[roodboi]

Refs HACK-1425. Follow-ups from the verifier reviews of #447 (merged as 686b6b8) and #449 (merged as 4c9eb3e). None of them blocked those PRs.

Root cause

1. No test pinned the production drain bound. Every drain test read NATIVE_EXIT_DRAIN_BOUND_MS itself, or injected a wide value through the exitDrainBoundMs test seam. The held-descendant cases only checked an upper bound (settled - exited < bound + margin). So setting the constant to 250 (or changing the ?? NATIVE_EXIT_DRAIN_BOUND_MS default to ?? 250) passed the whole suite.
2. Three bound drains had no late-writer case. The one-off drain and the two recovery drains (selection and cleanup) use the same code path as status, read, logs and exec. Only those four had a "closed code written after the reported exit" case.
3. Three wiring mutations in src/commands/project.ts survived the tests. tests/logs-loki-empty-fallback-cli.test.ts never passed --pretty, --query or --until through the CLI. Replacing pretty: args.options.pretty === true, or the selectors' query: opts.queryOpt or until: opts.until, with a constant did not fail any test.
4. The hack global logs-reset text was out of date. The command runs down -v, which (since fix(logs): persist Alloy positions, release SIGINT handlers and carry every selector in Loki hints #449) also deletes the alloy-data volume that holds Alloy's read positions. The confirm prompt and the CLI summary named only Loki and Grafana.

Fix

• tests/native-runtime-client.test.ts
  • New case the production exit drain bound keeps its documented floor and outlasts a late writer. It checks that NATIVE_EXIT_DRAIN_BOUND_MS >= 5000 (the documented floor, now written in the constant's doc comment) and that the bound is larger than the late-writer delay plus a clock tolerance.
  • The held-descendant exit cases (read, selection, cleanup) do not pass the seam. They now also check settled - exited >= NATIVE_EXIT_DRAIN_BOUND_MS - 50 ms, which proves the production default is the bound in effect.
  • LATE_WRITER_SELECTIONS gains one-off, selection and cleanup.
• tests/native-project-process.test.ts: the same lower bound on the foreground graph owner's two held-descendant cases. This covers that file's own ?? NATIVE_EXIT_DRAIN_BOUND_MS default.
• tests/logs-loki-empty-fallback-cli.test.ts gets two CLI cases. Both use the stub docker, the fake Loki, and a private HOME via privateCliEnv:
  • --loki --no-follow --pretty migrate: the hint repeats --pretty.
  • --no-follow --query '<custom LogQL>' --until 1m: Loki receives the custom query exactly as given, and the hint lists --query '…' and --until 1m as selectors that --compose cannot apply.
• src/commands/global.ts: the summary is now "Wipe Loki, Grafana and Alloy volumes (fresh logs, dashboards and Alloy read positions)". The prompt now says it deletes "ALL Loki logs, Grafana state and Alloy read positions". docs/reference/cli.md was regenerated with bun run docs:cli-reference. docs/architecture.md already described the volume removal, so it is unchanged.

There is no production behavior change apart from the logs-reset text and a doc comment.

Evidence

Mutation runs (each mutation applied on its own and restored from a backup afterward):

Mutation	Killed by
NATIVE_EXIT_DRAIN_BOUND_MS = 250	documented-floor case
runtime client selected ?? 250	read, selection and cleanup held-descendant exit cases
graph owner opts.exitDrainBoundMs ?? 250	ready and early failure held-descendant cases
fixed 250 ms window for one-off only	authored one-off keeps a closed code written after the reported exit
fixed 250 ms window for recovery-selection only	authored selection keeps a closed code …
fixed 250 ms window for recovery-cleanup only	authored cleanup keeps a closed code …
pretty: args.options.pretty === true → false	explicit --loki snapshot hint repeats --pretty
selectors query: opts.queryOpt → undefined	explicit --query snapshot hint names --query and --until
selectors until: opts.until → undefined	explicit --query snapshot hint names --query and --until

Local gates at this head (Bun 1.4.2):

• bun run check passed, including privacy:check.
• bun run --cwd packages/cli typecheck and root bun run typecheck reported 0 errors.
• Biome reported no issues on the changed files.
• gitleaks 8.30.1 on the origin/next..HEAD patch found no leaks.
• Focused tests: native-runtime-client, native-project-process, logs-loki-empty-fallback-cli, global-command and prerequisites-matrix gave 74 pass, 0 fail.

The full-suite evidence is hosted CI's sharded test job at this PR head.

Gaps

• The new lower-bound checks rely on the wall clock of two processes and a 50 ms tolerance. The lower direction does not depend on host load, because a timer cannot fire before its delay, so these checks should not add flakiness.
• The new prompt text has no test. confirmSafe declines in non-interactive runs (initialValue: false), so a CLI test cannot see the prompt without a TTY harness. This matches the coverage before this PR.
• The pin is a floor (>=), not an exact value. Raising the bound still passes, which is intended.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[linear-code]

HACK-1425