feat: add protected checkpoint recovery to Lab (#765)

CHANGELOG.md

@@ -14,8 +14,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   the attempt core. Finalized restore reuses the original terminal; before-unit
   restore rejects a recorded dispatch even without completion, preventing
   uncertain execution from silently rerunning. Source commands `runner checkpoint
-  create` and `runner checkpoint resume` emit capture-approved results; API/Lab
-  integration and full production-CLI/platform acceptance remain pending. See
+  create` and `runner checkpoint resume` emit capture-approved results. Protected
+  HTTP operations and the source Lab preserve original outcomes, unavailable
+  progress, and exact JSON exports without raw source inspection. Linux clean-CLI
+  boundary-kill/fresh-process controls cover offline mock work; full roadmap,
+  native-platform acceptance, and independent approval remain pending. See
   [checkpoint contracts and migration limits](docs/EXECUTION_CHECKPOINTS.md).
 
 - **Logical execution-attempt embedding API (unreleased partial #764)** - Recover

README.md

@@ -32,8 +32,8 @@ pending; this is not available in installed v0.40.1 artifacts.
 The [checkpoint embedding preview](docs/EXECUTION_CHECKPOINTS.md) adds compatible
 whole-unit recovery. Its unreleased source CLI adds `runner checkpoint create`
 and `runner checkpoint resume`, with protected HTTP create/resume operations.
-The Lab can inspect captured checkpoint results but does not orchestrate these
-operations; full process/platform acceptance remains pending.
+The source Lab can create/resume checkpoints and export protected results.
+Ordered evaluation-level lineage and full platform acceptance remain pending.
 
 **v0.40.0 adds** bounded
 runner/text execution, selected AgentV assertion conversion, task-level MCP

docs/DEMO-GUIDE.md

@@ -21,6 +21,19 @@ the [terminal outcome and diagnostics](../site/src/content/docs/guides/runner-te
 with an exact terminal JSON export. This demo remains mock-only; opted-in exact
 Copilot/text CLI execution has its own paid-call and disabled-tool/skill boundaries.
 
+For the unreleased checkpoint source preview, build a clean versioned CLI and
+open **Runner contract → Create or resume an exact-unit checkpoint**. Use an
+existing private `.waza-artifacts` store under the served results directory.
+Create the captured mock example, then resume its finalized digest: new calls
+must be zero while the original terminal outcome and ordinal remain unchanged.
+Failed originals remain failed grading. Unknown progress is unavailable, not
+zero; an incompatible source is withheld. Download the exact protected JSON.
+This is not a whole-evaluation or ordered-lineage demonstration. Stop and join
+the original worker before before-unit recovery; stopping a browser request
+does not prove worker termination. See the [checkpoint workflow](../site/src/content/docs/guides/immutable-artifacts.mdx#checkpoint-public-source-preview).
+
+![Recorded protected mock API checkpoint reuse, not live inference](images/dashboard-lab-checkpoint.png)
+
 For an offline controlled-comparison demo, follow
 [the source-bound planning and collection example](CONTROLLED-COMPARISONS.md).
 Its one-cluster mock design intentionally remains inconclusive (selected gate

docs/EXECUTION_CHECKPOINTS.md

@@ -5,8 +5,8 @@ This unreleased, partial #765 layer builds on the
 provides an embedding service and unreleased source commands
 `runner checkpoint create` and `runner checkpoint resume`. Installed v0.40.1
 artifacts do not contain these commands. Protected HTTP create/resume operations
-are registered; Lab integration, broader native-platform evidence, and full
-#755 acceptance remain pending.
+are registered and consumed by the source Lab. Broader native-platform evidence
+and full #755 acceptance remain pending.
 
 ## Supported boundaries
 
@@ -135,6 +135,33 @@ the existing `RESULTS_DIR/.waza-artifacts` store or its configured Azure provide
 request JSON cannot select host paths or provider locations. Unsupported active
 providers fail closed. No raw checkpoint GET endpoint is exposed.
 
+The source Lab's **Runner contract** page explicitly creates or resumes one unit,
+shows compatibility, source withholding, original terminal outcome/ordinal, retry
+parent or reuse source, and separate new/original engine-call counts. Unknown
+progress stays unavailable; recorded zero stays zero. It downloads the exact
+protected response, not a reconstruction or raw report. Edits, new requests and
+reload failures clear old results; demo mode has no operation inputs.
+Contradictory original terminal exits or new engine-call counts are rejected,
+not displayed or exported as successful results.
+Error observations may retain partial or empty grader verdicts; they are not
+reclassified as successfully graded work.
+Completion declarations must match the original execution ID and canonical
+reservation digest before display or download. This checks declaration linkage,
+not storage integrity or execution authority.
+Resume source declarations must also hash to the requested checkpoint digest,
+independent of response object-key order, before display or download.
+Compatible resumes also require the native request-policy digest for the current
+versions, required/optional capabilities and explicit paid-execution consent.
+Native-valid reordered policy arrays remain equivalent; consent-only changes
+do not. The decision's full selection must match the admitted current attempt,
+while the reused original reservation and terminal retain their older attempt.
+A completed resume, including a partial publication failure, without a source is
+rejected unless the failure explicitly withholds it. Source-withheld failures remain
+available with their recorded unavailable progress.
+Create results cannot contain resume-only source or reuse state.
+Stopping a browser request is not proof of worker termination. This panel is
+not an ordered evaluation-wide attempt history or a whole-evaluation resume UI.
+
 `Put` can return a digest **and an error** if a blob may remain after failure.
 Preserve that reference and inspect it; never treat it as verified publication.
 Finalized checkpoint publication can fail after the original execution completed.
@@ -201,6 +228,7 @@ refuses it without new records. That control does not kill a CLI mid-dispatch.
 These are offline mock-execution controls, not native runtime, LFS installation,
 live inference, dashboard delivery, or Windows execution evidence.
 
-Lab integration, full current-head broad/platform gates and independent review
-remain release blockers. Broad test failures and timed-out lint runs are not
-passing checks even when focused recovery controls pass.
+Ordered evaluation-level lineage, generic protected submissions, full current-head
+broad/platform gates and independent review remain release blockers. Broad test
+failures and timed-out lint runs are not passing checks even when focused recovery
+controls pass.

docs/PRD.md

@@ -85,6 +85,14 @@ and an explicit manifest; no billing, evidence or execution attestation is
 invented. Attempt ownership remains the separate #764 contract. See
 [publication and migration](IMMUTABLE_ARTIFACTS.md).
 
+The #765 source preview adds protected checkpoint create/resume in the CLI,
+HTTP API and Lab. Lab displays original outcomes, retry/reuse references and
+explicitly unavailable progress, and downloads the unchanged protected result.
+Finalized restore preserves the original terminal with zero new calls;
+before-unit restore requires a stopped original worker and a fresh retry child.
+This does not complete #764 ordered lineage, native-platform acceptance or the
+v0.41 release. See [checkpoint guarantees and limits](EXECUTION_CHECKPOINTS.md).
+
 External-schema interoperability (#789) remains partial. The v0.40.0
 adapter projects one explicitly selected, commit-pinned AgentV case-sensitive
 `contains` assertion to a single native text grader and exports the same