docs(quickstart): correct the commands that no longer work, and lead GitHub with demo-requests

[behinddwalls]

Summary

Why?

The quickstart was audited against a live stack on all three rungs, including a real GitHub run. Most of it holds. Four things did not.

Three commands named a compose project that no longer exists. Folding local-provider-* into local-submitqueue-* left one project, submitqueue, but the troubleshooting and log-tailing commands still said submitqueue-provider. Every one of them fails: docker logs submitqueue-provider-mysql-queue-1 finds no such container, and a down -v against that project silently does nothing — which is worse, because it is offered as the fix for a stack that will not start, and appears to work.

The GitHub rung led with the one command that is not the quickstart. Its first instruction was to open a pull request by hand and land it with make land PR=…, while make demo-requests — the command the other two rungs use, which opens the pull requests for you — was a sentence above it. The rung that needs the most setup was the one asking for the most manual work.

make local-submitqueue-stop claimed to preserve data it does not. Both MySQL services mount anonymous volumes, so down detaches them and the next up gets empty databases. The quickstart says so; the Makefile printed "Data volumes preserved", which reads as "your requests are still there". They are not.

A failing change was documented without saying which rung it belongs to. The ?sq-fake=build-fail URI is a git:// one pointing at nothing, which only resolves where the change provider is fake. It sits after the GitHub section, so it reads as applying there, where it fails for an unrelated reason.

What?

The stale project name is corrected in all three places, and the down -v recipe becomes make local-submitqueue-clean, which is the same operation with the right project and overlay. PROVIDER_LOCAL_PROJECT is deleted from the Makefile — its last use went with local-provider-*, and a variable that still defines the wrong answer is how the wrong answer gets copied again.

The GitHub rung now leads with make demo-requests, with real output from a scratch-repo run, and keeps make land PR=… below it under a heading that says what it is for — landing a pull request you opened yourself. Nothing was cut; the stack case moves to make demo-requests STACKED=true, which is what the other rungs already use for it.

It also now says plainly that your CI does not run these builds: the build runner is fake on the GitHub rung too, so landed there does not mean anything was tested. That was in the "Using real CI" section further down, which is too late for a reader who has just watched three pull requests merge.

The stop message says what actually happens, and the failing-change section says it belongs to the fake rung.

Test Plan

Run against a live stack, in this order:

• ✅ PROVIDER=github end to end for the first time, against a scratch repository: make demo-requests COUNT=3 opened pull requests 522–524, landed all three, and the API confirms "merged": true on each with their commits on main — so the rung the documentation described but nobody had run, works
• ✅ that run is what the new output block quotes, rather than an invented one
• ✅ the fake build runner claim is now checked two ways: demo-queue inherits buildRunner: {type: fake} in the GitHub profile, and the repository's newest workflow_dispatch run predates the run by six days. Zero check-runs on the landed head confirms nothing else fired either
• ✅ the failing change reaches error, as documented: submitted the exact URI from the doc and read status: error back
• ✅ make local-submitqueue-stop prints the corrected message
• ✅ make lint, make gazelle

Checked and found accurate, so left alone: every make demo-requests option (COUNT, FOLDERS, FILES, CONCURRENCY, STACKED, LAND) against the recipe's flags; land, land-status, land-list, land-watch and their variables; all six sq-fake tokens against the fakes that read them; SQ_TOKEN and -token-env; QUEUE_LOG_LEVEL; the sandbox path; and the anonymous-volume behaviour the clean-up section describes.

make land-list against the failed request also confirms the documented difference between list and watch — the row showed error alone, with no trail, because list does not fetch a history per request.