Dispatch cost inflation had three separate causes

The previous goalpost was to stop harness context from growing until a job hit forced compaction. The circuit breaker already contains a runaway job after the token bill explodes. It does not stop the dispatch from starting with a contaminated diff, a rewritten project snapshot, or a verbose test transcript that later turns keep resending.

What changed

Three independent root causes were still open.

Worktree base (gh:#2229). dispatch.stacking=auto treated every non-main checkout as a stack parent. A dispatch started from a long-lived workspace branch such as docs/claude/dev-preview-roadmap-execution cut the job worktree from that branch tip. The job diff against origin/main then included the workspace's unrelated history, and a PR opened from that branch carried the same files. Auto mode now stacks only on task branches (feat/, fix/, chore/, dispatch/, release/). Any other checkout resolves to the fresh origin/main tip. stacking=always and an explicit --base are unchanged.

Generated docs (gh:#2237). _docs_dir() returns the relative string project-docs. PR #2238 anchored the main todo.md write to _project_root(), but roadmap, costs, and memory rotation still archived through that relative path, so a job worktree could receive project-docs/archive/. Generated-doc writes and that rotation now go through _resolved_docs_dir(), which joins a relative docs dir to the shared project root. A real linked worktree regression covers both todo.md and the archive.

Verify contract (gh:#2240). When the worktree had changed tests, the injected How to Verify command was pytest <files> -v. A wide run prints one line per test, and that block is resent on later turns. The contract now asks for pytest -q, including the fallback text used when no test file list is known.

Circuit-breaker thresholds and .synlynk/policy.json are untouched. Those gates still contain a blowup after the fact. This change removes the three inputs that were inflating the bill.

Goalpost

A dispatched job should start from origin/main unless it is intentionally stacked on a task branch, should not dirty its worktree with a project-wide generated doc, and should verify with quiet pytest output. The next measurement is whether job cost stays near the dispatch estimate once these three sources are gone.