perf(skills): stop charging depth inference to every invocation

EXECUTION-PROFILES.md infers how hard a run should try. A stated depth, a
trivial ask, or a fast path already answers that, so forcing it on every
invocation of every dispatcher bought nothing and cost 751 tokens each
time. It is now conditional, read only when the depth is not already
obvious. Nothing was deleted.

Measured with the calibrated bill, before and after:
  plan   4.8K -> 4.1K tok  (-15%)
  qa     4.0K -> 3.3K tok  (-18%)
  review 3.7K -> 3.0K tok  (-19%)
  ship   3.8K -> 3.1K tok  (-18%)
  debug  3.6K -> 2.9K tok  (-19%)

Two ground-truth tests asserted the forced set was the old triad. They now
pin the new invariant instead, and additionally require that
EXECUTION-PROFILES stays reachable as a conditional read, so this cannot
decay into a silent deletion.

SHARED-JUDGMENT and AUTHORITY-POLICY stay forced. They bind every run:
evidence for claims, no fabrication, untrusted input, and approval before
merge or deploy.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: 2 people