project-loop-harness 0.4.0__tar.gz → 0.4.3__tar.gz

{project_loop_harness-0.4.0/src/project_loop_harness.egg-info → project_loop_harness-0.4.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: project-loop-harness
-Version: 0.4.0
+Version: 0.4.3
 Summary: Local control plane for agentic project development loops
 Author: Project Loop Harness Contributors
 License: MIT
@@ -245,6 +245,85 @@ Markdown.
 See [docs/context-pack.md](docs/context-pack.md) for the contract shape and
 boundaries.
 
+## Work Briefs
+
+An optional `work-brief/v1` artifact can capture a target's reviewed intent,
+acceptance criteria, constraints, non-goals, and assumptions without changing
+the default `pcl start` path:
+
+```bash
+pcl contract validate --type work-brief/v1 work-brief.json --json
+pcl brief add work-brief.json --summary "Reviewed execution input" --dry-run --json
+pcl brief add work-brief.json --summary "Reviewed execution input" --json
+pcl brief review E-0001 --actor "agent:codex" --actor-kind agent \
+  --reason "Self-review completed; human gate remains open" --json
+# Normally run by the agent after the human says "approve" in conversation:
+pcl brief approve E-0001 --actor "human:owner" --actor-kind human \
+  --recorded-by "agent:codex" --recorder-kind agent \
+  --source-kind conversation --source-ref "conversation:<approval-reference>" \
+  --reason "Human explicitly approved the presented review packet" --json
+pcl brief show --target task:T-0001 --json
+pcl route recommend --target task:T-0001 --json
+pcl policy explain --target task:T-0001
+pcl route override --target task:T-0001 --profile assure \
+  --actor "human:owner" --reason "Require independent review" --dry-run --json
+pcl route override --target task:T-0001 --profile assure \
+  --actor "human:owner" --reason "Require independent review" --json
+pcl route current --target task:T-0001 --json
+```
+
+Brief content is immutable Evidence. Review and approval are separate
+hash-bound events: agent/system review cannot satisfy the human approval gate.
+Task context, resume, and dashboard data expose the factual actor kind, actor,
+recorder, source reference, timestamp, target, and bound hash without treating
+assumptions as facts. Humans normally approve in conversation or Cockpit; the
+agent records that decision through PCL. Direct human CLI use remains a
+compatibility path, not the expected routine UX. See
+[docs/work-brief-v1.md](docs/work-brief-v1.md).
+Deterministic Direct/Discover/Assure recommendation is documented in
+[docs/route-recommendation-v1.md](docs/route-recommendation-v1.md).
+Multi-axis resolution and field-level source rules are documented in
+[docs/adaptive-policy-v1.md](docs/adaptive-policy-v1.md).
+An applied override preserves its original recommendation and policy
+resolution as separate hash-bound Evidence. Task context, completion packets,
+and resume handoffs expose additive references without rewriting historical
+artifacts.
+
+## Evidence Sets
+
+Use `evidence-set/v1` when one passing artifact is not enough and the target
+must retain which declared reports were included or excluded:
+
+```bash
+pcl evidence-set plan --target task:T-0001 --work-root work/lp \
+  --manifest work/lp/reports/report-manifest.json \
+  --required-kind visual_check \
+  --include visual_check=E-0001:acceptance --json
+pcl evidence-set record --target task:T-0001 --work-root work/lp \
+  --manifest work/lp/reports/report-manifest.json \
+  --required-kind visual_check \
+  --include visual_check=E-0001:acceptance \
+  --summary "LP verification evidence set" --json
+```
+
+Planning is read-only. Recording creates immutable, target-linked Evidence and
+keeps known exclusions visible. See
+[docs/evidence-set-v1.md](docs/evidence-set-v1.md).
+
+Use a domain-neutral completion policy when a Test must require an external
+JSON verdict rather than merely retain the report set:
+
+```bash
+pcl completion evaluate --policy completion-policy.json \
+  --evidence-set E-0003 --test TC-0001 --json
+pcl test pass TC-0001 --summary "Completion contract passed" \
+  --evidence-id E-0003 --completion-policy completion-policy.json --json
+```
+
+Only allowlisted JSON predicates are evaluated. The Evidence Set must target
+the exact Test, remain complete, and retain matching report hashes. See
+[docs/completion-policy-v1.md](docs/completion-policy-v1.md).
+
 ## Explainable Code Context
 
