Add independent reproduction and Node application study

.gitattributes

@@ -0,0 +1,5 @@
+/.gitattributes text eol=lf
+benchmark-evidence/** -text diff whitespace=cr-at-eol
+benchmark-evidence/**/*.log whitespace=-blank-at-eol,-blank-at-eof,cr-at-eol
+benchmark-evidence/**/*.gz -text -diff
+benchmark-evidence/**/*.pyc -text -diff

.gitignore

@@ -2,3 +2,5 @@
 # macOS AppleDouble in-progress turds
 .!*!*
 .gstack/
+__pycache__/
+*.pyc

benchmark-evidence/.gitignore

@@ -0,0 +1,6 @@
+# Credentials and account-specific cloud state never enter the public fork.
+*.env
+*.key
+*.token
+cloud/credentials/
+cloud/session/

benchmark-evidence/BASELINE_README.md

@@ -0,0 +1,77 @@
+# Independent RocketRide Benchmark Study
+
+This bundle contains three deliberately separate tracks at upstream commit
+`43be41acb58558dfae8e2e3deb86d8a00cb1b1c8`.
+
+| Track | Evidence status | Promotion status |
+|---|---|---|
+| A: unchanged upstream reproduction | Complete on supported Linux; Windows appendix incomplete | Independent, unsubmitted |
+| B: Node application extension | Complete, 90 results and 30 RocketRide traces | Blocked by the fixed 10 s protocol deadline |
+| C: RocketRide Cloud appendix | Control complete and audited; exact upstream pipe blocked | Operational, unsubmitted |
+
+## Track A
+
+The primary reproduction ran the unchanged benchmark source, fixtures,
+competitors, pool sizes, retries, and timeouts on Ubuntu 22.04 under WSL2. All
+10 fault-isolation, 10 concurrent-processing, 10 data-isolation, and one
+authoring-effort cells completed. The earlier Windows run is retained as an
+incomplete exploratory appendix rather than substituted for the supported Linux
+result.
+
+This supports only an independent reproduction claim. It does not support a
+general framework-superiority claim or an official RocketRide claim without an
+external acceptance receipt.
+
+## Track B
+
+Track B compares native worker threads, RocketRide 1.2.0 with local engine
+3.2.1.30, and real `langchain-core==0.3.86`. All executors return candidate-only
+results; application validation, version checks, CAS, proposals, review, and
+final writes remain application-owned.
+
+The N=3 matrix completed for five frozen fixtures in normal and injected
+hard-failure variants. Its orchestration gate passed, but the post-run protocol
+audit found that all 30 RocketRide envelopes exceeded the fixed 10,000 ms total
+deadline. Evidence is complete; protocol admission and promotion are blocked.
+
+## Track C
+
+The `JULY2026BENCHMARK` promotion was redeemed for Starter at 100 percent off.
+The subscription readback reports active Starter, USD 0 due at checkout, and
+renewal disabled at period end. A pre-registered Cloud-native control passed 48
+normal requests, 9 unaffected failure-phase requests, 3 expected requests to a
+terminated task, and 0 cross-task leaks across 12 resident tasks.
+
+The unchanged upstream fault-isolation pipeline did not run on Cloud because
+the hosted catalog lacks its custom `workload` service. Track C is therefore an
+operational control, not an upstream benchmark score and not a comparison with
+LangChain. The failed first receipt and passing rerun are both retained.
+
+## Frozen Applications
+
+| App | Frozen base | Adapter commit |
+|---|---|---|
+| NodeRoom | `ca25e347dc467bc37f06918e1a18656f7336ee28` | `353ddcce5b4ee791c4a0d70713a47ab302a643b9` |
+| NodeBenchAI | `6ed0a58eeda993ff2a937ea4bacc2856756dd521` | `651020807ed13ea3f36cdf6be601f5a874705b29` |
+| NodeSlide | `dd67e4c642c40e6bb414af617a67a31dbed507c5` | `702020d5cdd9785bfb4e6b79a30246a3c88a23b5` |
+| NodeVideo | `bb79bc385de93c90cee89b160fc801d18372d89e` | `cf2c696e1c0ccea6ebb43617dc81a23f05e785da` |
+
+## Evidence Map
+
+- Track A primary summary: `baseline/linux/linux-20260715T122439Z/aggregate/summary.json`
+- Track A Windows appendix: `baseline/concurrent-work/aggregate/summary.json`
+- Track B pre-registration: `extensions/node-suite/PRE_REGISTRATION.md`
+- App verifier: `extensions/node-suite/app-verification.json`
+- Post-run app verifier: `extensions/node-suite/app-verification-post-run.json`
+- Track B original scorecard: `extensions/node-suite/runs/node-suite-20260715T150250Z/scorecard.json`
+- Track B superseding protocol audit: `extensions/node-suite/runs/node-suite-20260715T150250Z/audit.json`
+- Track C pre-registration: `cloud/PRE_REGISTRATION.md`
+- Track C failed first run: `cloud/runs/cloud-20260715T225614Z/audit.json`
+- Track C passing control audit: `cloud/runs/cloud-20260715T230356Z/audit.json`
+- Track C refreshed billing UI receipt: `cloud/ui-verification.json`
+- Rerun history: `failures.jsonl` and `deviations.json`
+
+No model API was called. RocketRide Cloud was called only in Track C under the
+redeemed promotion. Paid model cost and Cloud checkout charge are both USD 0;
+the passing control consumed 15.3 promotional compute tokens. Raw credentials,
+task tokens, account identifiers, and local environment files are not included.

