feat(cli): add configurable Insights analysis and evidence workflows

[liam-langchain]

Summary

Adds configurable Insights analysis and evidence drill-down using the existing LangSmith Go SDK. No API or SDK changes.

Command	Capability
insights create	Start a bounded report from flags, an analysis JSON file, or a saved configuration ID.
insights runs ID	Read a page of report evidence, optionally restricted to a category.
insights list	Read a bounded page, with offset and saved-config filtering.
insights get ID	Read status and results with project/workspace context.

Creation supports --preview-run with --dry-run, --file/-f, --categories, --attributes, --cluster-model, --summary-model, --config-id, --dry-run, and --output/-o. Categories/attributes flags accept JSON file paths. Analysis files use API field names and reject unknown fields, scheduler controls, and credential-validation overrides. File and saved-config modes reject conflicting analysis flags.

Samples accept 1–1000. Custom summary prompts require trace variables; categories reject duplicate trimmed names, and attribute names cannot contain whitespace. JSON files reject duplicate keys and excessive nesting.

Dry-run creates no job. Optional --preview-run UUID reads one project-scoped root run and returns variable bindings, missing paths, and unchecked paths without generating a summary. It does not verify the run matches the analysis window/filter or will be sampled; thread messages and feedback remain unchecked. It validates the local request but does not resolve a saved configuration, authorize a write, verify model availability, freeze a trace sample, or guarantee cost. Creation is not automatically retried after an ambiguous response.

Evidence includes extracted summaries/attributes and available run metadata; IO is explicitly marked as previews. Full IO uses run get --full. Attribute sorting is page-local, and evidence pagination exposes next_offset.

Listing behavior change: default limit is now 20, with a maximum page size of 100 and explicit --offset. The existing JSON array format is preserved. Full pages do not guarantee another page.

Saved-config authoring, generation, scheduling, and cancellation are not added. This independent PR targets main; #308 is the reference-only combined draft.

Manual testing

Build with make build, select an authenticated profile/workspace, and replace placeholders. See the README for analysis.json and the full workflow.

bin/langsmith --format json insights create --project-id PROJECT_ID --file analysis.json --dry-run
bin/langsmith --format json insights create --project-id PROJECT_ID --file analysis.json --dry-run --preview-run RUN_ID
# Review first: the next command incurs workspace model usage.
bin/langsmith --format json insights create --project-id PROJECT_ID --file analysis.json
bin/langsmith --format json insights get JOB_ID --project-id PROJECT_ID
bin/langsmith --format json insights runs JOB_ID --project-id PROJECT_ID --cluster-id CLUSTER_ID --limit 20
bin/langsmith --format json insights list --project-id PROJECT_ID --config-id CONFIG_ID --limit 20 --offset 0

Test Plan

• Request construction, category/attribute validation, strict bounded file parsing, conflicting modes, and saved-config-only payloads.
• Prompt/sample boundaries and preview flag conflicts; preview performs only a scoped run query and preserves null/false/zero with explicit missing and unchecked paths.
• Dry-run sends no creation request; creation retains no-retry behavior.
• Evidence pagination, project/category scope, page-local sorting parameters, and false/zero preservation.
• Bounded report listing preserves its JSON array and makes one page request.
• Live paid creation with manual categories/custom models and restricted-role/deployment coverage.

Corridor analysis was blocked by session policy. Local lint is deferred to CI because golangci-lint is unavailable.