-program.md`](research/proportional-process-and-change-program.md).
@@ -855,7 +855,7 @@ Repositorios fixture reciben 10–30 cambios consecutivos. Se miden duplicación
 8. ~~K5: budgets (incl. autoridad/efectos) / failure / recovery~~ — hecho: archivado y publicado en v2.45.13 (remediaciones v2.45.7→v2.45.13).
 9. ~~K6a: primitivas de ejecución aislada (`CreateWorkspace`…`DisposeWorkspace`); no conoce Repair~~ — hecho: archivado y publicado en v2.46.7; frontera de procesos cerrada en v2.47.1; endurecimiento de frontera (política inmutable, fs mutante, live-identity, `worker_threads`) en v2.47.2.
 10. ~~K4b: orquesta Repair shadow (consume K6a; freeze Candidate vía K3)~~ — hecho: publicado en v2.48.0; corrección en v2.48.1; invariantes de integración en v2.48.2; cierre mode-only/baseline en v2.48.3.
-11. ~~K6b: verifier + provenance + Assurance Graph (proyección)~~ — verifier, strategies/provenance y proyección implementados; autoridad del grafo, K6c ChallengePlan y K6d complexity delta siguen pendientes.
+11. ~~K6b: verifier + provenance + Assurance Graph (proyección)~~ — hecho: publicado en v2.50.0; integridad (MUST, assessment/v1, collector, graph_id canónico, proyección fail-closed) cerrada en v2.51.0. K6c ChallengePlan es `next-eligible`; K6d complexity delta sigue pendiente.
 12. K7: ReviewAdapter + ReviewReducer + lineage; K8: CandidateEvaluationAttestation (emisión CAS).
 13. K9: shadow/replay/A-B; promoción de **un** profile (checkpoints intermedios ya validados).
 14. K10-delivery: DeliveryAuthorization **solo** del profile promovido; relación Candidate por etapas; resto fixed/deferred.

docs/roadmaps/harness-evolution.md

@@ -1,7 +1,7 @@
 # Roadmap general — kernel, grafo y evidencia
 
 > **Autoridad:** única fuente operativa del backlog transversal.
-> **Versión de referencia:** v2.50.0, 2026-08-27.
+> **Versión de referencia:** v2.51.0, 2026-08-27.
 > **Arquitectura:** [`../architecture/harness-evolution.md`](../architecture/harness-evolution.md).
 > **Investigación no normativa:** [`../architecture/research/harness-kernel-graph-evidence-roadmap-fusion.md`](../architecture/research/harness-kernel-graph-evidence-roadmap-fusion.md) (P0–P27). Proporcionalidad de proceso y Change Program: [`../architecture/research/proportional-process-and-change-program.md`](../architecture/research/proportional-process-and-change-program.md).
 > **Regla de estado:** los hechos se contrastan con código/OpenSpec; este roadmap no cambia el estado de un change ni sustituye sus artefactos.
@@ -78,7 +78,7 @@ Las iniciativas anteriores no se descartan. O20A, O13A–C, O15, O18, O19A/B y R
 | `done` | **K5** | Budgets (incl. autoridad/efectos), failures y recovery; remediaciones v2.45.7→v2.45.13 (authoritative enforcement, authority boundary/CAS concurrency, reconciliación, remediación técnica del núcleo y blindaje de concurrencia); archivado y publicado en v2.45.13 |
 | `done` | **K6a** | Worker isolation y work-order capsule; primitivas de ejecución aislada, integración con WorkerTransport, contención de filesystem y WorkResult canónico; archivado en v2.46.0, frontera de procesos cerrada en v2.47.1 y endurecida en v2.47.2 |
 | `done` | **K4b** | Repair shadow execution (WO→WR→integrate→Candidate); despacho exclusivo K6a, integración estricta, cápsula mínima, base derivada y registro 1:N; remediación de invariantes en v2.48.2 y cierre mode-only/baseline en v2.48.3 (`2026-08-26-k4b-mode-only-and-baseline-projection`) |
-| `done` | **K6b** | Verifier independiente, evidence strategies/provenance y Assurance Graph como proyección derivada (no autoridad); archivado y publicado en v2.50.0 (`2026-08-27-k6b-verifier-evidence-assurance-graph`) |
+| `done` | **K6b** | Verifier independiente, evidence strategies/provenance y Assurance Graph; integridad cerrada en v2.51.0 (`k6b-verification-integrity-remediation`) |
 | `next-eligible` | **K6c** | ChallengePlan policy-selected |
 | `pending` | K6d–K8 | Complexity delta, review authority, **Evaluation Attestation** |
 | `pending` | K9 | Gate de promoción shadow/replay/A-B (checkpoints intermedios ya validados) |
@@ -186,7 +186,7 @@ Campo canónico de binding al candidato: **`candidate_id`** (no `candidate_diges
 Entregado:
 G0/G0.1 ─ O2A ─ O3 ─ O4+O5/O4.1 ─ O4.2 ─ O6A ─ O2B → K1 → K2 → K2.1 → K2a → K3 → K4a → K5 → K6a → K4b → K6b
                                                                                                       ↓
-Next-eligible:                                                                                       K6c
+Next:                                                                                                K6c  (next-eligible)
                                                                                                       ↓
 Pending:     K6d → K7 → K8
                                                                    ↓
@@ -1011,7 +1011,7 @@ Vertical Repair shadow produce Candidate congelado. Gate de invariantes cerrado
 
