fix(gepa): fix 6 optimizer bugs — state advancement, cache key, eval heuristics, gold sets

Fix 1 (CRITICAL): state.currentGeneration now always advances in scoreAndSelect()
regardless of improvement. Previously stayed at N on no-improvement, causing mutate()
to recompute nextGen=N+1 and overwrite the same generation directory every round.

Fix 2 (CRITICAL): expand implement.jsonl from 1 → 13 entries covering scope creep,
ESM missing .js, over-engineering, timeout/excessive reads, missing tests, wrong file
target, console.log, duplicate abstraction, backwards-compat shim, fixture mutation,
config rewrite, factory pattern over-engineering.

Fix 3: replace keyword-presence heuristics in eval-phases.js with regex sentence-level
checks. A mutation appending glossary keywords no longer scores 100% — the prompt
must contain a meaningful sentence matching the pattern (e.g. "never use --no-verify"
not just "--no-verify" floating in a header).

Fix 4: eval cache key now hashes full variantContent instead of first 500 chars,
preventing cache collisions between variants that share a common preamble.

Fix 5: coding-tasks.jsonl expanded from 5 → 12 entries, adding ESM imports, scope
discipline, test-first, no-console, immutable update, error boundary, commit format.

Fix 6: crossover gen-1 skip is now logged explicitly instead of silently no-op.

Author: StackMemory Bot (CLI)