pi-pstack-native 0.0.0-stage → 1.0.0

package/docs/upstream-provenance-0.15.17.json

@@ -0,0 +1,62 @@
+{
+  "repository": "cursor/plugins",
+  "version": "0.15.17",
+  "commit": "978ca6e9c0e012047cd5058b8e829240336e704e",
+  "verifiedAt": "2026-10-12T11:34:30Z",
+  "commitAPI": "https://api.github.com/repos/cursor/plugins/commits/978ca6e9c0e012047cd5058b8e829240336e704e",
+  "treeAPI": "https://api.github.com/repos/cursor/plugins/git/trees/978ca6e9c0e012047cd5058b8e829240336e704e?recursive=1",
+  "archiveURL": "https://api.github.com/repos/cursor/plugins/tarball/978ca6e9c0e012047cd5058b8e829240336e704e",
+  "treeTruncated": false,
+  "manifestBlob": "7d786382d4da27f0b9c7a59cd7c94457dffea3f6",
+  "verifiedPstackBlobs": 165,
+  "verification": "SHA-1 of Git blob header plus archive bytes matched every pstack blob in the official commit tree",
+  "previousBaseline": {
+    "version": "0.15.15",
+    "commit": "ccb5507cec1546dc88135c1139c811e6c59115ba"
+  },
+  "wordingReviewCommit": "00b14f1df36b32a8dd1529324a3a459f6563b127",
+  "changedIn01516": [
+    "pstack/.cursor-plugin/plugin.json",
+    "pstack/skills/architect/SKILL.md",
+    "pstack/skills/arena/SKILL.md",
+    "pstack/skills/benchmark-checklist/SKILL.md",
+    "pstack/skills/figure-it-out/SKILL.md",
+    "pstack/skills/how/SKILL.md",
+    "pstack/skills/interrogate/SKILL.md",
+    "pstack/skills/interrogate/references/reviewer-prompt.md",
+    "pstack/skills/interrogate/references/rubric.md",
+    "pstack/skills/poteto-mode/SKILL.md",
+    "pstack/skills/poteto-mode/playbooks/opening-a-pr.md",
+    "pstack/skills/poteto-mode/playbooks/trace-forensics.md",
+    "pstack/skills/principle-build-the-lever/SKILL.md",
+    "pstack/skills/principle-encode-lessons-in-structure/SKILL.md",
+    "pstack/skills/principle-explain-the-number/SKILL.md",
+    "pstack/skills/principle-minimize-reader-load/SKILL.md",
+    "pstack/skills/principle-model-the-domain/SKILL.md",
+    "pstack/skills/principle-never-block-on-the-human/SKILL.md",
+    "pstack/skills/principle-redesign-from-first-principles/SKILL.md",
+    "pstack/skills/principle-subtract-before-you-add/SKILL.md",
+    "pstack/skills/principle-test-behavior-not-implementation/SKILL.md",
+    "pstack/skills/principle-type-system-discipline/SKILL.md",
+    "pstack/skills/reflect/SKILL.md",
+    "pstack/skills/reflect/references/divergent-reviewer.md",
+    "pstack/skills/reflect/references/judgment-reviewer.md",
+    "pstack/skills/reflect/references/synthesizer.md",
+    "pstack/skills/reflect/references/tooling-reviewer.md",
+    "pstack/skills/technical-writing/SKILL.md",
+    "pstack/skills/unslop/SKILL.md",
+    "pstack/skills/why/SKILL.md",
+    "pstack/skills/why/references/epistemics.md",
+    "pstack/skills/why/references/investigator-prompt.md",
+    "pstack/skills/why/references/sources/code-archaeology.md",
+    "pstack/skills/why/references/sources/datadog.md",
+    "pstack/skills/why/references/sources/sentry.md",
+    "pstack/skills/why/references/synthesizer-prompt.md"
+  ],
+  "changedIn01517": [
+    "pstack/.cursor-plugin/plugin.json",
+    "pstack/README.md",
+    "pstack/skills/ketchup/SKILL.md",
+    "pstack/skills/poteto-help/SKILL.md"
+  ]
+}

