Replace invented Guide patterns loop with canonical 3-phase

[PunchTheDev]

Summary

Rewrites Guide → Agent architecture patterns against the canonical source agents. The section previously claimed "best agents use a structured observe → plan → act loop rather than a single prompt" and showed a 6-phase loop (OBSERVE→PLAN→ACT→VERIFY→REFLECT→EXPORT) plus four library tiles, half of which are not used by any canonical agent.

Motivation

Audit against forge/agents/ and forge/examples/:

• agents/baseline/agent.py is deterministic geometry, no LLM call. Docstring is explicit: "This agent accepts the LLMClient parameter but does not use it … This is permitted; the harness requires the parameter to be present in the signature, not necessarily used." Guide framing ("best agents use a loop") implied this is wrong — it's actually one of two legitimate patterns.
• examples/llm-agent/agent.py uses literal 3-phase comments: ── Observe ──, ── Plan (LLM proposes dimensions) ──, ── Act (build geometry) ──. The Guide's 6-phase loop with VERIFY + REFLECT + EXPORT is not in any canonical agent.
• Library tiles: gmsh and numpy/scipy are not imported by any agent in agents/ or examples/. Only build123d (both LLM examples) and raw OCP (baseline) are verified.
• examples/llm-agent/ and examples/metric-aware-agent/ — the references the template literally tells users to read — were never linked from the Guide.
• Canonical LLMClient.chat(messages, max_tokens=...) API was never surfaced.

Pattern matches steps 359 / 365 / 367 / 368 / 369 / 370 / 371 / 372 / 373 — canonical-source check before tooltip / before claim.

Changes

• QuickstartGuide.tsx L639–743: lead reframed as "two canonical patterns, pick one" rather than "best agents use a loop"
• Pattern A card — Deterministic geometry, references agents/baseline/agent.py, 247-char tooltip quoting the baseline docstring verbatim
• Pattern B card — Observe → Plan → Act, references both LLM examples, 200-char tooltip pointing to the literal comment markers in examples/llm-agent/agent.py
• 3-phase loop block replaces the 6-phase loop, with concrete field names (build_volume_mm, load_point_mm, bolt_pattern_mm, min_wall_thickness_mm) and the actual llm.chat([{"role":"user","content": prompt}], max_tokens=256) call from examples/llm-agent/agent.py:53
• Harness-injection sentence with 213-char tooltip naming forge/sdk/llm.py and tying back to the Whitelisted models section
• Library tiles reduced from 4 to 2 — build123d (recommended, used by both LLM examples) and Raw OCP (BRep) (used by baseline). Each tile gets a tooltip citing the specific OCP / build123d imports in the canonical files. gmsh and numpy/scipy tiles removed (not used by any canonical agent).
• Three routed GitHub source links: examples/llm-agent/agent.py, examples/metric-aware-agent/agent.py, agents/baseline/agent.py

Verification: Puppeteer 1440×900 on /guide#patterns confirms 5 tooltips (lengths 247/200/213/176/135), 3 GitHub links pointing at the right files, old 6-phase / gmsh / numpy claims all gone, canonical 3-phase OBSERVE/PLAN/ACT + llm.chat( API + all 3 reference agent paths present. 0 new console errors (1 pre-existing 404).