docs(claude-md): trim root, move eval suite to evals/AGENTS.md (#3288)

* docs(claude-md): trim root, move eval suite to evals/AGENTS.md

Root CLAUDE.md goes from 316 to 85 lines. Net -264 / +33.

Keeps the load-bearing parts: Architecture: Rendering with the
placement table compressed to 7 rows, direction-aware mirroring
rule, one boundary-check rg command, Style Resolution Boundary
with the 'why', Document API Contract, Commands, Testing matrix.

Cuts content that's either derivable, redundant with the placement
table, or has a nested home: Project Structure tree (ls works),
the old 'Where to Look' table (subsumed), 'When to Modify Which
System' (subsumed), per-test-suite detail (already in tests/
behavior/AGENTS.md and tests/visual/AGENTS.md).

Eval suite (~60 lines) moves intact to a new evals/AGENTS.md so
the documentation isn't dropped; evals/CLAUDE.md symlinks to it
for cross-agent compatibility.

Brand, comments, JSDoc types, and generated-artifact tables are
move candidates for follow-up PRs (path-scoped rules under
.claude/rules/ or per-package AGENTS.md). Not in this PR to keep
it small.

Background: based on an N=3 exploratory benchmark + static audit
comparing variants of the root CLAUDE.md (baseline / lean /
lean+scoped-rules / organic bloat) against two SuperDoc tasks
(add doc-api operation, refactor RTL paragraph feature module).
Lean root matched baseline correctness in the final scored sample
at ~17% lower cost on the complex doc-api task; architecture-
violation rate was 0% in the final scored sample. C-lean-rules
had one incomplete-implementation run that flagged a real failure
mode (operation not wired into public API). N=3 per cell, Claude
Sonnet only, grep-based correctness. Directional only.

* docs(evals): use --filter form for pnpm scripts so commands work from any cwd

Per PR review: eval scripts live in evals/package.json, not the
root, so bare 'pnpm run eval'/'eval:e2e'/'view' fail with
ERR_PNPM_NO_SCRIPT from the repo root. Switched all references
to 'pnpm --filter @superdoc-testing/evals run ...' form.
Bare-form usage from inside evals/ called out separately for
operators already cd'd in.

Author: caio-pizzol