Proposal: feat(api): experimental pipeline-run comparison

[mvanhorn]

Motivation

Tangle's core loop is experiment iteration: submit, monitor, clone, edit, resubmit. Clone exists. What does not exist is a read-only API that answers "what changed between these two runs?" Shopify's own launch post lists the failure mode: a teammate asks you to reproduce a result and you cannot tell which parameters, artifacts, or step timings differed.

MLflow, Vertex Experiments, and W&B ship run comparison as a first-class surface. Tangle already stores the ingredients:

• Pipeline-run annotations (PipelineRunAnnotation + PipelineRun.annotations)
• Per-execution artifact nodes with ArtifactData.hash, total_size, uri
• Container started_at / ended_at
• Cache reuse recorded as extra_data["reused_from_execution_node_id"]
• List/get already optionally include execution_summary / execution_status_stats

Those compose into a compare endpoint without a new table, a new launcher, or a new "pipeline run status" (which was rejected on #177).

Proposed solution

Add a read-only experimental endpoint:

GET /api/experimental/pipeline_runs/compare?ids={id1},{id2}

Exactly two IDs in this slice (N-way is out of scope). Auth: the caller must be allowed to get both runs (same created_by check as PipelineRunsApiService_Sql.get / admin skip). 404 if either is missing. 400 if ids malformed, duplicate, or count != 2.

Response (JSON, no new table): left/right run payloads plus annotation_diff and per-task diffs (status, cache reuse, timings, output artifact hashes). Match tasks by task_id_in_parent_execution. Do not emit a single run-level status field.

Default remains: existing GET/list bytes unchanged.

Out of scope: UI in tangle-ui, N-way compare, metric-file parsing, HTML report, writing annotations, triggering clone.

Happy to open a draft PR if this shape looks right. Looking for maintainer signal before writing code.