feat(llm-obs): add agent-insights and model-pricing commands

[gsvigruha]

Summary

Closes the Agent Insights and model-pricing gaps between the dd-source LLM Obs MCP server (domains/ml-observability/shared/libs/mcp/tools/) and pup. Of that server's 41 LLM Obs tools, 34 already had pup equivalents at full parameter parity; this PR adds 5 of the remaining 7.

Still unimplemented after this PR: launch_llmobs_experiment and get_llmobs_bits_session.

Changes

• pup llm-obs agent-insights list|get|update-status|submit-feedback — proxies the four /agent-insights/* endpoints (src/commands/llm_obs.rs:1097)
• pup llm-obs model-pricing — proxies /pricing/model for deterministic cost projection (src/commands/llm_obs.rs:1247)
• Lifecycle statuses and usefulness verdicts are validated client-side against the server's declared enums, so a typo reports the valid values rather than a 400 (src/commands/llm_obs.rs:1123)
• --feedback takes target_key=usefulness[=reasoning]. Real target keys embed colons — suggested_evaluator:<eval_name> — so = is the only separator that round-trips them; a colon-separated entry is rejected rather than silently mis-parsed (src/commands/llm_obs.rs:1190)
• is_write_command_name now treats submit-* as a write (src/main.rs:10945), so submit-feedback is blocked under --read-only and reports read_only: false in the agent command catalog
• Optional filters are omitted from request bodies rather than sent as defaults, so the server's own defaults (status=for_review, limit=25) still apply
• docs/COMMANDS.md updated

Testing

18 new tests, all mock-based via the existing test_support helpers.

Positive: every filter populated and asserted body-for-body against the MCP tool definitions; omitted-argument bodies asserted to be empty; feedback parsing with and without reasoning; colon-bearing target keys and =-bearing reasoning preserved; exactly 25 feedback items accepted.

Negative: unknown status (both list and update-status), unknown usefulness, missing usefulness, empty target key, colon instead of =, 26 feedback items, model-pricing with neither --provider nor --model, plus 404/403/400/500 and no-auth paths.

Read paths were also verified against the live API: list with every filter, get (including the 404 path), model-pricing by exact model and by partial query across all providers, and cursor pagination returning a distinct second page. --read-only blocking both writes was verified end-to-end. The two write endpoints themselves were exercised against mocks only, to avoid mutating real insight state.

cargo fmt --check, cargo clippy --all-targets -- -D warnings, and cargo test are clean for this change. Five unrelated tests (cases::test_cases_timeline, dbm, monitors::test_monitors_diff_detects_changes, security::test_security_iocs_get, traces::test_spans_metrics_list) fail intermittently on main too — they race on the shared PUP_MOCK_SERVER env var and fall through to a DNS lookup of unused.local; each passes when run alone.

---

🤖 Generated with Claude Code