benchmark-evidence/LIMITATIONS.md

@@ -0,0 +1,50 @@
+# Limitations
+
+- No RocketRide team acceptance, official publication, or external scorer receipt
+  exists. Every result is independent and unsubmitted.
+- The primary Linux host is Ubuntu 22.04 under WSL2 on x86-64, not the Apple
+  Silicon host used for upstream committed results. Timing remains host-dependent.
+- The Windows appendix is incomplete: data isolation never completed its first
+  repetition within the unchanged retry and timeout contract.
+- Track B covers five deterministic fixtures with frozen candidate artifacts. It
+  measures binding, scheduling, failure isolation, and admission behavior, not
+  model quality or framework-specific artifact generation.
+- Broader cancellation, checkpoint/resume, duplicate callback, and malformed
+  candidate scenarios remain outside the scored V1 matrix even where app verifier
+  tests cover related invariants.
+- Track B V1's `totalMs` includes RocketRide pool warmup and remains protocol
+  blocked. V2 is the separately pre-registered always-warm study and cannot
+  replace that unfavorable result.
+- V2 excludes RocketRide engine and pool readiness from the request clock while
+  retaining those values as standing and recovery metrics. Native and LangChain
+  retain subprocess startup in their totals, so cross-framework speed rankings
+  are not symmetric and are not claimed.
+- V2 pool warmup remained substantial: 6,976.188-11,940.908 ms across the 20
+  provisioned pools. Meeting the request deadline therefore does not erase
+  deployment and failed-worker replacement cost.
+- Native controls are additive deterministic workers, not a claim about every
+  production deployment topology in the four applications.
+- Application validation, version checks, CAS, proposals, and review are
+  application guarantees. No framework receives credit for them, and Track B
+  attempted no final writes.
+- Local compute and engineer time are not monetized. Only model and cloud API spend
+  are stated as USD 0.
+- RocketRide Cloud was run only for the separately pre-registered deterministic
+  Track C control. It used no model provider and cannot replace the pinned local
+  baseline or supply a LangChain comparison.
+- The exact upstream Cloud pipeline remains blocked because its custom
+  `workload` service is absent from the hosted catalog. Cloud validation did not
+  expose that incompatibility before execution.
+- Track C injects failure by terminating a hosted task through the API. It does
+  not prove isolation from an operating-system-level hard crash inside a hosted
+  worker.
+- Hosted task startup was 25.867-44.476 seconds in Track C. Request latency
+  excludes that standing cost but reports it separately.
+- The Starter promotion is 100 percent off for the checkout period only. Renewal
+  was disabled and verified with `cancelAtPeriodEnd=true`; future account state
+  remains external to this repository.
+- Physical and non-blank line counts are descriptive authoring evidence, not a
+  semantic complexity score.
+- Failed provisioning and harness attempts are preserved. Later successful runs
+  do not erase the missing-provider, unbounded-send, or connection-lifecycle
+  failures recorded in `failures.jsonl` and `deviations.json`.

