docs: shift FinWorldGraph pipeline to agent-conditioned world-state compilation

[MagicFitz]

Motivation

• Reframe the project center from building a standalone financial graph to compiling an agent-perceivable world state so the pipeline produces agent-conditioned observations as a primary output.
• Emphasize that WorldGraph is an intermediate representation and PerceptionPack (P_{a,t}) is a first-class output for downstream belief updates and simulation evaluation.
• Align the repo with an NLP-facing task definition that supports heterogeneous agents, temporal grounding, and evidence-grounded narratives.

Description

• Rewrote README.md to include a formal task definition (D_{<=t}, M_{<=t}, A → W_t, P_{a,t}) and a layered pipeline spec (Layer 0–8) covering configuration, normalization, extraction, canonicalization, world-state compilation, agent-conditioned projection, simulation interface, and evaluation.
• Added the agent-conditioned projection formulation P_{a,t} = Π(W_t, C_a, H_{a,<t}), a PerceptionPack-first schema, legacy-to-new module mapping, and recommended relational tables for an MVP.
• Included a scoped V1 implementation plan (20 stocks, 1 month of news, 3 agent archetypes), phased guidance to defer GNN-heavy training, and one-sentence contribution statements in English and Chinese.

Testing

• Verified repository file listing with rg --files | head -n 40 which executed successfully.
• Verified the updated documentation content with cat -n README.md which displayed the new pipeline and specs.
• Created the pull request via the repository make_pr helper and the automated PR creation returned a successful response.

---

Codex Task