 Build a local code context snapshot when an agent handoff needs auditable code
@@ -318,6 +397,12 @@ review/continuation handoff and presumes a change already exists for the target.
 }
 ```
 
+When no active work or real human decision exists, the same schema returns
+`type: "idle"`, `command: null`, `requires_human: false`, and
+`run_policy: "idle"`. If a user has already supplied explicit work, pass that
+literal intent to `pcl start "<intent>"` instead of inventing a Goal title or
+asking for a second approval merely to register it.
+
 Use:
 
 ```bash

{project_loop_harness-0.4.0 → project_loop_harness-0.4.3}/README.md

@@ -211,6 +211,85 @@ Markdown.
 See [docs/context-pack.md](docs/context-pack.md) for the contract shape and
 boundaries.
 
+## Work Briefs
+
+An optional `work-brief/v1` artifact can capture a target's reviewed intent,
+acceptance criteria, constraints, non-goals, and assumptions without changing
+the default `pcl start` path:
+
+```bash
+pcl contract validate --type work-brief/v1 work-brief.json --json
+pcl brief add work-brief.json --summary "Reviewed execution input" --dry-run --json
+pcl brief add work-brief.json --summary "Reviewed execution input" --json
+pcl brief review E-0001 --actor "agent:codex" --actor-kind agent \
+  --reason "Self-review completed; human gate remains open" --json
+# Normally run by the agent after the human says "approve" in conversation:
+pcl brief approve E-0001 --actor "human:owner" --actor-kind human \
+  --recorded-by "agent:codex" --recorder-kind agent \
+  --source-kind conversation --source-ref "conversation:<approval-reference>" \
+  --reason "Human explicitly approved the presented review packet" --json
+pcl brief show --target task:T-0001 --json
+pcl route recommend --target task:T-0001 --json
+pcl policy explain --target task:T-0001
+pcl route override --target task:T-0001 --profile assure \
+  --actor "human:owner" --reason "Require independent review" --dry-run --json
+pcl route override --target task:T-0001 --profile assure \
+  --actor "human:owner" --reason "Require independent review" --json
+pcl route current --target task:T-0001 --json
+```
+
+Brief content is immutable Evidence. Review and approval are separate
+hash-bound events: agent/system review cannot satisfy the human approval gate.
+Task context, resume, and dashboard data expose the factual actor kind, actor,
+recorder, source reference, timestamp, target, and bound hash without treating
+assumptions as facts. Humans normally approve in conversation or Cockpit; the
+agent records that decision through PCL. Direct human CLI use remains a
+compatibility path, not the expected routine UX. See
+[docs/work-brief-v1.md](docs/work-brief-v1.md).
+Deterministic Direct/Discover/Assure recommendation is documented in
+[docs/route-recommendation-v1.md](docs/route-recommendation-v1.md).
+Multi-axis resolution and field-level source rules are documented in
+[docs/adaptive-policy-v1.md](docs/adaptive-policy-v1.md).
+An applied override preserves its original recommendation and policy
+resolution as separate hash-bound Evidence. Task context, completion packets,
+and resume handoffs expose additive references without rewriting historical
+artifacts.
+
+## Evidence Sets
+
+Use `evidence-set/v1` when one passing artifact is not enough and the target
+must retain which declared reports were included or excluded:
+
+```bash
+pcl evidence-set plan --target task:T-0001 --work-root work/lp \
+  --manifest work/lp/reports/report-manifest.json \
+  --required-kind visual_check \
+  --include visual_check=E-0001:acceptance --json
+pcl evidence-set record --target task:T-0001 --work-root work/lp \
+  --manifest work/lp/reports/report-manifest.json \
+  --required-kind visual_check \
+  --include visual_check=E-0001:acceptance \
+  --summary "LP verification evidence set" --json
+```
+
+Planning is read-only. Recording creates immutable, target-linked Evidence and
+keeps known exclusions visible. See
+[docs/evidence-set-v1.md](docs/evidence-set-v1.md).
+
+Use a domain-neutral completion policy when a Test must require an external
+JSON verdict rather than merely retain the report set:
+
+```bash
+pcl completion evaluate --policy completion-policy.json \
+  --evidence-set E-0003 --test TC-0001 --json
+pcl test pass TC-0001 --summary "Completion contract passed" \
+  --evidence-id E-0003 --completion-policy completion-policy.json --json
+```
+
+Only allowlisted JSON predicates are evaluated. The Evidence Set must target
+the exact Test, remain complete, and retain matching report hashes. See
+[docs/completion-policy-v1.md](docs/completion-policy-v1.md).
+
 ## Explainable Code Context
 
