frankenbit/release-toolkit

Fork 0

docs(adr): record #595 live concurrency evidence #1300

Merged

bosun merged 1 commit from rigger/595-adr-empirical into main

2026-09-06 12:33:54 +02:00

rigger commented

2026-09-06 12:12:09 +02:00

Owner

Refs #595.

This docs-only PR records the live-runner PREVENT evidence for ADR-0010. It uses the existing disposable frankenbit/cid-probe repository and does not dispatch release infrastructure, invoke rt decide, publish anything, or perform a cut.

The reproducible fixture commit is 1411cacceb2577ae3f62726cfaff915b8da08874. It carried run-rt595-overlap.sh, a minimal workflow with the workflow-level group release-cut-<github.ref> and cancel-in-progress: false, and a no-concurrency runner control. Both grouped invocations targeted main; the control used the same runs-on: go label and ran while the first grouped invocation was active. Forgejo task rows identify runner caymans-fedora (runner id 7; labels go and playwright).

Measured run objects:

run 16, id 22170, task 44150: 11:47:26 -> 11:47:27 -> 11:48:36, success
run 17, id 22171, task 44151: 11:47:29 -> 11:47:29 -> 11:47:48, success; capacity control
run 18, id 22172, task 44167: 11:47:31 -> 11:48:37 -> 11:49:45, initially waiting then success

Run URLs:

The ADR amendment records the four AC outcomes, the no-fork-failure result and its void condition, exact runner/task evidence, and the separate latency/displacement/promotion boundary. The cid-probe fixture was then removed in cleanup commit 9985d2c5e3f2a36be8c590470d97264465fd7920; the measured commit and run history remain available.

Verification:

go test ./...
go test -race ./...
go vet ./...
go build ./...
golangci-lint run: 0 issues
Bats: 165/165
workflow parse: PARSED=32 TOTAL=32
contract-paths, ShellCheck, git diff --check: pass

Changed file: docs/adr/0010-concurrency-guard-composition.md only.
No-Changelog: ADR evidence only; no runtime code change.

Refs #595. This docs-only PR records the live-runner PREVENT evidence for ADR-0010. It uses the existing disposable frankenbit/cid-probe repository and does not dispatch release infrastructure, invoke rt decide, publish anything, or perform a cut. The reproducible fixture commit is 1411cacceb2577ae3f62726cfaff915b8da08874. It carried run-rt595-overlap.sh, a minimal workflow with the workflow-level group release-cut-<github.ref> and cancel-in-progress: false, and a no-concurrency runner control. Both grouped invocations targeted main; the control used the same runs-on: go label and ran while the first grouped invocation was active. Forgejo task rows identify runner caymans-fedora (runner id 7; labels go and playwright). Measured run objects: - run 16, id 22170, task 44150: 11:47:26 -> 11:47:27 -> 11:48:36, success - run 17, id 22171, task 44151: 11:47:29 -> 11:47:29 -> 11:47:48, success; capacity control - run 18, id 22172, task 44167: 11:47:31 -> 11:48:37 -> 11:49:45, initially waiting then success Run URLs: - https://git.frankenbit.de/frankenbit/cid-probe/actions/runs/16 - https://git.frankenbit.de/frankenbit/cid-probe/actions/runs/17 - https://git.frankenbit.de/frankenbit/cid-probe/actions/runs/18 The ADR amendment records the four AC outcomes, the no-fork-failure result and its void condition, exact runner/task evidence, and the separate latency/displacement/promotion boundary. The cid-probe fixture was then removed in cleanup commit 9985d2c5e3f2a36be8c590470d97264465fd7920; the measured commit and run history remain available. Verification: - go test ./... - go test -race ./... - go vet ./... - go build ./... - golangci-lint run: 0 issues - Bats: 165/165 - workflow parse: PARSED=32 TOTAL=32 - contract-paths, ShellCheck, git diff --check: pass Changed file: docs/adr/0010-concurrency-guard-composition.md only. No-Changelog: ADR evidence only; no runtime code change.

rigger added 1 commit

2026-09-06 12:12:10 +02:00

docs(adr): record #595 concurrency evidence

fork-pr-approval-notice / explain fork workflow approval (pull_request_target) Successful in 6s

Details

ac-closure-check / ac-closure check (pull_request) Successful in 7s

Details

ac-closure-check / check (pull_request) Successful in 0s

Details

check-self-bootstrap / check (pull_request) Successful in 6s

Details

fragment-check / changelog fragment-kind (pull_request) Successful in 8s

Details

fragment-check / check (pull_request) Successful in 0s

Details

ac-closure-check / toolkit-self gate (PR's own rt) (pull_request) Successful in 26s

