ci(ci): run the e2e journeys against pnpm dev on linux

[systemfsoftware-maker]

Lake 1, U8, stacked on #41. Trimmed per Kiro's 2026-10-06 ruling: the journeys run against pnpm dev, on Linux, with no local stack.

• bin/journeys starts pnpm --filter @endgame/site dev in its own process group, waits for http://localhost:1337/, runs the site-e2e journeys against it, stops dev with one SIGINT (so alchemy dev stops its sidecar instead of orphaning it), and prints the last 60 lines of the dev log when a run is red. Root script: pnpm journeys.
• alchemy.run.ts pins the dev port (dev: { port: 1337, strictPort: true }): a busy port fails the run instead of serving on 1338 while the journeys test 1337.
• ci.yml gains a journeys job on ubuntu-latest that uploads apps/site-e2e/artifacts/ and .journeys/dev.log.
• AGENTS.md gains START-5: the e2e journeys pass against pnpm dev (gate: pnpm journeys).

Removed by the trim: the process-compose local stack (#38, closed), bin/stack-ci, stack:ci, and the OpenTelemetry collector, Tempo and Grafana.

Gate at e917151

$ pnpm check:ci
checkci=0
$ pnpm journeys
journeys=0

Two consecutive pnpm journeys runs both passed and left no listener on 1337-1340.

Cycle 35 (Kiro rulings on starter-verify's review)

• F6 (3f3a504, cycle 35): added a busy-port refusal to bin/journeys. Deleted in cycle 61 (below).

Linux only (Kiro, cycle 52)

• 9bb82c7: the journeys job runs on ubuntu-latest only, and the title drops "and macos". This edits ci.yml, an evaluator surface, at Kiro's direction (CONST-W3); Kiro owns the instrument.

Gate at 71841ad, clean worktree, Linux:

$ pnpm install --frozen-lockfile    # exit 0
$ pnpm check:ci                     # exit 0: 12/12 turbo tasks, dist 2/2
$ pnpm journeys                     # exit 0
Tests  2 passed (2)

CI run 37573358625 on 71841ad: all 6 jobs pass (the five check legs and journeys).

Cycle 61 (conductor rulings on starter-verify's c59 re-check)

• F4 (7767212): the busy-port guard is gone, with no replacement. Inside the sandbox's own network namespace (from build(repo): run all dependency code in the sandbox, systemfsoftware packages from Nix #52 up) it could never fire, and that isolation already keeps a stale host server from answering the journeys.
• F6 (4e8a46c): turbo.json's test task logs with outputLogs: "new-only", so CI logs show which tests ran.

[systemfsoftware-maker]

Fix 15d131d: stack:ci moves to bin/stack-ci, which starts process-compose with --keep-project (the first hosted run lost the server when Tempo died, so down got connection refused) and on any failure writes local-stack/data/stack-report.md: process list -o wide plus the last 40 log lines per process. The job appends it to the step summary and uploads it; a separate step realises the dev shell first so the 300 s readiness budget measures startup (the cold fetch took 21 s; there is no repo Nix cache action, so none was added). Sabotage (Tempo marker dir under /proc): report shows tempo Completed exit 1, otel-collector/grafana/site Skipped, and the live-store mkdir error. CI run 37407682838: 7/7 green, stack 92 s.

[kiro-systemf]

Conductor verdict: 8/8 green on a1aee1d incl. changeset check (strategy: changesets) and journeys; tsgo 0.51.0; 0 threads; hunt clean.