 Build a local code context snapshot when an agent handoff needs auditable code
@@ -284,6 +363,12 @@ review/continuation handoff and presumes a change already exists for the target.
 }
 ```
 
+When no active work or real human decision exists, the same schema returns
+`type: "idle"`, `command: null`, `requires_human: false`, and
+`run_policy: "idle"`. If a user has already supplied explicit work, pass that
+literal intent to `pcl start "<intent>"` instead of inventing a Goal title or
+asking for a second approval merely to register it.
+
 Use:
 
 ```bash

project_loop_harness-0.4.3/agent-tasks/0141-idle-routing-repair.md

@@ -0,0 +1,71 @@
+# 0141: Idle routing without a redundant human gate
+
+- **Status:** Approved implementation slice
+- **Milestone:** v0.4.1 Integrity Migration
+- **Priority:** P0
+- **Estimated size:** S
+- **Dependencies:** 0136 (`pcl start`), 0140a–0140c (v0.4.0 Integrity Gate)
+
+## Problem
+
+When a project has no active work, `pcl next` currently recommends a fixed
+`Reach feature coverage` Goal and marks that recommendation as a human
+decision. If a user has already supplied a literal implementation intent, this
+creates a meaningless confirmation round trip. The v0.4.0 `pcl start`
+command is already the guarded entry point for registering that intent.
+
+## Goal
+
+Represent a genuinely idle project as neutral read-only state, without
+inventing a durable mutation or human approval. Teach agents to pass an
+explicit user intent literally to `pcl start`.
+
+## Contract
+
+When no active Goal, Task, Defect, Workflow, Decision, Escalation, checkpoint,
+or uncovered Feature action exists, `pcl next --json` returns the ordinary
+guided-action keys with:
+
+- `type: "idle"`;
+- `command: null`;
+- `target: null`;
+- `blocking: false`;
+- `requires_human: false`;
+- `safe_to_run: false` because there is no command to execute;
+- `run_policy: "idle"`;
+- factual guidance to call `pcl start "<intent>"` only when explicit intent is
+  available.
+
+The `command` field remains a string for executable and human-gated actions.
+Its nullability is limited to action shapes that intentionally have no command.
+
+## Scope
+
+- Replace the terminal fixed-Goal fallback in `src/pcl/commands.py`.
+- Keep `pcl next --explain`, dashboard data, dashboard HTML, MCP output, and
+  distribution smoke behavior coherent with the nullable idle command.
+- Update the three byte-identical `project-control-loop` Skill copies.
+- Update direct tests and the generated baseline snapshot with an intentional
+  contract-drift note.
+
+## Invariants
+
+- `pcl next` remains read-only.
+- Real human gates retain their existing priority and fields.
+- Existing active-work, defect, task, workflow, checkpoint, and uncovered
+  Feature routing is unchanged.
+- `pcl start` continues to treat intent as literal text and preserves its
+  duplicate-active-work guard.
+- No schema migration, dependency, LLM call, agent launch, or remote operation.
+
+## Acceptance criteria
+
+- A freshly initialized empty project returns the idle contract above.
+- `pcl next --explain` does not print `None` as an executable command.
+- An open Decision still returns the existing human-gated action before idle.
+- Example and fresh-wheel distribution tests expect the idle contract.
+- The current baseline snapshot is regenerated and its intentional delta is
+  documented.
+- The canonical Skill copies remain byte-identical.
+- Targeted tests, full `pytest`, `ruff check .`, strict validation, and render
+  pass.

