tech debt: optimize_agents is forked into two modules sharing one import name, with a test each #14873

Found while fixing #14518.

optimize_agents exists twice, forked, with a separate test file each:

Neither is deleted here, per "consolidate, never fork — measure before picking
the survivor". This issue is the measurement and the survivor decision.

Why it is not merely cosmetic

Both modules want the top-level name optimize_agents. In a shared pytest
process the first one imported binds sys.modules["optimize_agents"] — and
scripts/optimize_agents_test.py:51 does exactly that binding explicitly. The
sibling test was routed through the utilities. namespace in the #14518 pass
specifically so the two cannot collide, but that is a containment measure, not
a fix: whichever module loses the race would otherwise have its tests run
against the other implementation and pass.

That is the same failure shape as the rest of this cluster — a test that is
green while exercising something other than what it names.

Acceptance criteria

• One survivor chosen, with the measurement recorded here (which callers
  each has, which behaviours only one has — the newer copy has atomic
  writes and a dry-run guard the older lacks)
• The other's unique behaviour is merged into the survivor before it goes
• One test file, exercising the survivor
• No module name is bound by two different files

Related

• bug(infra): 44 imports under shared/scripts whose target does not resolve — F821-clean scripts that still cannot run #14518 — where this surfaced
• safety(infra): optimize_agents.py now rewrites .claude/agents/*.md in place with no --dry-run, after the path fix made it reachable #14546 — added the atomic-write / dry-run behaviour to one copy only