package/docs/upstream-provenance-0.15.32.json

@@ -0,0 +1,11 @@
+{
+  "repository": "cursor/plugins",
+  "version": "0.15.32",
+  "commit": "99b609a72323b70204eab407c622f436e858e112",
+  "verifiedAt": "2026-10-26",
+  "treeTruncated": false,
+  "manifestBlob": "2d5accfaed9959e5a15226df48b17a59a16a4132",
+  "verifiedPstackBlobs": 200,
+  "verification": "Every extracted Git blob was hashed against the pinned target tree; vendored subset has its own SEMANTICS.json blob registry",
+  "previousBaseline": "978ca6e9c0e012047cd5058b8e829240336e704e"
+}

package/docs/upstream-provenance-0.15.36.json

@@ -0,0 +1,11 @@
+{
+  "repository": "cursor/plugins",
+  "version": "0.15.36",
+  "commit": "c742dd61706fbacc1b95f1a20b7b988cf096ce6c",
+  "verifiedAt": "2026-10-28",
+  "treeTruncated": false,
+  "manifestBlob": "1295f46878218d568a037344877f0b603a887696",
+  "verifiedPstackBlobs": 573,
+  "verification": "Every Git tree entry was independently hashed against its checked-out blob; the vendored executable and Markdown subset is enumerated in SEMANTICS.json.",
+  "previousBaseline": "99b609a72323b70204eab407c622f436e858e112"
+}

