feat: route galaxy-brain workhorse writers to gpt-5.6-luna by default

[scode]

Six treeward features and a planted bug, six models, one attempt each: every model passed every hidden test first
time, the gate's work-done checks rejected three results (none luna's), and luna medium ended as the cheapest clean
record — $0.64 against $5.57 for terra and $21.73 for sonnet
(https://claude.ai/code/artifact/43a3d4f1-fd32-41df-84bc-d62d6fb1f248). The routing table still started clear-spec
implementation at terra, and the rule that keeps a delegate in the orchestrator's own model family sent a Claude
orchestrator to sonnet for the same work at roughly thirty times the price. Both defaults were guesses from before the
eval.

Luna's measured weaknesses are judgment on questions the spec leaves open and long-context retrieval. The checkpoint
protocol added earlier moves the first to the orchestrator before code exists — the guidance eval's checkpoint arm
went 8 for 8 across the four cheap models, luna included, on a feature the same models had gotten right once in eight
runs without it — and the existing terra exception covers the second. So clear-spec implementation now starts at luna,
and tree-editing delegates in the two workhorse profiles — mechanical work and clear-spec implementation — default to
luna from every orchestrating harness, overriding the same-family rule for writers. Read-only fan-out is deliberately
not covered: cheap native sub agents beat per-delegate shell-out overhead there, and the eval measured writers. The
text also names what the default trades — delegates run under the shell-out path's permission-bypass flags rather
than the native permission system, and spend moves from subscription to metered API — and tells the orchestrator when
to surface each to the user.

The same eval showed luna logging zero decisions where muse logged eleven on the same diff. That is not evidence the
work was simpler, and the checkpoint text now says so, so an orchestrator does not read an empty log as a settled
design and skim the diff. The evidence is one eval on one small, well-documented Rust repository at n = 1 per cell,
with no cell in which any model failed a hidden test — it separates prices, not failure rates. What would reverse
this is a run on a larger or messier codebase where luna fails where terra does not; the text says so next to the
rule.