project_loop_harness-0.4.3/agent-tasks/0142-lifecycle-repair-planner.md

@@ -0,0 +1,172 @@
+# 0142: Lifecycle repair planner for existing projects
+
+- **Status:** Approved implementation slice
+- **Milestone:** v0.4.1 Integrity Migration
+- **Priority:** P1
+- **Estimated size:** L
+- **Dependencies:** 0140b (lifecycle findings and mutation guards), 0141
+- **Parallel-safe with:** none of 0143–0145; this task establishes the repair-plan contract used by 0143 and touches shared CLI/validator surfaces
+- **DB schema:** remains 8
+
+## Problem
+
+v0.4.0 prevents new false-completion states, but an existing project can still
+contain terminal rows that predate the stronger lifecycle contract. Strict
+validation can identify those rows, yet operators currently have to infer a
+repair sequence from prose errors. Blindly changing statuses or choosing a
+Story, Evidence item, Verification, or human decision would manufacture
+semantics that the harness does not know.
+
+## Goal
+
+Add a deterministic, wholly read-only lifecycle repair planner. It classifies
+each inconsistent row and shows concrete inspection or later repair commands,
+but owns no mutation path or link mutation service.
+
+## CLI contract
+
+```text
+pcl repair lifecycle [--dry-run] [--json]
+```
+
+- Bare `pcl repair lifecycle` and explicit `--dry-run` are equivalent
+  read-only operations.
+- Planning never writes the DB, events/outbox/JSONL, reports, dashboard, copied
+  Evidence, or any other file.
+- The command never executes a command string from its own output.
+- All mutation modes belong to task 0143, which consumes this task's public
+  action model in one direction.
+
+The JSON result uses an additive, versioned `lifecycle-repair-plan/v1` shape:
+
+```json
+{
+  "contract_version": "lifecycle-repair-plan/v1",
+  "mode": "plan",
+  "mutated": false,
+  "summary": {
+    "structural": 0,
+    "semantic": 1,
+    "human_review": 0,
+    "unsupported": 0
+  },
+  "actions": [
+    {
+      "action_id": "LR-0001",
+      "finding_code": "test_story_required",
+      "classification": "semantic",
+      "action_kind": "inspect_story_candidate",
+      "sort_key": [1, "test_case", "TC-0001", "inspect_story_candidate", "test_story_required"],
+      "entity": {"type": "test_case", "id": "TC-0001"},
+      "related": [{"type": "user_story", "id": "US-0001"}],
+      "safe_to_apply": false,
+      "requires_human": true,
+      "command": "pcl story read US-0001 --json",
+      "reason": "A Story link requires an explicit operator choice."
+    }
+  ]
+}
+```
+
+`classification` is a closed public enum with exactly `structural`, `semantic`,
+`human_review`, and `unsupported`. `summary` contains exactly those four count
+keys, including zero values. `action_kind` is a stable public snake-case action
+identifier such as `add_missing_evidence_link`, `inspect_story_candidate`, or
+`record_goal_verification`; consumers must not infer it from prose.
+
+Every action publishes its canonical `sort_key` array:
+
+```text
+[classification_rank, entity.type, entity.id, action_kind, finding_code]
+```
+
+The fixed ranks are structural=0, semantic=1, human_review=2, unsupported=3.
+Actions are sorted lexicographically by this array before sequential
+`LR-0001` IDs are assigned. IDs are deterministic within the plan and are not
+persisted domain IDs. An action with no primary entity uses empty strings for
+both entity sort components. Text output uses the same order and four labels.
+
+## Structural versus semantic boundary
+
+An action is structural only when all source and target IDs already exist and
+the relationship is unambiguous from authoritative stored data. Initial
+supported examples are:
+
