test(cpp): cross-runtime conformance suite and documentation

[kovtcharov]

Phase 5 · Validation · P5.2 — the gate that makes "parity" a measured claim rather than an assertion.

Conformance

• A test running the same SKILL.md corpus (tests/fixtures/skills/ + tests/fixtures/openclaw_skills/, 32 skills) through both the Python and C++ parsers, asserting identical accept/refuse verdicts and identical prompt bytes. In CI, not as a spot check.
• A gaia eval agent adapter for gaia-code following cpp/agents/bash/eval/bash_eval_adapter.py, with a committed baseline scorecard.

Documentation

Several docs currently assert the opposite of what this milestone ships. A functional change must update every doc that describes it:

• cpp/README.md — feature matrix lists "Audio / RAG / Stable Diffusion" as Python-only; the env-var table documents only the deprecated GAIA_CPP_* names when the code reads GAIA_BASE_URL / GAIA_MODEL_ID / GAIA_CONTEXT_SIZE; the project-structure tree omits 8 existing headers; only 2 of 6 agents are described.
• docs/cpp/overview.mdx — states "RAG and audio are available in the Python SDK".
• docs/cpp/api-reference.mdx, docs/cpp/custom-agent.mdx — new subsystems.
• New docs/cpp/skills.mdx, registered in docs/docs.json.
• docs/plans/cpp-webui-integration.md — its capability matrix asserts rag: ❌ and "MCP tools only (no RAG, no file tools)".
• docs/roadmap.mdx — the v0.22.5 C++ entry lists cancellation, event callbacks, tool policies, streaming, and runtime config as planned when all five already shipped.

Acceptance

• Conformance test is a required CI check.
• Byte-equality of the skills prompt block is asserted, not inspected.
• Committed eval baseline for gaia-code.
• Every doc listed above updated; grep for the stale claims returns nothing.

Plan: docs/plans/cpp-framework-parity.md

[kovtcharov]

Correction: the CI gate must not reference a nonexistent path

Found while reviewing the plan PR (#2806). This issue specified the conformance corpus as tests/fixtures/skills/ + tests/fixtures/openclaw_skills/ = 32 skills. tests/fixtures/openclaw_skills/ is not on main — it lands with PR #2693.

Making a required CI check point at a path that does not exist is how a gate silently passes by testing nothing.

Build the gate to discover the corpus from the fixtures directory rather than hardcoding a list. It then ships against the 6 skills on main today and widens to 32 automatically when #2693 merges — no follow-up PR, and no window where the check is green because it found zero inputs. Assert a non-zero corpus size so an empty discovery fails loudly.