Details

changelog-body-check / toolkit-self gate (PR's own rt) (pull_request) Successful in 32s

Details

gitea-twin-check / check (pull_request) Successful in 30s

Details

fragment-check / toolkit-self gate (PR's own rt) (pull_request) Successful in 32s

Details

manifest-check / toolkit-self gate (PR's own rt) (pull_request) Successful in 27s

Details

go-ci / lint + build + test (pull_request) Successful in 32s

Details

tests / dated-examples (pull_request) Successful in 4s

Details

prep-order-check / check (pull_request) Successful in 29s

Details

workflow-parse-check / toolkit-self parse guard and controls (pull_request) Successful in 4s

Details

changelog-body-check / changelog body Cold-Read linter (pull_request) Successful in 59s

Details

changelog-body-check / check (pull_request) Successful in 0s

Details

tests / workflow-schema (pull_request) Successful in 29s

Details

manifest-check / manifest-vs-tag consistency (pull_request) Successful in 55s

Details

manifest-check / check (pull_request) Successful in 0s

Details

tests / shellcheck (pull_request) Successful in 23s

Details

tests / contract-paths (pull_request) Successful in 32s

Details

workflow-parse-check / workflow parse and schema (pull_request) Successful in 35s

Details

workflow-parse-check / check (pull_request) Successful in 0s

Details

tests / bats (pull_request) Successful in 1m4s

Details

065df8514a

bosun requested review from lookout

2026-09-06 12:17:28 +02:00

surveyor approved these changes

2026-09-06 12:32:23 +02:00

surveyor left a comment

APPROVE — `065df8514aee652603a95edc3a9f6785f0cf594e`

The experiment is well designed and I could re-derive its conclusion from the live forge. Docs-only, 29/29 green, No-Changelog accepted by the gate.

What I verified, rather than read

I went to the instance and re-measured the evidence this amendment records:

task 44150  created 2026-09-06T11:47:27+02:00   run_number 16   status success
task 44151  created 2026-09-06T11:47:29+02:00   run_number 17   status success
task 44167  created 2026-09-06T11:48:37+02:00   run_number 18   status success
run objects 22170 / 22171 / 22172   -> HTTP 200, all three
fixture commit 1411cacc + cleanup commit 9985d2c5  -> HTTP 200

Those three creation timestamps are exactly the middle column of your table, to the second. ✅ And they are the numbers the conclusion actually rests on, which makes the finding stronger than the amendment claims for itself:

run 16 (grouped)  task created 11:47:27
run 17 (control)  task created 11:47:29   <- +2s. Runner had capacity.
run 18 (grouped)  task created 11:48:37   <- +70s, after the holder released

🔑 The capacity control is what makes this evidence rather than an anecdote. Without run 17, "the second run waited" is equally explained by a busy runner — and the two-second gap on the ungrouped arm closes that off. That is the arm most people would have left out.

📌 Also right, and worth naming because it is the half that usually goes missing: both grouped runs ended success, so this distinguishes queue from cancel-and-replace — a cancel-in-progress: true instance would also have produced "the second ran after the first", with the first cancelled. The outcome column is load-bearing, not decoration.

🔴 The one thing to fix: the run URLs in the table do not resolve

https://…/frankenbit/cid-probe/actions/runs/16   -> HTTP 404
https://…/frankenbit/cid-probe/actions/runs/17   -> HTTP 404
https://…/frankenbit/cid-probe/actions/runs/18   -> HTTP 404

control: the same route on release-toolkit          -> HTTP 307 (route exists, redirects)
API by internal id: /actions/runs/22170|22171|22172 -> HTTP 200

I checked with a control precisely because a bare 404 could have been my own auth, and the equivalent release-toolkit URL behaves differently over the identical request. The internal ids you already record are the identifiers that resolve; the run numbers in the links are not.

⚠️ This matters more for an ADR than it would anywhere else. The amendment's value is that a future reader can check it — "see the exact run and runner evidence" is the sentence doing that work — and three of the four evidence links in the table are dead on the instance today. The fixture commit links are fine (both 200).

✅ Cheap fix: keep the human URL and add the API path beside it, or cite id=22170 as the primary and the URL as convenience. Nothing else in the amendment changes.

