feat(dream): add reviewed Memory-to-Experience-to-Skill workflows

[Teingi]

Which issue or RFC does this PR close?

Closes #1509.

RFC: English · 中文.

Rationale for this change

Related task evidence can be repeated across Memory entries and Experience revisions without a reliable way to revisit the exact inputs, deduplicate their underlying observations, and propose a reviewed improvement. Artifact Dreaming adds an asynchronous workflow for turning selected Memory evidence into Experience and approved Experience into a managed Skill.

What changes are included in this PR?

• Implement refine_experience and derive_skill with durable runs, immutable input manifests, idempotency, leases, bounded generation, recovery fencing, and atomic Candidate/run commits.
• Resolve exact Memory entry versions and Experience lineage, group repeated root evidence, and recheck source eligibility, current authorization, entry activation, and target revisions. Prompt configuration remains lineage only.
• Preserve entry citations through Candidate revision and approval, expose provenance in Review Inbox, and reuse the existing Experience/Skill approval and package lifecycle.
• Keep ordinary Review validation separate from Dream generation budgets. Experience revision chains remain reviewable, and deriving a Skill from an Experience with 32 Sources preserves its direct Artifact reference without flattening those Sources into the Candidate. Iterative validation still checks transitive authorization and Memory validity; ordered Memory head locks protect approval against deactivation. Explicit Memory citations retain their required Source dependencies.
• Add HTTP, Python SDK, CLI, MCP, capability discovery, and generated integration operation tables, with bilingual RFC and usage documentation.
• Preserve upstream Topic Memory behavior, Prompt publication restrictions, and the reorganized documentation paths.

Are there any user-facing changes?

• New endpoints: POST and GET /v1/scopes/{scope_id}/dream, plus GET /v1/scopes/{scope_id}/dream/{run_id}. New work returns 202; replaying a terminal idempotent request returns 200.
• Add pc_dream_runs and nullable memory_citations columns on pc_artifacts and pc_artifact_candidate_versions, with additive initialization/migration for SQLite and OceanBase. Existing data remains readable. Explicit MySQL transactions preserve atomicity even with session autocommit enabled.
• Experience proposals and candidate revisions accept exact Memory citations. Omitting citations or setting them to null during revision retains them; an explicit empty array clears them. The Candidate's combined direct-reference limit remains 32.
• Dream results require review before becoming committed Artifacts. Generation is available only in the all process role; split api/background roles do not accept new Dream work. Dream and Topic Memory currently enforce separate execution concurrency limits.
• Supervisor integration, automatic input discovery, Profile refinement, cross-Scope evidence expansion, and automatic installation are outside this scope.

How was this change tested?

• make check: passed, including dependency lock consistency, pre-commit hooks, and type checks.
• make contract-test: 46 passed; generated Python and JavaScript contracts match OpenAPI.
• Review, Profile, HTTP authorization, and Dashboard regression tests: 95 passed, 2 optional-backend tests skipped.
• Three new regression scenarios on SQLite and isolated OceanBase databases: 6 passed. They exercise 12 successive Experience revisions, transitive Memory deactivation beyond Dream's depth budget, and Skill proposal/revision/approval from an Experience with 32 Sources, including permission revocation on an indirect Source. Dream's own depth budget remains enforced.
• Eight affected Dream evidence scenarios on both backends: 16 passed. They cover exact provenance, review edits, authorization, deactivation during generation, transitive Memory checks, root deduplication, atomic rollback, and Prompt lineage.
• After integrating upstream changes, Prompt API, MySQL schema, integration manifest, documentation contract, and packaging tests: 46 passed.
• Real configured LLM over a local HTTP server on SQLite and an isolated OceanBase database: Memory → Experience candidate → approval → recall → Skill candidate → approval, idempotent replay, and rejection of deactivated entries passed. Memory extraction used a deterministic fixture. Each generated Artifact used one model call; temporary databases were removed. These checks establish workflow correctness, not comparative quality or cost improvements.
• make docs-test: passed; 770 static pages and their internal links verified.
• git diff --check: passed.

AI usage statement

OpenAI Codex assisted with research, RFC writing, implementation, review fixes, merge integration, test execution, and PR preparation. Validation results and limitations are recorded above; no credentials or local environment files are included.

[Teingi]