package/docs/upstream-sync/IMPLEMENTATION.md

@@ -0,0 +1,114 @@
+# Continuous upstream sync implementation
+
+This records the original M1–M4 delivery. The subsequent candidate ports and report revision behavior are documented in [PORT-0.15.23.md](PORT-0.15.23.md).
+
+The reviewed baseline remains pstack 0.15.17 at
+`978ca6e9c0e012047cd5058b8e829240336e704e`. No product runtime, Skill,
+Playbook, UPSTREAM.json or PARITY.json was changed by this implementation.
+
+## Architecture and milestones
+
+- M1: `core.mjs` pins Git commits, compares only `pstack/`, records raw Git
+  modes/blob evidence and renames, inventories names, classifies all file types,
+  and scans old/new keyword seams without executing upstream data. Reports are
+  fingerprinted independently of generatedAt. `parity.mjs` replaces version/count
+  constants with consistent metadata and validated evidence while retaining
+  pending/unsupported/file-existence/non-equivalence gates.
+- M2: every change has a task ID, local mapping, advisory disposition, risk,
+  owner placeholder, acceptance requirements and missing/pending evidence.
+  Markdown reports/task exports prohibit treating upstream instructions as
+  authorization. Report declarations are in `report.d.ts`; the parser is
+  exercised with valid and invalid fixtures.
+- M3: daily/manual read-only Actions scanning plus artifacts; explicitly enabled
+  notification uses a separate issues-write job, trusted local notifier and
+  all-state marker deduplication. No npm/LLM invocation in the scan job.
+- M4: explicit prepare → exported developer prompt → local implementation →
+  verify → dry-run/authorized draft proposal. Prepared branches use independent
+  worktrees. Dirty/existing worktrees, missing gates, stale verification,
+  history divergence and automatic baseline changes block proposals. Task
+  packages, verification evidence, worktrees, branches and PR bodies are keyed by
+  report revision, so two mapping revisions of one upstream pair never share an
+  artifact or a branch. Optional
+  commands reuse the native artifact fingerprint, including hidden-index-flag
+  rejection and pre/post-test source equality. No
+  provider/model is selected and no automatic inference is started.
+
+Seven local gates retain existing behavior; required Node CI matrix is unchanged.
+Upstream test suite is included in `npm run check` and therefore both matrix jobs.
+
+## Real pinned upstream example
+
+`reports/978ca6e9c0e0_837067b75aa2.json` compares the reviewed baseline to
+`837067b75aa20ab58d7b2a2f6c0531e37118c596`, manifest version 0.15.23.
+It contains 35 changes: 1 manifest, 2 agents, 1 other, 1 docs, 12 Skills,
+6 references, 2 Playbooks and 10 scripts. Skill `in-your-own-words` is added.
+There are 59 old/new keyword seam observations. These are observations from
+this pinned tree, not hard-coded assertions or completed semantic ports.
+
+Reproduce using a full upstream Git clone or network mode:
+
+```bash
+npm run upstream:scan -- --candidate-ref 837067b75aa20ab58d7b2a2f6c0531e37118c596
+npm run upstream:scan -- --from /path/to/cursor/plugins/pstack --candidate-ref 837067b75aa20ab58d7b2a2f6c0531e37118c596
+npm run upstream:plan -- --report docs/upstream-sync/reports/978ca6e9c0e0_837067b75aa2.json
+```
+
+Network fetch and offline scans produced the same sync ID/fingerprint. A repeated
+scan preserves the original timestamp and reports instead of duplicating them.
+The real candidate tasks remain pending; report gates are `not_run` because they
+refer to porting this upstream candidate, not testing the detector implementation.
+
+## Validation evidence
+
+See `VALIDATION.md` for final command outcomes and remaining limitations. The
+new offline tests exercise fixed Git commits, unchanged input, same-version
+changes, auxiliary-only changes, equal-count renames, deletions, new names,
+missing refs/manifest, history divergence, binary/oversized/special Git modes,
+symlink/traversal output safety, prompt-injection data, report tampering,
+parity blockers, all-state Issue deduplication and API errors. CLI/worktree
+integration tests execute actual npm gates in an isolated fixture and show that
+failed builds cannot be proposed and a verified candidate only produces a
+proposal, never integration or baseline updates.
+
+## Limitations and remaining risks
+
+- This delivers the synchronization mechanism, not the 35-item 0.15.23 semantic
+  port. Every real change needs adaptation evidence and maintainer approval.
+- Keyword seams and file mappings are advisory; text-only changes are not
+  assumed low risk. Unsupported behavior stays visible. No behavioral or
+  byte-level equivalence is inferred from names, counts or CI.
+- Verification fingerprints are integrity checks, not signed attestations.
+  CI must independently rerun checks and reviewers must inspect task evidence.
+- Worktrees and cooperative path checks are not an OS sandbox against malicious
+  code or concurrent filesystem tampering. Local tests run candidate code;
+  untrusted Agent execution requires stronger external isolation.
+- Live LLM inference, real notification writes, candidate push/draft creation
+  and real merge E2E are `not_run`. Issue/API behavior is tested with a fake API;
+  this cannot prove live credentials, rate-limit policy or repository settings.
+- The new scheduled workflow starts after merge into the default branch. Write
+  notifications require explicit repository configuration. GitHub Actions runs
+  are subject to repository billing/permissions and external availability.
+- No unattended M5 campaign, custom model integration, automatic retries of
+  remote writes, baseline promotion CLI or credential provisioning is included.
+  A failed run retries from the original baseline on the next scan. Proposal
+  lookup is deduplicated; concurrency of manually launched proposals is a
+  maintainer responsibility. Scheduled notification runs are serialized.
+- Reports with one SHA pair but inconsistent local mappings fail explicitly;
+  a deliberate remap requires separately reviewed report management.
+- Existing normal shipping policy prefers ready PRs; only the optional upstream
+  candidate workflow uses drafts as required by this PRD, without changing the
+  existing shipping runtime or playbooks.
+
+## Changed files
+
+- `.github/workflows/upstream-sync.yml`
+- `README.md`, `package.json`
+- `scripts/sync-upstream.mjs`, `scripts/check-parity.mjs`
+- `scripts/upstream/{core.mjs,cli.mjs,parity.mjs,report.d.ts}`
+- `scripts/upstream/{core.test.mjs,notify.cjs,notify.test.mjs}`
+- `docs/upstream-sync/{IMPLEMENTATION.md,VALIDATION.md}`
+- `docs/upstream-sync/reports/978ca6e9c0e0_837067b75aa2.{json,md}`
+- `docs/upstream-sync/tasks/978ca6e9c0e0_837067b75aa2.md`
+
+Historical provenance/report documents are retained. The current upstream
+baseline and Pi-native runtime source remain unchanged.