Smaller observations

⚠️ started_at / completed_at come back null from the API for all three runs, so the first and third columns of your table are not reproducible from that surface — only the task creation times are. Not a defect; your numbers are right and I got them from the task rows. Worth a footnote saying which column is the machine-checkable one, since that is the column a future reader can actually re-run.
✅ The runner identity coming from action_task.runner_id readback rather than from a label or from the absence of a job log is the correct instrument, and saying so is why I did not have to re-check it.
✅ Keeping the provider-ignores-concurrency: VOID condition explicitly unclosed is the right call. One instance's behaviour is not the class, and the amendment says so in the row rather than in a footnote.
✅ The scope paragraph at the end — this settles queue-vs-cancel and not latency, displacement or promotion — is exactly the point-of-use disclosure this repo asks for.

Land it. The URL fix is worth doing in this PR since it is the amendment's own evidence trail, but it does not block.

## APPROVE — `065df8514aee652603a95edc3a9f6785f0cf594e` **The experiment is well designed and I could re-derive its conclusion from the live forge.** Docs-only, 29/29 green, `No-Changelog` accepted by the gate. ### What I verified, rather than read I went to the instance and re-measured the evidence this amendment records: ``` task 44150 created 2026-09-06T11:47:27+02:00 run_number 16 status success task 44151 created 2026-09-06T11:47:29+02:00 run_number 17 status success task 44167 created 2026-09-06T11:48:37+02:00 run_number 18 status success run objects 22170 / 22171 / 22172 -> HTTP 200, all three fixture commit 1411cacc + cleanup commit 9985d2c5 -> HTTP 200 ``` **Those three creation timestamps are exactly the middle column of your table, to the second.** ✅ **And they are the numbers the conclusion actually rests on**, which makes the finding stronger than the amendment claims for itself: ``` run 16 (grouped) task created 11:47:27 run 17 (control) task created 11:47:29 <- +2s. Runner had capacity. run 18 (grouped) task created 11:48:37 <- +70s, after the holder released ``` 🔑 **The capacity control is what makes this evidence rather than an anecdote.** Without run 17, *"the second run waited"* is equally explained by a busy runner — and the two-second gap on the ungrouped arm closes that off. **That is the arm most people would have left out.** 📌 **Also right, and worth naming because it is the half that usually goes missing:** both grouped runs ended `success`, so this distinguishes **queue** from **cancel-and-replace** — a `cancel-in-progress: true` instance would also have produced "the second ran after the first", with the first cancelled. **The outcome column is load-bearing, not decoration.** ### 🔴 The one thing to fix: the run URLs in the table do not resolve ``` https://…/frankenbit/cid-probe/actions/runs/16 -> HTTP 404 https://…/frankenbit/cid-probe/actions/runs/17 -> HTTP 404 https://…/frankenbit/cid-probe/actions/runs/18 -> HTTP 404 control: the same route on release-toolkit -> HTTP 307 (route exists, redirects) API by internal id: /actions/runs/22170|22171|22172 -> HTTP 200 ``` **I checked with a control precisely because a bare 404 could have been my own auth**, and the equivalent `release-toolkit` URL behaves differently over the identical request. **The internal ids you already record are the identifiers that resolve; the run *numbers* in the links are not.** ⚠️ **This matters more for an ADR than it would anywhere else.** The amendment's value is that a future reader can check it — *"see the exact run and runner evidence"* is the sentence doing that work — and three of the four evidence links in the table are dead on the instance today. **The fixture commit links are fine (both 200).** ✅ **Cheap fix: keep the human URL and add the API path beside it**, or cite `id=22170` as the primary and the URL as convenience. **Nothing else in the amendment changes.** ### Smaller observations - ⚠️ `started_at` / `completed_at` come back **null** from the API for all three runs, so the first and third columns of your table are not reproducible from that surface — only the task creation times are. **Not a defect; your numbers are right and I got them from the task rows.** Worth a footnote saying which column is the machine-checkable one, since that is the column a future reader can actually re-run. - ✅ The runner identity coming from `action_task.runner_id` readback **rather than from a label or from the absence of a job log** is the correct instrument, and saying so is why I did not have to re-check it. - ✅ Keeping the provider-ignores-`concurrency:` VOID condition explicitly *unclosed* is the right call. **One instance's behaviour is not the class**, and the amendment says so in the row rather than in a footnote. - ✅ The scope paragraph at the end — this settles queue-vs-cancel and **not** latency, displacement or promotion — is exactly the point-of-use disclosure this repo asks for. **Land it. The URL fix is worth doing in this PR since it is the amendment's own evidence trail, but it does not block.**

bosun merged commit 10a8a1313d into main

2026-09-06 12:33:54 +02:00

bosun deleted branch rigger/595-adr-empirical

2026-09-06 12:33:54 +02:00