Scope session-close background-process check to the current session

[bborbe]

/vault-cli:session-close Phase 5 scanned the machine-wide process table (ps aux | grep -E '(dark-factory|daemon|watch)') and flagged anything it found. In practice that surfaces sibling Claude sessions' daemons as "outstanding before close" — a false positive that makes the verdict untrustworthy and asks the user to clean up work that isn't this session's.

Hit live this session: a dark-factory daemon started by a different session (different shell snapshot) was reported as an open item on a session that never touched dark-factory.

Changes

• Phase 5 now builds its candidate list from the conversation — Bash calls made with run_in_background: true, Monitor watches armed this session — instead of a process-table scan. Empty list → phase skipped silently, no ps aux at all. ps -p <pid> is used only to confirm liveness of PIDs already on the list. Explicit rule added: never flag a process this session didn't spawn; each session cleans up its own.
• Phase 6 gets the matching scoping note — "in scope" means a repo this session touched, so another session's dark-factory project is skipped even when its daemon is running.
• Phase 9 example output no longer models the bug (the sample line now says the daemon was spawned by this session).

Docs-only change to the command definition; make precommit green.

[ben-s-pull-request-reviewer]

The diff is purely documentation (CHANGELOG.md + commands/session-close.md). Let me now do the in-session classification and adjudication.

Step 4b-i — Candidate rules (triggers matching CHANGELOG.md or commands/session-close.md):

Rule ID	Owner	Trigger match
changelog/conventional-prefix-required	agent-auditor	CHANGELOG.md ✓
changelog/unreleased-entry-required	agent-auditor	(script, no glob)
changelog/preamble-frozen	agent-auditor	(script, no glob)
agent-cmd/command-frontmatter	slash-command-auditor	commands/**/*.md ✓
agent-cmd/command-thin	slash-command-auditor	commands/**/*.md ✓
agent-cmd/gap-driven-feedback	agent-auditor	commands/**/*.md ✓
agent-cmd/no-user-prompts	agent-auditor	commands/**/*.md ✓
agent-cmd/scripts-in-claude-dir	agent-auditor	commands/**/*.md ✓
agent-cmd/single-source-of-truth	agent-auditor	commands/**/*.md ✓

Step 4c-sel — Classify:

• changelog/conventional-prefix-required → applicable (CHANGELOG has ## Unreleased with fix: prefix — correct ✓)
• agent-cmd/command-frontmatter → skipped — existing file, not created; rule triggers on creation
• agent-cmd/command-thin → skipped — change clarifies behavior, doesn't make command thin; "thin" = sparse/underspecified, this makes it more precise
• agent-cmd/gap-driven-feedback → skipped — existing command receiving a bug fix, not a new capability introducing a gap
• agent-cmd/no-user-prompts → skipped — purely docs; no behavioral prompts introduced
• agent-cmd/scripts-in-claude-dir → skipped — purely docs; no script dependencies
• agent-cmd/single-source-of-truth → applicable — correctly identifies the conversation's run_in_background/Monitor calls as the authoritative source; documents this explicitly

Step 4d-sel — Adjudicate:

1. changelog/conventional-prefix-required: CHANGELOG has ## Unreleased with - fix: session-close Phase 5... — fix: is a recognised prefix. PASS.
2. agent-cmd/single-source-of-truth: The doc now correctly states candidates come from the conversation's run_in_background/Monitor calls, not a machine-wide ps aux scan. This is an accurate and well-scoped single source of truth. PASS.

Both applicable rules pass. No violations.

Step 3a — LICENSE: Present at repo root. ✓

---

{
  "verdict": "approve",
  "summary": "Documentation-only PR clarifying session-close Phase 5 behavior: candidates now come from the conversation's own background-process tracking, not a machine-wide process scan. CHANGELOG entry uses correct `fix:` prefix. No violations of applicable judgment rules.",
  "comments": [],
  "concerns_addressed": []
}