package/docs/upstream-sync/PORT-0.15.23.md

@@ -0,0 +1,142 @@
+# Proposed pstack 0.15.23 behavior ports
+
+Reviewed baseline: 978ca6e9c0e012047cd5058b8e829240336e704e (0.15.17).
+Candidate: 837067b75aa20ab58d7b2a2f6c0531e37118c596 (0.15.23).
+Native implementation base: ee8acfe5f65a44545a1334eae7e3a4de7f14c75b.
+
+Status is **proposed**, not integrated. UPSTREAM.json and PARITY.json retain the
+reviewed baseline. [The candidate receipt](candidate-0.15.23.json) accounts for
+all 35 pinned changed files, linking upstream blobs to implementation files,
+rationale, retained differences and pending semantic review. Available candidate
+Skills do not imply reviewed parity. No source count or CI pass is an equivalence
+percentage. Existing stronger native isolation and SHA verification are retained.
+
+## 1. Resource access and evidence boundaries
+
+Read-only children now have an explicit active SDK allowlist of confined
+read/ls/find/grep plus pstack_skill and pstack_evidence. The old empty allowlist
+prevented even the custom read tools from being activated. A real SDK + local
+protocol fixture verifies the model receives the six safe tools without shell,
+edit/write tools. Another fixture actually reads a bundled Skill, reads a parent
+issue snapshot and attempts an outside-worktree read; the outside secret is not
+returned. Traversal, symlink, script and arbitrary-resource cases are separately
+tested. No filesystem root was added to the ordinary read/search tools.
+
+The Skill tool provides only curated bundled Markdown, including scoped how
+and why templates. Resources are paginated with totalLines/hasMore. Poteto
+profiles require full Skill/playbook/leaf reading and preserve parent ownership,
+fresh-task default and costly-state recovery. This is a prompt contract, not a
+claim that every real model always follows the instructions.
+
+External evidence is a bounded, typed, per-category parent snapshot, persisted
+in AgentRecord/WorkflowCall and forwarded on recovery. Available/empty/
+unavailable/inconclusive are distinct. Seven categories have native source
+playbooks and epistemic/synthesis contracts. Excerpts are untrusted data,
+provenance is parent-supplied and not independently verified, and the parent
+must verify decisive citations against authorized sources. No MCP token/tool
+forwarding, network browsing in a read-only child or credential discovery occurs.
+Credential-bearing URLs/patterns in supplied packets are refused. Native gate
+logs now redact userinfo-only URLs and bare known token prefixes; the previously
+TODO credential regression is enabled and passes.
+
+## 2. PR-stack delivery and native watcher
+
+The legacy autopilot-stack DAG remains usable as local preparation and now
+returns deliveryContract=remote-stack-required. A completed DAG is not a
+completed PR stack. The new stack-plan/stack-open CLI operates on separate,
+clean, already published, linearly based worktree branches. Every lane requires
+current accepted behavior evidence against its exact parent; plans contain
+stable patch IDs and known remote refs. Only the first PR targets trunk.
+Opening is dry-run by default; --approve is explicit remote creation authority,
+never merge authority. Existing open ready PRs are reused only when head/base
+match. Unknown/moved refs, wrong ancestry, dirty/hidden-index flags, unrelated
+repository identities, empty patches, old ledgers and topology drift block.
+No automatic branch pushing, rebasing, force-pushing or retargeting is added.
+Those topology operations stay with the root and require their existing authority.
+
+The GitHub adapter paginates all review threads, refuses incomplete/looping
+cursors or conflicting page observations, and rechecks head/base after the read.
+Frontier blocks unresolved or unqueried review threads, draft state, bad merge
+state and stale verification even when CI/review rollups look green. It checks
+both parent SHA and parent branch when available. The pinned trunk and review/
+behavior policy are included in the authorization digest. After a lower merge,
+the next frontier must be prepared against trunk; it cannot silently land into
+the old parent branch. Native snapshots are waiting/reviewable/advance/complete/
+closed, not synthetic Cursor READY verdicts; auto-merge requested is not queue
+readiness. Server-side branch protection remains the final merge constraint.
+
+A real Git fixture creates two dependent worktrees, executes native gates, opens
+a simulated two-PR chain, then retries without duplicate writes. Wrong parent,
+moved refs, missing/unresolved review evidence, and the post-merge trunk
+transition are negative cases. Adapter fixtures exercise a second-page unresolved
