docs(journald-migration): retract a runtime claim #1412 shipped, and close out the merged handoff

[ZacxDev]

Follow-up to #1412 (merged as f06b106f). Three corrections found after that merge, two of them to claims #1412 itself shipped.

1. A runtime claim #1412 shipped is unsupported — retracted

scripts/diagnose-nix-disk.sh's header said "THIS TAKES TENS OF MINUTES" and cited "section 3 alone had not finished after 13 minutes".

The run launched to verify that finally returned: 4h05m wall clock, having reached section 3 of 10. So the shipped figure understates it by roughly an order of magnitude — but it is a floor, not a runtime, for two reasons that both matter:

• the box was running the test gate throughout and the process sat at 3% CPU (95s user + 452s system across four hours) — I/O-starved, not working;
• the run did not stop of its own accord. It died on a syntax error at a line that is fine on disk (bash -n on the committed copy is clean; that line is a df -hT /). Bash reads a script by byte offset as it executes, and the file was edited five times during those four hours, shifting underneath the running interpreter. The artefact looked exactly like a defect in the script. The header now warns about it.

The honest state, and the header now says so: no clean full run has ever been observed, and the script's own closing all 10 sections attempted banner is untested end to end. The paragraph explicitly tells the next reader not to quote a runtime from it as if measured, and to replace it once someone completes a run on an idle box.

2. The merged handoff doc still described the PR as open

It said "PR #1412 OPEN" with a gate "IN FLIGHT" pointing at a superseded run. It now records the squash (verified by CONTENT — a squash is never an ancestor of its base), both tiers passing on the merged tree 53d8b962, and the five-round audit whose ladder stopped on a stated criterion.

3. The rename list was wrong, and the fix is deliberately NOT a script

Next-step 4 listed three deprecated renames; only two are live — services.dnsmasq.servers is commented out at configuration.nix:71. The list came from an eval warning that was true when captured; /etc/nixos changed underneath it, and a too-narrow first grep missed it.

The two live ones (gnome.tracker → tinysparql, gnome.tracker-miners → localsearch) get the exact edit written into the doc and no apply-*.sh script. They are warning-only (mkRenamedOptionModule). A second root-privileged /etc/nixos mutator would ship a copy of the rebuild-and-rollback trap that took five audit rounds to get right — three of its four blockers introduced by the previous round's own fix. RULES.md → "One rule, one place": a duplicated predicate is typically wrong at N−1 sites in the same direction. The doc says to extract a shared nix/system/lib/ helper first if a third such change ever appears. That is a judgement to overrule, not a gap to fill.

Gate: NOT GREEN, and not because of this change

🔴 Do not merge this on my say-so — the full pytest tier could not be obtained. Two consecutive runs hit the 3600s timeout and were Terminated (exit 124 / RESULT exit=143). That is load, not an assertion, by the documented discriminator:

• node tier PASS 1449/1449 on both runs;
• every pytest target that completed before the wall passed;
• load average went 72 → 92 across the two runs, with 257 concurrent python3.12 processes and 54 other Claude session wrappers on the box;
• the same suite passed twice earlier today on this branch's base (collected=21164 failed=0).

The first timeout was my fault — I ran a /nix-walking diagnostic (~80M inodes) concurrently with the gate. I stopped it by resolved PID and confirmed the PIDs were gone rather than zombies. The second timeout happened with that gone, on other sessions' load alone.

What I can vouch for: this diff is one comment block in a .sh and one .md, so it cannot reach Python. Targeted runs of every test that reads those paths — the three content gates, the CI-claim marker test, and the journald suite — are 254 passed. That is a narrower claim than "the gate passed", and I am not substituting one for the other.

Someone should run scripts/gate.sh --tier both on a quiet box before merging.

[ZacxDev]

Targeted evidence completed — 861 tests, still NOT a substitute for the gate

The PR body listed 254. The remaining tests that read claudedocs/ have now finished too:

test_closing_condition_single_source.py + test_drift_check.py
  + test_claim_work.py + test_claude_sessions.py     607 passed in 427.40s
content gates + CI-claim marker + journald suite     254 passed in  47.31s
                                                     ---
                                                     861 passed, 0 failed

That is every test in scripts/tests/ that reads either path this diff touches (scripts/diagnose-nix-disk.sh, claudedocs/), found by grep -rl.

🔴 This still is not "the gate passed." It is a claim about the targets I could enumerate as reachable, on one tier, run individually — not the full suite on the merged tree, and not the sandbox tier at all. The full pytest tier remains unobtainable on this box: two consecutive 3600s timeouts under load 72→92 from other sessions.

The merge condition is unchanged: run scripts/gate.sh --tier both on a quiet box, and nix build .#checks.x86_64-linux.{pytests,nodetests} one at a time. Neither has been done for this branch.