+- restore a missing `test_case` / `acceptance` `evidence_links` row when the
+  terminal Test already stores that same healthy `evidence_id`;
+- restore a missing completion-packet link when the validated packet itself is
+  targeted to the same entity and no conflicting target/link exists.
+
+The following are always plan-only semantic or human actions, even when only
+one candidate appears to exist:
+
+- approve, review, or waive a Story;
+- choose or change a Test-to-Story relationship;
+- choose, create, copy, or replace Evidence;
+- change a lifecycle status;
+- record or approve a Verification;
+- close or reopen a Goal, Feature, Test, Decision, or Escalation;
+- choose between conflicting candidates or reinterpret legacy inline text.
+
+An exactly-one Story heuristic is therefore not sufficient for automatic
+linking or approval. Unsupported corruption is reported, not normalized.
+
+## Scope
+
+- Add `src/pcl/lifecycle_repair.py` and the `pcl repair lifecycle` parser and
+  dispatch path.
+- Reuse lifecycle predicates from `src/pcl/validators.py`; do not parse legacy
+  validation message strings to discover entities.
+- Add synthetic existing-project fixtures and
+  `tests/test_lifecycle_repair.py`.
+- Extend `docs/recovery-playbook.md` with the plan-first migration path.
+
+## Invariants
+
+- Default and dry-run modes cause zero mutation, including no projector flush
+  or generated report.
+- Semantic actions are never auto-executed, auto-approved, or converted to a
+  structural action by candidate count.
+- 0142 does not add an apply flag, internal link mutation service, transaction,
+  event, or post-repair validation path.
+- No raw SQL or generated HTML becomes an operator repair interface.
+- Existing validation severity policy, same-status no-op behavior, CLI flags,
+  and JSON fields remain backward compatible.
+- No migration, dependency, LLM call, agent launch, or remote operation.
+
+## Non-goals
+
+- Generic database repair or audit-log rebuild (covered by 0129).
+- Any repair application, including unambiguous structural link mutation.
+- Internal/common link mutation services and dedicated link commands; these
+  belong to 0143.
+- Post-repair strict-pass orchestration or acceptance; this belongs to 0143.
+- Automatic application of suggested shell commands.
+- Schema-backed generic Verification targets.
+- Enabling enforced lifecycle policy for unrepaired existing projects.
+
+## Acceptance criteria
+
+- Bare and explicit dry-run produce the same deterministic plan and leave DB,
+  event/outbox/JSONL, report, dashboard, and Evidence hashes unchanged.
+- A passing Test with no Story remains a semantic operator action; an
+  exactly-one candidate is not linked or approved.
+- A missing link derivable from an existing healthy `test_cases.evidence_id`
+  is classified structural and described with a stable `action_kind`, but is
+  not applied.
+- Conflicting, missing, drifted, cross-target, or wrong-role Evidence is not
+  auto-repaired.
+- `classification` rejects values outside the fixed four-value enum; summary
+  always contains exactly the four classification keys.
+- The published sort key, fixed ranks, `action_kind`, and post-sort action IDs
+  produce identical output for identical state.
+- Text and JSON output contain the same action classes and concrete IDs; stdout
+  remains pure JSON in JSON mode.
+- Targeted tests, full `pytest`, `ruff check .`, strict validation, and render
+  pass in an initialized verification project.
+
+## Evidence required to close
+
+- Before/after hashes and row/event counts proving dry-run zero mutation.
+- JSON fixtures for all four classifications, public action kinds, canonical
+  sort keys, zero-filled summary keys, ambiguous cases, and idempotent reruns.