+thread, cursor errors and a head move. These are not live GitHub-stack E2E.
+
+Native head/base changes always invalidate ledgers, even if patch IDs match.
+This deliberately preserves the stricter native rule instead of importing the
+upstream repeated-build/noise exception. Exact Cursor watcher/merge-queue verdict
+protocol and Bugbot pass-based dismissal heuristics remain unsupported. All
+unresolved threads are conservatively blocked, including Bugbot threads.
+
+## 3. Candidate content ports
+
+- Added in-your-own-words as a restatement-only Pi Skill with valid frontmatter.
+- Ketchup carries the newer plain/brief/checked-number/available-visual/action
+  rules while retaining Pi event-authorship and read-only boundaries.
+- Verification Skill creation requires usable empty-data startup, migrations/
+  seed preconditions, preserved proof after teardown and unverified stand-in
+  setup differences. A missing precondition fails the journey; it is not waived.
+- Poteto and comment-sicko profiles now carry the stronger delegation and scoped
+  findings contracts. The native role name was already comment-sicko.
+- How/why have readable scoped explorer/investigator/synthesizer/category
+  templates and explicit null-source accounting through the evidence boundary.
+- Root-cause guidance adds the cheap always-on deciding-fact signal.
+- Help/recipes explain advice versus starting paid work, the new restatement
+  Skill, resource/evidence boundaries and actual stack commands.
+- The reviewed upstream check-plan tool/fence regressions are reused with Node
+  assertions and a native loop marker. Bun bootstrap/lifecycle files are not
+  imported. Its strict ten-lane format is opt-in and grants no model spend.
+- Explicit-base worktree-audit reports head/base/dirty/ahead/merged without
+  deleting files. Cursor pinned-chat/transcript/sidebar and simulator cleanup
+  remain unsupported; owned/active/dirty cleanup rules stay in place.
+- Mostly textual upstream principle reductions retain native scoped guidance;
+  Cursor manifests are excluded from Pi package registration. These decisions
+  are mapped as equivalent/platform exclusions pending maintainer review.
+
+The scan reporter now permits intentional local mapping revisions for one SHA
+pair: it retains original evidence and writes a fingerprint-suffixed revision.
+Changed upstream evidence under the same ID remains an error. Read blob bytes
+are independently checked against Git SHA-1. Workflow artifacts resolve the
+validated reported revision path instead of assuming the original filename.
+Runtime status/workflow metadata now reads the reviewed UPSTREAM version rather
+than the obsolete hard-coded 0.15.15; it does not use the candidate version.
+
+## Validation and limitations
+
+Local release checks passed on Node 24.19.0, including 147 runtime tests,
+11 upstream-tool tests, types/build/SDK/benchmark and actual installed tarball
+smoke. Offline E2E has 9 passing tests and zero TODOs; fake E2E has 6 passing
+cases. Node 22.19.0 is also run through the existing release gates. Final
+remote matrix evidence is recorded in the PR description after it completes.
+Frontier hardening is followed by targeted shipping/verification tests and a
+fresh build. No existing gate or assertion is removed to manufacture success;
+read-only fixture assertions now require exactly the six safe tools instead of
+the old broken empty allowlist, with no shell/edit/write exposure.
+
+Live external LLM inference, comparative Pi/Cursor tasks, real native CLI stack
+creation/merge, and live Cursor-only platform behavior remain not_run. Maintainer
+semantic approval is pending. Worktree/path checks are not an OS sandbox against
+malicious code or concurrent filesystem attacks. Parent snapshots are not signed
+source attestations. Some new content has structural discovery/contract tests,
+not independently proven model conversational behavior. This PR does not claim
+full upstream semantic equivalence or advance the reviewed baseline.
+
+## Final CI regression correction
+
+The first final remote matrix failed in the relocated/symlinked benchmark test:
+the copied package fixture omitted UPSTREAM.json, which is now required to
+record the reviewed runtime version. The fixture now copies that shipped
+metadata too. Its existing assertions still require a real benchmark run,
+AUTHENTICATION_UNAVAILABLE, 20 case results and zero verified cases; no auth or
+verification check is bypassed. The corrected fixture passes on Node 22.19.0
+and 24.19.0. A fresh full remote matrix runs on the follow-up commit; the failed
+run is retained rather than relabeled or silently retried as a flake.