benchmark-evidence/RESULTS.md

@@ -0,0 +1,147 @@
+# Results
+
+## Claim Status
+
+- **Track A:** evidence complete, independent unsubmitted reproduction.
+- **Track B V1:** evidence complete, orchestration gate passed, protocol
+  admission failed, promotion blocked.
+- **Track B V2:** evidence complete, orchestration and protocol admission
+  passed, eligible for external submission.
+- **Track C:** Cloud-native operational control passed and independently
+  audited; exact upstream Cloud execution blocked by missing `workload` service.
+- **Official status:** no external RocketRide acceptance or publication receipt.
+- **Paid spend:** USD 0 model cost and USD 0 cloud cost.
+
+## Track A: Supported Linux
+
+The unchanged upstream suite completed at commit
+`43be41acb58558dfae8e2e3deb86d8a00cb1b1c8` with no aggregate gate issues.
+
+| Dimension | RocketRide | LangChain control |
+|---|---|---|
+| Hard-failure isolation, N=10 | Held 10/10 | Lost all work 10/10 |
+| Concurrent work, pool 8 | 10/10 clean; wall median 1.285 s, p95 1.548 s | `batch_shared` crashed; `abatch` median 7.522 s; sequential median 7.530 s |
+| Concurrent work, pool 16 | 10/10 clean; wall median 0.867 s, p95 1.040 s | Same controls as above |
+| Data isolation, N=10 | 10/10 clean; 0 lost, duplicated, or leaked; wall median 5.508 s | Not a direct equivalent result in this cell |
+| Warm deployment cost | Pool 8 median 9.160 s; pool 16 median 15.590 s; data median 30.085 s | Reported separately from measured work |
+| Authoring descriptor | 0 imperative lines in the authored pipe | 14-17 imperative lines; 5 decision points for the correct idiom |
+
+The thread-count appendix emitted a median 31 node errors and is not used as a
+headline result.
+
+## Track A: Windows Appendix
+
+The unchanged Windows path is evidence-incomplete. Fault and concurrency cells
+completed, but data-isolation repetition 1 timed out on all five unchanged
+360-second attempts.
+
+- RocketRide pool 8 was clean 10/10 but had wall median 45.514 s versus
+  LangChain `abatch` 7.292 s.
+- RocketRide pool 16 was clean 0/10, retained 152 rows where 80 were expected,
+  and had wall median 60.821 s. Open SQLite cleanup on Windows is a recorded
+  inference, not a patched or promoted result.
+- This appendix cannot replace the supported Linux reproduction.
+
+## Track B V1: Node Applications
+
+The app verifier passed all four immutable adapter commits, base ancestry,
+canonical fixture parity, application tests, and shared protocol hash
+`b9b8b9f293c486fcd68f532f177e40f86c07c9310f31a64aa316ba469c1b1d59`.
+
+The completed matrix contains five fixtures, two variants, three repetitions,
+three frameworks, 30 requests, 90 result envelopes, 30 RocketRide traces, and
+30 aggregate rows.
+
+| Dimension | Native | RocketRide | LangChain |
+|---|---|---|---|
+| Normal correctness | 4/4 units in all 15 runs | 4/4 units in all 15 runs | 4/4 units in all 15 runs |
+| Injected hard failure | 3 unaffected units in all 15 runs | 3 unaffected units in all 15 runs; engine healthy after every run | Shared interpreter completed 0/4 in all 15 runs |
+| Normal total median range | 190.836-203.875 ms | 25,966.952-31,814.868 ms | 1,220.388-1,390.712 ms |
+| Hard-failure total median range | 176.464-222.259 ms | 31,934.275-37,909.386 ms | 1,057.141-1,139.453 ms before process loss was recorded |
+| Candidate digest | Frozen candidate parity held | Frozen candidate parity held | Frozen candidate parity held |
+| Model/cloud cost | USD 0 / USD 0 | USD 0 / USD 0 | USD 0 / USD 0 |
+
+### Promotion Blocker
+
+The original scorecard's completion/isolation gate reports `passed`. The later
+audit mirrors the shared app protocol and supersedes that status for promotion:
+all 30 RocketRide envelopes had `totalMs > deadlineMs` (10,000 ms). They would be
+rejected before candidate admission. The audit therefore reports:
+
+```text
+evidenceStatus: complete
+protocolAdmissionStatus: failed
+promotionStatus: blocked
+officialStatus: separate_application_study_unsubmitted
+```
+
