docs(worktree): scratch space is not isolated between workers

[stellarfeline]

The harness hands every worker a scratchpad path it calls "session-specific, isolated". The isolating token is the planner session's id — so every worker fanned out from one dispatch gets the same string. One flat directory, no per-agent segment.

Measured, 2026-08-09. Four concurrent workers, four independent before/after trees, one namespace:

worker	names chosen
1	base/ · after/
2	out-base/ · out-new/ (both nobodys-cave-island)
3	zh-base/ · zh-new/
4	out/base-<campaign> · out/new-<campaign>

Worker 1's base/ was replaced mid-run by worker 2's repo checkout — verified: $SP/base/campaigns now resolves into wt-state. Worker 1's own setup line was rm -rf $SP/base && mkdir -p $SP/base, so with the arrival order reversed it would have deleted worker 2's tree. Same shape as the git add -A near-miss the constitution already records, one layer over.

This was not bad luck. A before/after byte comparison is the evidence this project demands of everyone who touches emission, and base is the first word all of them reach for.

Why it is worse than a lost directory

The loud failure is the safe one — worker 1 got ENOENT and noticed. The quiet one is the danger. A byte-identity proof is find base/ | shasum against the same over after/. Had the foreign tree landed before the hash instead of after, the run would have hashed someone else's repo. Worse: worker 2 was writing build outputs of the same campaign. Had those names collided, a worker could have compared worker 2's before-tree against worker 2's after-tree and reported its own change byte-identical — hundreds of files, all matching, a number nobody can re-derive from the PR.

That is a green gate that binds to nothing, and review cannot see it, because the output is indistinguishable from a pass.

The asymmetry against the worktree-collision precedent this mirrors: a code leak fails CI; a corrupted evidence tree fails nothing. It emits a sentence in a PR description.

What this PR does, and what it does not

Adopts the convention — each worker gets its own scratch subdirectory, named in the dispatch prompt, same token as its worktree — and records the measurement so the next session does not rediscover it.

It does not build the real fix, which is named in the file: a baseline hash manifest that records the git SHA and delvec --version it was produced from, with the comparison asserting them. That turns a swapped tree into a loud failure and covers every piece of before/after evidence, not just the ones in a shared directory. Filed separately.

What CI proves

Doc-only. check-doc-dupes and check-reference-versions green.