[physics515]

Three post-mortems from this wave, written up because each is a fact about the tooling that the
code gives no hint of.

• PR [PAC-11] host-install: CI for the bare-metal host packaging contract #596 ran almost no checks for three pushes because its merge state was CONFLICTING;
  only its pull_request_target workflow ran, while CI, Host install, Container and
  crash matrix produced no run at all — not queued, not awaiting approval. Merging
  origin/main and resolving the CHANGELOG started all four within seconds. New check order for
  "CI did not run": gh pr view --json mergeable, then the trigger's event type, then runner
  availability.
• Both W8 lanes timed out mid-build with their work unpushed, and the harness deleted the
  worktree and the branch afterwards: only
  worktree-diffs/step-0/task-0-parallel-worker.patch survived, and it is against the harness
  base commit (4ee2629, a local commit) rather than origin/main, so it has to be applied
  per-file with --include or it "adds" files main already has. Records why a fresh worker with
  a written brief beat action: "resume" on a 400k-token session whose worktree no longer
  existed.
• [skip ci] suppresses pull_request workflows as well as pushes. The unpublished
  dictionary-doubling fix carried it in its subject, which would have landed a memory-bug fix
  and four version bumps with zero checks run. Dropped on the rebase onto main; [weftdb] Stop pipeline dictionaries doubling on every run; release 0.3.2 #612 is a draft
  until CI is green, and tagging 0.3.2 stays a human action (release.yml accepts only a CI run
  on main, and crates.io is unreachable from this machine, so publication state cannot be
  confirmed here at all).

Docs only — no code, no behaviour change.