+This is still a useful integration result: RocketRide isolated worker failure,
+but the tested local resident-pool lifecycle is too slow for the pre-registered
+application contract. The appropriate next experiment is a newly pre-registered
+warm-service topology or deadline budget study, not retroactively changing this
+run.
+
+## Track B V2: Resident Runtime
+
+V2 was separately pre-registered and committed before implementation. It keeps
+the five frozen candidates and 10,000 ms request deadline, adds a fixed 2,000 ms
+per-unit failure-detection budget, and binds every request and result to one
+resolved-definition digest containing the merged app commit plus protocol,
+adapter, domain-tool, validator, and fixture hashes.
+
+The completed V2 matrix again contains 30 requests, 90 results, 30 RocketRide
+traces, and 30 aggregate rows. The independent audit reports complete evidence,
+passed protocol admission, and `eligible_for_external_submission`.
+
+| Dimension | Native | RocketRide | LangChain |
+|---|---|---|---|
+| Normal correctness | 4/4 in all 15 runs | 4/4 in all 15 runs | 4/4 in all 15 runs |
+| Injected hard failure | 3 unaffected units in all 15 runs | 3 unaffected units in all 15 runs; engine healthy after every run | Shared interpreter completed 0/4 in all 15 runs |
+| Normal total median range | 136.618-156.006 ms | 1,043.073-1,090.228 ms | 790.117-897.225 ms |
+| Hard-failure total median range | 137.180-161.455 ms | 3,008.731-3,025.511 ms | 733.217-930.532 ms before process loss was recorded |
+| Deadline overruns | 0/30 | 0/30 | 0/30 |
+| Model/cloud API cost | USD 0 / USD 0 | USD 0 / USD 0 | USD 0 / USD 0 |
+
+RocketRide standing cost remains material and is not hidden: engine cold start
+was 1,569.703 ms, normal pool warmup median was 7,721.659 ms, and replacement
+pool warmup after destructive fault median was 10,109.896 ms. Native and
+LangChain totals conservatively retain subprocess startup while RocketRide V2
+uses the pre-registered prewarmed request clock, so V2 supports protocol
+admission and fault-isolation claims, not a raw speed-winner claim.
+
+All four adapter PRs are merged, their post-merge checks passed, and the exact
+main revisions are deployed to production. Deployment IDs, immutable URLs,
+aliases, and HTTP checks are recorded in
+`extensions/node-suite/production-deployment-v2.json`.
+
+## Track C: RocketRide Cloud
+
+Track C was pre-registered before its measured hosted run. It uses the official
+Cloud endpoint and a deterministic built-in `webhook -> parse -> response_text`
+pipeline with no model provider. Each of three repetitions created four
+resident tasks, warmed them, processed 16 unique documents, terminated task 0,
+and then sent one unique failure-phase request to every task.
+
+| Check | Audited result |
+|---|---:|
+| Normal requests correct | 48/48 |
+| Unaffected failure-phase requests correct | 9/9 |
+| Terminated-task requests failed as expected | 3/3 |
+| Cross-task leaks | 0 |
+| Successful termination calls | 12/12 |
+| Normal response latency | 827.990 ms median; 1,109.528 ms p95 |
+| Unaffected failure response latency | 520.824 ms median; 844.998 ms p95 |
+| Hosted task start time | 25.867-44.476 s; 35.319 s median |
+| Warm-up response time | 1.539-2.414 s; 1.920 s median |
+
+Hosted startup is therefore material and is not hidden by the request-latency
+figures. The unchanged upstream fault-isolation `.pipe` was also attempted and
+failed after 7.823 seconds with `The service workload was not found`. Hosted
+validation had accepted the pipeline during setup, so this is a catalog/runtime
+portability issue rather than a successful official benchmark cell.
+
+Billing readback after the passing run reports `JULY2026BENCHMARK` valid at 100
+percent off, zero cents due, Starter active, and `cancelAtPeriodEnd=true`. The
+passing control consumed 15.3 promotional compute tokens, leaving 9,969.3. Its
+audit reports complete evidence and passed control admission, while retaining
+`cloud_operational_appendix_unsubmitted` as the official status.