package/docs/upstream-sync/PR-PORT-0.15.23.md

@@ -0,0 +1,28 @@
+## Why
+
+Read-only Pi children had an empty active-tool allowlist, so they could not inspect their own worktree or bundled Skills. The local autopilot-stack DAG did not produce the linear PR stack its name suggested, and newer upstream pstack behavior was still unmapped. This implements the first three follow-up priorities while preserving existing isolation and the reviewed baseline.
+
+## What changed
+
+- Enable exactly confined file tools plus a safe bundled-Markdown registry and bounded parent-supplied evidence snapshots in read-only children; preserve missing-source uncertainty and recover packet state. Strengthen Poteto parent/delegate/fresh-task contracts.
+- Add verified linear stack-plan/stack-open CLI commands, per-parent patch/ledger checks and idempotent ready-PR opening. Mark the existing DAG as local preparation. Add all-page review threads and conservative native Frontier snapshots, including pinned-trunk transitions and policy-bound approval digests.
+- Port or explicitly map every one of the 35 pinned 0.15.23 changed files: new restatement Skill, ketchup/verification/Poteto/help rules, how/why references, audited plan-fence checker and explicit-base worktree audit. Preserve original reports and support content-addressed local mapping revisions.
+- Fix the existing URL-userinfo/bare-token evidence leak and convert the two real SDK/E2E defect TODOs into passing assertions. Runtime upstream status now uses the reviewed metadata instead of hard-coded 0.15.15.
+
+## Scope
+
+Pinned candidate 837067b75aa20ab58d7b2a2f6c0531e37118c596 (0.15.23), based on main ee8acfe5f65a44545a1334eae7e3a4de7f14c75b. UPSTREAM.json and PARITY.json remain unchanged at reviewed 0.15.17. Candidate ports stay proposed/pending semantic review; no automatic merge, publication or baseline advancement. See docs/upstream-sync/candidate-0.15.23.json for the per-file receipt and PORT-0.15.23.md for behavior and limits.
+
+## Blast Radius
+
+Child tool selection, bundled Skill prompts, GitHub read/PR preparation and compiled package APIs. Ordinary read/search confinement is unchanged; no MCP token/tool forwarding or arbitrary resource root is added. Unknown review-thread state now blocks shipping. Stack creation requires --approve, existing published clean worktrees and current behavior ledgers; it never pushes/rebases/retargets/merges automatically. Head/base changes still always require verification, even if patch IDs match.
+
+Exact Cursor watcher/merge-queue verdicts, Bugbot pass-based dismissal heuristics and Cursor sidebar/transcript/simulator cleanup remain unsupported. Parent snapshots are not independent source attestations. Worktrees/path checks are not OS sandboxes.
+
+## Verification
+
+Local seven release gates passed on Node 24.19.0 and Node 22.19.0. 147 runtime tests, 11 upstream tests, playbook/discovery checks, type/build/SDK/benchmark and actual tarball install passed. Offline E2E is 9/9 with zero TODOs; fake E2E is 6/6. Final Frontier hardening received 72 targeted regression tests on both Node runtimes, fresh types/build and offline/fake E2E. Delivery-contract/evidence snapshot persistence and read-boundary hardening received additional targeted regressions before the final build. Remote CI must independently run the full final matrix.
+
+A real SDK + loopback provider actually reads a bundled Skill and supplied issue evidence, refuses an outside-worktree secret and exposes no bash/edit/write tools. Real Git fixtures create/verify two separate dependent worktrees and simulate correctly based PR creation, reuse, drift and post-merge blockers. Adapter fixtures catch an unresolved thread on page two and incomplete/repeated cursors. These do not prove live LLM task quality or live native stack delivery.
+
+Live external LLM inference, Pi/Cursor task comparisons, native CLI live stack create/merge and maintainer semantic approval remain not_run/pending. No behavioral equivalence percentage is claimed.