 **Dependencias:** K4b + K6a + K3.
 
-**Estado:** `done`. Change OpenSpec `k6b-verifier-evidence-assurance-graph` archivado y publicado en v2.50.0: verifier independiente, strategies/provenance y Assurance Graph como proyección derivada. OpenSpec/Git/Candidate siguen siendo la única autoridad semántica. El grafo no concede lifecycle, approval ni delivery. K6c es next-eligible.
+**Estado:** `done`. Publicado en v2.50.0 (`k6b-verifier-evidence-assurance-graph`); integridad cerrada en v2.51.0 (`k6b-verification-integrity-remediation`): cobertura MUST del Obligation Manifest, assessments persistibles `assessment/v1`, provenance de collector del harness, `graph_id` canónico y proyección fail-closed. Este change no reabre el macro-slice. OpenSpec/Git/Candidate siguen siendo la única autoridad semántica. El grafo no concede lifecycle, approval ni delivery. **K6c es `next-eligible`.**
 
 **Absorbe/rebasa:** P12/P16; O15; separación apply/verify vigente.
 
@@ -1103,9 +1103,9 @@ external-unverified
 - equivalence manifest queda listo para K9;
 - ningún consumer trata el Assurance Graph como segunda fuente de verdad frente a OpenSpec/Git/Candidate.
 
-**Gate terminal:** verifier, strategy selector, provenance y Assurance Graph (proyección) conformes; checkpoint evidencia/challenges → `continue` | `revise` | `reject`. K6c no empieza antes.
+**Gate terminal:** verifier, strategy selector, provenance y Assurance Graph (proyección) conformes; checkpoint evidencia/challenges → `continue` | `revise` | `reject`. Cerrado en v2.51.0. Desbloquea K6c.
 
-### K6c — adversarial challenges (policy-selected) — **pending**
+### K6c — adversarial challenges (policy-selected) — **next-eligible**
 
 **Dependencias:** K6b (+ `PolicySnapshot` / strategy de evidencia).
 
@@ -2060,3 +2060,5 @@ Un Change Program (objetivo → children OpenSpec + cursor, ver investigación `
 - 2026-08-27: reconciliación aditiva (sin mover ruta crítica ni next-eligible): se nombra la distinción proceso intra-change vs Change Program inter-change; first-match de la tabla viva es compatibilidad, no K10; K6b/R4/K10 no absorben ese hueco. Investigación no normativa `docs/architecture/research/proportional-process-and-change-program.md`. Arquitectura: corte conceptual de la misma fecha; deuda stale K3/K4a/K5/K6a/K4b reconciliada.
 - 2026-08-27: K6b (`k6b-verifier-evidence-assurance-graph`) entra en apply: verifier independiente, evidence strategies/provenance y Assurance Graph proyección `implemented`; autoridad independiente del grafo, K6c/K7/K8 permanecen `target`. K6c queda next-eligible.
 - 2026-08-27: K6b cierra con verify PASS, 4R approved y archive transaccional; publicado en v2.50.0. K6c queda next-eligible.
+- 2026-08-27: K6b entra en `revise` (`k6b-verification-integrity-remediation`): cobertura MUST, assessments persistibles, provenance de collector, `graph_id` canónico y proyección fail-closed. K6c pasa a `blocked-by-K6b-remediation` hasta archive de esa remediación.
+- 2026-08-27: K6b (`k6b-verification-integrity-remediation`) cierra integridad con verify PASS, 4R successor approved y archive transaccional; publicado en v2.51.0. K6b queda `done`; K6c queda next-eligible.