feat: channels + SubAgent finishing + harness-project split

[TYRMars]

Summary

Integration PR bundling several in-flight feature threads on feat/channels-next:

• harness-channel crate extracted — ChannelDispatcher trait, OutboundMessage / ChannelInstance / ChannelBinding value types, store traits. Shared vocabulary between harness-tools (channel.send), harness-store (persistence), and harness-server (adapter registry).
• Channel adapters in harness-server: DingTalk, Feishu, WeCom (webhook + app), OAuth callback routes, channels_inbound_routes. Companion apps/jarvis/src/channels_cli.rs for CLI registration.
• harness-tools: new channel.send tool wired to ChannelDispatcher.
• jarvis-web: OAuth result page, channels settings, sidebar/composer refinements, hint UI under built-in subagents in Settings → Agents.
• harness-project crate split out of harness-core — Project / Requirement / RequirementRun / Activity / Comment / Label / DocProject / ProjectMemory + their 8 store traits. Decouples Work-feature persistence from the agent loop.
• harness-observability crate split out of harness-core — Eval* (suite runs / verdicts / baselines) + Observed* (run / span / metric / dashboard) value types and persistence traits.
• SubAgent system finishing (this session): extra_args: Vec<String> field on ClaudeCodeConfig + composition-root wire-up for JARVIS_SUBAGENT_CLAUDE_CODE_ARGS; POST /v1/requirements/:id/review manual reviewer trigger reusing auto_mode::spawn_background_run (promoted to pub); env-override hint under built-in subagent cards in Settings → Agents; docs alignment in docs/proposals/subagents.zh-CN.md and CLAUDE.md (JARVIS_REVIEW_MODE → actual JARVIS_REVIEWER_AUTO_ACCEPT; removed phantom Activity kinds; added manual endpoint reference).
• Cleanup: remove unused svelte-learn/ playground.

Why bundled

These threads grew together on the same branch and share workspace-level Cargo manifest changes (new crate members, dep edits). Splitting now would mean three PRs whose merge order is fragile against the same Cargo.toml / Cargo.lock lines. The harness-project / harness-observability splits are also load-bearing for the channel work (Project store traits live in harness-project now), so they ride together.

Test plan

• cargo clippy --workspace --all-targets --exclude jarvis-desktop -- -D warnings — clean
• cargo test --workspace --exclude jarvis-desktop — exit 0
• New claude_code::tests::extra_args_* (3 tests: serde roundtrip, legacy-JSON default, Command args assembly order) — pass
• Web preview verified Settings → 子智能体 env-hint footer renders all 5 JARVIS_SUBAGENT_* codes
• Manual end-to-end of POST /v1/requirements/:id/review against a Review-status Requirement (needs JARVIS_DB_URL + seeded data — recommend reviewer runs jarvis-dev-build + curl 202 / 409 / 404 / 503 paths)
• Manual smoke of each channel adapter (DingTalk / Feishu / WeCom / WeCom App / OAuth callback) against staging webhook endpoints

Reviewer notes

• crates/harness-core/src/{activity,comment,doc,label,project,project_memory,requirement,requirement_run}.rs are git mv'd to crates/harness-project/src/ — diff should be mostly clean rename.
• crates/harness-core/src/observability.rs moved to crates/harness-observability/src/lib.rs.
• crates/harness-core/src/channel_{binding,instance}.rs deleted (moved into new harness-channel crate).
• Composition root (apps/jarvis/src/serve.rs + apps/jarvis/src/subagents.rs) is the only place env vars are read; library crates remain env-free per CLAUDE.md.

🤖 Generated with Claude Code