Route dispatch runbook to --derive-held, not the hand-typed held set

[fdaviddpt]

Closes #586

skills/manager/phases/dispatch.md (the runbook the loop reads when it dispatches) and skills/manager/SKILL.md's spine copy both instructed the maintainer to hand-type the --against PATTERN set when checking a new lane's files against everything already running via scripts/lane_setup.py. #558 shipped --derive-held, which derives that same held set from open pull requests and the live lane registry -- but nothing routed the reader to it, so the fix existed one flag away from a runbook that never mentioned it.

Both documents now name --derive-held as the route for the disjointness check (a new lane against everything currently running), name could-not-derive-the-held-set as the third state -- not a plain "if it fails, type it by hand" -- under which the hand-typed --against remains the documented fallback, and note the mutual exclusion the script itself enforces.

The candidate-issue side is untouched: #267 still holds -- an issue's own files are not derivable from its body, so naming the candidate issue to check stays the maintainer's call, never the script's.

A second commit fixes a stale cross-reference the review round found in the first one. The bundling-check paragraph (below the disjointness paragraph) used to say it ran "the same ... call" as the paragraph above it -- true while both used plain --against, false the moment the disjointness paragraph was rewritten to lead with --derive-held. The bundling check compares one candidate's declared lane against one named running lane, not the --derive-held aggregate, so it still needs --against and now says so without the stale cross-reference; SKILL.md's table row is split into two rows for the same reason, so it stops conflating the two calls.

Below-bar, not filed: while auditing the diff, the reviewer confirmed everything else in both files reads correctly against scripts/lane_setup.py's own --derive-held/--against argparse wiring and the availability states in lane_report; nothing else adjacent to this diff was found stale.

Both byte budgets (scripts/skill_phases.py) still clear with headroom: SKILL.md 58,655 / 64,600, dispatch.md 25,186 / 26,100.

Test: tests/test_dispatch_runbook_derive_held_586.py, new, 7 assertions -- red against the pre-fix blobs (verified against 3eefc87~1), green after.