benchmark-evidence/baseline/application-shas.json

@@ -0,0 +1,9 @@
+{
+  "capturedAt": "2026-07-15T09:26:22.398Z",
+  "repositories": {
+    "HomenShum/NodeRoom": "ca25e347dc467bc37f06918e1a18656f7336ee28",
+    "HomenShum/NodeBenchAI": "6ed0a58eeda993ff2a937ea4bacc2856756dd521",
+    "HomenShum/NodeSlide": "dd67e4c642c40e6bb414af617a67a31dbed507c5",
+    "HomenShum/NodeVideo": "bb79bc385de93c90cee89b160fc801d18372d89e"
+  }
+}

benchmark-evidence/baseline/concurrent-work/aggregate/SUMMARY.md

@@ -0,0 +1,30 @@
+# Independent Windows Baseline
+
+Evidence status: **evidence_incomplete**. External status: **independent_unsubmitted_reproduction**.
+This is an unchanged local reproduction, not a RocketRide-accepted official result.
+
+## Completeness
+
+| Benchmark | Expected | Observed | Complete |
+|---|---:|---:|---|
+| fault-isolation | 10 | 10 | true |
+| concurrent-processing | 10 | 10 | true |
+| data-isolation | 10 | 0 | false |
+| authoring-effort | 1 | 1 | true |
+
+## Dimensional Results
+
+- Fault isolation: RocketRide held in 10/10 runs; the in-process LangChain probe lost all work in 10/10 runs.
+- RocketRide pool M=8: warm execution median 45.514 s, p95 80.232 s; warm-up median 26.775 s; clean 10/10.
+- RocketRide pool M=16: warm execution median 60.821 s, p95 81.444 s; warm-up median 63.655 s; clean 0/10.
+- LangChain batch_shared: statuses crash; wall median 0.042 s, p95 0.056 s.
+- LangChain abatch_blocking: statuses ok; wall median 7.292 s, p95 7.344 s.
+- LangChain seq: statuses ok; wall median 7.435 s, p95 7.612 s.
+- Data isolation: RocketRide clean 0/0; lost-doc median None; duplicated/leaked median None.
+- Preserved retry signals: {"concurrent-processing": 9, "data-isolation": 5}.
+
+Cold start, warm-up, and warm execution remain separate. Resource fields are reported only where the upstream result emits them.
+
+## Evidence Gaps
+
+- data-isolation: expected 10 results, observed 0