project_loop_harness-0.4.3/agent-tasks/0143-terminal-link-repair.md

@@ -0,0 +1,155 @@
+# 0143: Dedicated terminal link repair commands
+
+- **Status:** Approved implementation slice
+- **Milestone:** v0.4.1 Integrity Migration
+- **Priority:** P1
+- **Estimated size:** L
+- **Dependencies:** 0142 (repair-plan action model)
+- **Parallel-safe with:** none of 0144–0145; shared `cli.py`, `validators.py`, and Evidence services require a serial merge
+- **DB schema:** remains 8
+
+## Problem
+
+The v0.4.0 same-status contract deliberately makes repeated terminal commands
+no-ops. Consequently, rerunning `pcl test pass` must not be overloaded to
+repair a missing Story or Evidence link on an existing passing Test. Operators
+need an explicit, narrow, audited path for relationship repair that does not
+replay the lifecycle transition or rewrite historical Evidence.
+
+## Goal
+
+Add dedicated Test and Evidence link commands that validate the complete
+relationship before mutation, repair links atomically, and leave lifecycle
+status and historical records untouched. Consume the 0142 action model to add
+the only structural apply path for lifecycle repair.
+
+## CLI contract
+
+```text
+pcl test link TC-0001 [--story US-0001] [--evidence-id E-0007] --summary "..."
+pcl evidence link E-0007 --target test_case:TC-0001 --role acceptance --summary "..."
+pcl repair lifecycle --apply-structural [--json]
+```
+
+`pcl test link` requires at least one of `--story` or `--evidence-id`. When both
+are supplied, both changes are validated before one transaction commits. The
+JSON response reports `changed`, before/after relationship values, the event
+ID, and any stable warnings.
+
+`pcl evidence link` accepts an existing Evidence ID, a parsed
+`<target-type>:<target-id>` reference, a role, and an audit summary. It is a
+generic link insert only for known schema-8 target types and compatible roles;
+it does not silently update a Test's `evidence_id`. For a terminal Test
+`acceptance` link, the stored Test pointer must already match, otherwise the
+command returns a typed error directing the operator to `pcl test link`.
+
+`pcl repair lifecycle --apply-structural` first builds the current
+`lifecycle-repair-plan/v1` plan through 0142, selects only actions with
+`classification: structural` and `safe_to_apply: true`, and dispatches their
+public `action_kind` values through the internal link mutation service owned by
+this task. Unknown action kinds, non-structural actions, or changed
+preconditions fail closed. No command string from the plan is shell-executed.
+
+## Validation rules
+
+- Test, Story, Evidence, and target IDs must exist before mutation.
+- A Story linked to a Test must belong to the same Feature. A passing Test may
+  link only an `approved` or explicitly `waived` Story.
+- Acceptance Evidence for a direct passing Test must be an allowed healthy
+  hash-pinned Evidence type and target the same Test.
+- Feature `acceptance` Evidence and completion-packet roles must satisfy the
+  existing terminal guard predicates.
+- Completion-packet links require a valid packet whose embedded target matches
+  the requested target.
+- Known exclusive target-bound roles reject an existing conflicting target;
+  generic supporting links retain their existing many-target behavior.
+- All requested relationships are checked before opening the mutation window,
+  then re-checked inside it to prevent a time-of-check/time-of-use repair.
+
+## Scope
+
+- Add the parser and dispatch paths in `src/pcl/cli.py`.
+- Add one internal relationship mutation service used by `pcl test link`,
+  `pcl evidence link`, and structural lifecycle apply. The service may be
+  organized through `src/pcl/stories.py` and `src/pcl/evidence.py`, but 0142
+  remains read-only and does not import it.
+- Extend `src/pcl/lifecycle_repair.py` only with the explicit
+  `--apply-structural` consumer of the already-public 0142 plan model.
+- Keep `src/pcl/validators.py` consistent with the write-side rules.
