fix(eval): harden GEPA discovery and reuse

[LukasParke]

What

• thread optimization scope into runtime field discovery
• use stable tool-index paths in discovery and mutation
• traverse current composite step kinds without restoring pre-feat!: unify Noetic public API naming #68 names
• bound per-field teacher proposals with the shared run pool
• reuse LLM-judge harnesses for identical provider/model configuration
• add timeout and kill grace to the Pi coding-agent RPC path

Why

GEPA discovery ignored the requested scope, mutable tool names made candidate paths unstable, composite traversal missed current step shapes, and judge harnesses were rebuilt for every score call.

This ports item 5, the remainder of fork/port/openrouter-fixes commit aedf712f, on top of item 4 #72.

Test plan

• 49 focused discovery/GEPA tests pass
• eval typecheck
• root lint
• sentrux check .

Dependency

Depends on #72 for the shared run pool and removed dead budget option. The branch currently includes that prerequisite and must be rebased through main after merge.

[vercel]

@LukasParke is attempting to deploy a commit to the Matt Apperson's projects Team on Vercel.

A member of the Team first needs to authorize it.

[LukasParke]

Agent: Code CI/structural/DCO are green where configured; remaining external red checks are compat with an empty upstream OPENROUTER_API_KEY and Vercel authorization. Where core CI is red, the only failure is the unrelated timing-sensitive executeRunCode retry-backoff test; affected local/full suites passed. I cannot rerun upstream Actions without admin rights.

[LukasParke]

Agent: Root-caused the repeated core CI failure to retry tests proxying global setTimeout and capturing sibling-test timers. Pushed a deterministic fix that tests the pure retry-delay function without mutating globals. DCO and Structural gate pass; CI rerun is in progress.

[LukasParke]

Agent: Root cause was Bun mock.module() leakage from optimizer.test.ts: mock.restore() does not evict the mocked dynamic-import module, so later GEPA environment-resolution tests could receive the stale fake bridge. Commit 112cde45 adds an explicit bridge-loader seam for optimizer tests, removes module mocking, and adds a full-suite regression. Verification: eval 284/284, eval test:ci, eval typecheck, root lint, and root test 1,446 pass / 5 skip. DCO and Structural gate are green; CI rerun is in progress.