feat(skill): perf-trend render bench history report across recent main CI runs

[dryotta]

What

Adds a new perf-trend skill plus the supporting scripts/perf-trend.mjs helper and npm run perf:trend wrapper. The skill pulls the bench-output artifact from recent successful ci.yml runs on main, parses the criterion bencher output, joins it with each run's commit SHA + PR title, and renders a markdown trend report.

Why

The bench job's per-run job-summary table shows only latest vs budget. It tells you nothing about:

• whether a number is moving over time,
• whether a regression is new or chronic,
• whether other budgeted benches are silently missing because an earlier harness in cargo bench panicked.

The perf-trend skill answers all three on demand.

Report sections

1. ** Missing budgeted benches** names from the workflow's BUDGETS table that produced no output in the window. The CI summary step is silent about these; surfacing them is the headline reason this skill exists.
2. Summary per-bench row: latest, budget, window median, Δ vs median, status ( / over budget), ASCII sparkline (oldestnewest).
3. Movers top regressions / improvements (10% off median, 3 samples), each citing the commit SHA + PR title of the latest run.
4. Per-bench history <details> block with the full series.

What it found on current main

Running on the 14-day window already surfaces two latent issues the per-run CI summary has been silent about:

• hot_path/get_file_comments_large has been ~34 over its 20 ms budget (~7083 ms) for at least 8 consecutive runs.
• matching_bench panics with "yaml anchors/aliases not allowed in sidecars" on every run, short-circuiting cargo bench and dropping 4 of 5 budgeted benches from the output (matching/50_comments_1000_lines, fold_regions/large_100kb_jsonlike/default, parse_kql/pipeline_50_steps, strip_json_comments/large_100kb).

Both are pre-existing repo issues outside the scope of this PR, but the report makes them impossible to miss going forward.

Design

• Budgets parsed live from .github/workflows/ci.yml single source of truth; if the workflow's BUDGETS table changes, the script follows.
• 14-day default window matches the bench-output artifact retention.
• --summary flag appends the report to \ so a future scheduled-workflow caller can post the same renderer's output to a job page without rewriting it.
• Read-only never writes to the repo, never opens issues. Acting on findings is for the user or a follow-on skill (per user direction during design).

Verified

• npm run lint:skills (11 skills now, all npm run references resolve)
• npm run lint:markdown-surfaces
• npx eslint scripts/perf-trend.mjs
• node scripts/perf-trend.mjs --help
• node scripts/perf-trend.mjs --days 14 --runs 10 renders full report with Missing-benches section
• node scripts/perf-trend.mjs --summary with \ set writes 35 lines to the summary file