package/docs/upstream-sync/VALIDATION.md

@@ -0,0 +1,76 @@
+# Validation evidence — continuous-sync implementation
+
+Local environment: Node 22.19.0 (npm-exec supplied runtime) and Node 24.19.0.
+No live model/provider credentials were used. Commands run from the repository.
+
+| Check | Node 22.19.0 | Node 24.19.0 |
+|---|---|---|
+| `npm run check` (source/package/parity/playbooks, upstream tests, runtime tests, review) | passed | passed (final sequential repeat) |
+| `npm run check:types` | passed | passed |
+| `npm run build` | passed | passed |
+| `npm run check:host` | passed | passed |
+| `npm run check:e2e` | passed (offline/fake only) | passed (offline/fake only) |
+| `npm run benchmark:smoke` | passed (20 fixtures; no external model) | passed |
+| `npm run smoke:installed` | passed (actual tarball install + SDK 1.2.0) | passed |
+| `node --test docs/parity.test.mjs` | included in playbook check | passed: 10 tests |
+| `npm run test:upstream` | passed: 10 tests | passed: 10 tests |
+
+The aggregate command was `npm run check:release` on both runtimes; no gates
+were removed or skipped. Offline E2E has 7 passing tests and 2 existing TODOs;
+fake-provider E2E has 6 passing tests. These TODOs remain unimplemented, not
+counted as passes. The Node 22.19.0/24.x GitHub CI matrix remains in place and
+will independently run after PR creation.
+
+## Failures retained for review
+
+The first Node 24 aggregate attempt failed in the pre-existing
+`test/protocol.test.ts` coordinator test (`8 !== 12` reconciled units). A
+subsequent full run passed all 135 runtime tests and all release gates. Another
+Node 24 run executed concurrently with the Node 22 release suite failed in the
+same existing test (`PROGRAM_CLOSE_BLOCKED: unit4,unit5`). No runtime/test
+assertions were weakened or changed. This indicates intermittent existing
+coordinator/verification behavior under load; its exact cause has not been
+established. The final sequential Node 24 release command passed all seven gates. Passing reruns
+do not establish that the intermittent failure has been fixed.
+
+## Real upstream and parsing evidence
+
+Network and local Git scans both returned:
+
+```json
+{"status":"changes_detected","syncId":"978ca6e9c0e0_837067b75aa2","changes":35}
+```
+
+Pinned candidate is `837067b75aa20ab58d7b2a2f6c0531e37118c596` (0.15.23).
+The result includes added `in-your-own-words`, all auxiliary changes and 59
+old/new seam hits. Repeated scans preserve fingerprint and timestamp;
+`upstream:plan` regenerates the same task package. The real report was parsed
+and fingerprint-validated by `validateReport`. Workflow YAML parsed successfully;
+`git diff --check` passed.
+
+## Not run / not claimed
+
+- Real upstream 35-item semantic port: pending; report gates `not_run`.
+- Live credentialed LLM/GitHub E2E, actual candidate draft/push and merge:
+  `not_run`; smoke and fake APIs do not substitute for these.
+- New scheduled workflow and opt-in Issue notification in the target repository:
+  `not_run` before merge/default-branch deployment.
+- Maintainer semantic approval and integration: `not_run`; UPSTREAM.json and
+  PARITY.json remain unchanged at the reviewed 0.15.17 baseline.
+
+Full local logs are retained in the implementation session. This document
+records exit results/test counts rather than arbitrary tool output or credentials.
+
+Final hardening reuses the native artifact fingerprint for optional candidate commands, including index-flag/submodule rejection and pre/post-gate source equality. The upstream CLI regression suite and syntax checks were rerun after this change on Node 24, and the upstream suite was rerun on Node 22.19.0.
+
+## Remote delivery blocked
+
+GitHub connector create-tree and create-branch both returned HTTP 403
+`Resource not accessible by integration`. Direct Git push failed because the
+workspace has no GitHub login credentials. No remote branch/PR or remote CI
+run was created, and none is claimed as successful. The connected plugin is
+installed but its current integration cannot perform these repository writes.
+A maintainer must supply an appropriately authorized GitHub connection to
+complete remote delivery. No credentials were requested in prompts or logged.
+
+After the user reconnected GitHub, creating the isolated feature branch succeeded. Remote delivery is now being resumed; the earlier HTTP 403 failures are retained above as historical evidence.