feat(dsh-storage): complete session mirror (Prisma 7, 4 DBs) + CI suite

[Weaxs]

Complete dsh-storage implementation

• Projector verified against the dsh persistence catalog: envelope {type, seq, time, data}; per-step replacement usage accounting (usage chunks and assistant/message, cache buckets included, failed steps covered, persisted across resume); tool/result callId from message.source.callId with errors projected to exactly {name, code}; only the three surface events produce rows; session/title folds into the session row.
• Effect-based lifecycle (this cordis fork has no ready/dispose events): backend init at plugin load; in-flight mirror writes drained at the awaited session/flush checkpoint and on fiber unload, so one-shot headless runs never lose tail rows.
• Deterministic-PK upserts (sha256(sessionId + messageId)): no JSON-path filtering (PostgreSQL/MySQL-only feature), portable across connectors and race-free under event bursts.
• Four database providers (MySQL / PostgreSQL / SQLite / SQL Server): one Prisma schema variant each, driven through Prisma 7 driver adapters (adapter-mariadb, adapter-pg, adapter-libsql, adapter-mssql). Clients are pre-generated at package build time and shipped — consumers never run prisma generate. SQL Server has no Prisma Json type, so its variant maps JSON columns to text and the backend serializes on write (derived from provider, no config knob). archivedAt was dropped from the schema by owner decision.
• 56 vitest cases: pure projector logic, plugin wiring on a real cordis Context (mocked backend), and the backend against mocked prisma/adapters — covering the no-throw seam, write draining, per-session serialization, resume seeding, usage replacement, URL decoding, and credential redaction.

CI suite (all GitHub-hosted runners; free for this public repo)

• integration.yml — the matrix is auto-discovered from .github/scripts/integration/<pkg>.mjs. The base leg boots the real dsh headless profile, asks deepseek-v4-flash to read a seeded marker file, and asserts the full user → assistant(tool-call) → tool/result → assistant chain in SQLite (marker content, callId linkage, message_count >= 4). A sibling <pkg>.providers file adds backend-only legs (mysql/postgresql/sqlserver service containers: db push + real backend writes/reads) that need no secrets and therefore also run on fork PRs.
• labeler.yml — rule labels (area/*, type/*, size:*, platform/*) on every PR; P0–P3 proposed via the integration gateway for same-repo PRs only. (Starts firing once this workflow lands on main — pull_request_target restriction.)
• dsh-review.yml — two-axis (Standards/Spec) DSH review: trusted base checkout only, PR content wrapped as untrusted data, all model tools disabled, P0/P1 findings fail the job. (Same pull_request_target activation note.)
• ci.yml — moved to ubuntu-latest; lint (biome + pinned actionlint on workflow files), typecheck, build, and the vitest suites. Fork PRs get every secrets-free stage.

First verified end-to-end run

SCENARIO_OK (messages=5, toolCalls=1, title="读取文件并复述内容") — the base leg made a real deepseek-v4-flash round-trip, read the marker file through a tool call, and every database assertion passed. All five integration legs are green.

Verified locally vs. only provable in CI

• Verified locally: four real databases (sqlite + docker mysql/postgresql/sqlserver) through DatabaseBackend; dry-run of the full headless flow up to the LLM call.
• Only provable in CI (now green): the real LLM round-trip and workflow behavior on real PRs.