runner: restate the run's configuration in every PR comment

[adriangb]

Result comments link back to the triggering comment, but the link alone doesn't say what the run did — you have to click through to find out which benchmark, which refs, which env vars. Every runner comment (running, completed, and failed) now opens with the comparison line and a collapsed Run configuration block:

🤖 Benchmark completed (GKE) | [trigger](…#issuecomment-100)

**Instance:** `c4a-standard-48` (12 vCPU / 65 GiB)

Comparing my-branch (aaaa111bbbb) to v45.0.0 [diff](…/compare/cccc222dddd..aaaa111bbbb)

<details><summary>Run configuration</summary>

```yaml
run benchmark tpch
env:
  CARGO_BUILD_JOBS: "1"
baseline:
  ref: "v45.0.0"
changed:
  env:
    DATAFUSION_RUNTIME_MEMORY_LIMIT: "2G"

… (lscpu, Details, Memory Pool Peaks, Resource Usage unchanged) ```

Why re-rendered rather than quoted verbatim

The trigger comment and the run it produces describe different things. One run benchmark tpch clickbench_1 fans out into a job — and a pod, and a comment — per name, and a bare run benchmarks expands to the repo's default suite. Quoting the trigger into each result comment would list benchmarks that comment's own run never touched.

So the block is re-rendered from the job, in the same YAML the trigger accepts. Pasting it back into a PR reproduces that single run; yaml_re_parses_as_the_same_request asserts the rendering round-trips through detect_benchmark. Empty sections are omitted, so an unconfigured run renders as just run benchmark tpch. Values are always quoted so "1" doesn't read back as an int.

Changes

• controller/src/runner/trigger.rs (new) — Comparison::line() and config_block(), with tests
• job_manager.rs — passes SHARED_ENV_VARS as a JSON map. These were already set individually on the pod, but the runner couldn't tell which of its env vars came from the trigger's shared env: block and which the controller injected (sccache, tokens, …). The map is for reporting only; it changes nothing about what the benchmark sees.
• runner/config.rs — new shared_env_vars field
• bench_datafusion.rs, bench_arrow.rs — both comments use the shared helpers; the running comment's ad-hoc BENCH_NAME= / BENCH_FILTER= lines are now covered by the block
• bin/runner.rs — failure comments get the config block too, without a comparison line (a failure can precede either checkout)
• README.md — documents the block and why it isn't a verbatim copy

One subtlety: BENCH_FILTER reaches the runner both as bench_filter and as a shared env var when set through env: in a trigger comment (as the README documents), so it renders once — as a # BENCH_FILTER: note only in the main-tracking workflow case, where it's set outside the env block.

Testing

cargo test -p benchmark-controller — 154 passed, 0 failed. cargo clippy --all-targets clean, cargo fmt --check clean.

🤖 Generated with Claude Code