+- Add direct and regression coverage in `tests/test_stories.py`,
+  `tests/test_evidence_add.py`, and lifecycle validation tests where needed.
+- Update recovery documentation and command help examples.
+
+## Mutation and event contract
+
+- One successful `pcl test link` call uses one transaction and appends one
+  `test_links_repaired` event containing old/new Story and Evidence IDs plus
+  the operator summary.
+- One successful `pcl evidence link` call uses one transaction and appends one
+  `evidence_link_added` event containing Evidence, target, role, and summary.
+- One successful structural plan application re-checks every selected action
+  inside one transaction and appends one
+  `lifecycle_structural_repair_applied` event containing plan action IDs,
+  action kinds, and before/after relationships.
+- Repeating an exact existing relationship returns `changed: false` and adds no
+  event or outbox row.
+- Replacing the current Test Evidence pointer does not delete or mutate the
+  previous Evidence row or historical `evidence_links`; the event makes the
+  pointer change reviewable.
+
+## Invariants
+
+- `pcl test pass` and every other same-status terminal call remain no-ops; they
+  do not acquire hidden repair behavior.
+- Link commands never change status, `created_at`, prior transition events,
+  Evidence content/hash, Verification, Decision, or approval state.
+- Any invalid or conflicting input produces a typed error and zero domain,
+  event, outbox, or JSONL mutation.
+- A stale structural plan or changed relationship precondition fails with a
+  typed error and zero mutation; partial batch repair is not observable.
+- Structural apply never dispatches semantic, human-review, or unsupported
+  actions and never executes plan command strings.
+- Every mutation goes through application services and the event outbox; no
+  operator-facing raw SQL path is added.
+- JSON changes are additive. Schema 8, dependencies, and legacy flags remain
+  unchanged.
+
+## Non-goals
+
+- Story approval/waiver, Goal verification, or lifecycle status repair.
+- Link deletion, bulk relinking, or automatic conflict resolution.
+- Treating mutable legacy path strings as terminal proof.
+- Generic Verification targets or a schema migration.
+- Changing the 0142 classification/action/sort contract.
+
+## Acceptance criteria
+
+- A same-Feature approved Story and healthy acceptance Evidence can repair an
+  existing passing Test in one audited transaction.
+- Cross-Feature Story, draft/review Story on a passing Test, missing/drifted
+  Evidence, incompatible role, packet target mismatch, and exclusive-role
+  target conflict each fail with zero mutation.
+- A combined command cannot commit the Story half if Evidence validation fails.
+- `pcl evidence link` refuses to create a half-repaired terminal Test when its
+  `evidence_id` pointer differs.
+- Exact reruns are event-free no-ops; changing a Test pointer preserves all old
+  Evidence rows and links.
+- Structural apply accepts only recognized structural `action_kind` values,
+  is atomic and audited, and is an event-free no-op when the current plan has no
+  applicable structural actions.
+- Existing tests prove repeated `pcl test pass` remains a same-status no-op.
+- Repaired direct and Workflow-backed fixtures satisfy the existing lifecycle
+  validator without weakening its predicates.
+- After structural apply and explicit operator execution of every semantic or
+  human-review action in the supported synthetic fixture,
+  `pcl validate --strict --json` passes.
+- Targeted tests, full `pytest`, `ruff check .`, strict validation, and render
+  pass.
+
+## Evidence required to close
+
+- Before/after Test and `evidence_links` JSON plus the single emitted event.
+- Zero-mutation counts/hashes for every rejection case.
+- Same-status and exact-link idempotency regression output.
+- A complete 0142 plan → 0143 structural apply → explicit semantic actions →
+  strict-pass transcript.